diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_0/index_tensor([1319, 1319])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_0/index_tensor([1319, 1319])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5dfb11e7cdef4102d745c5b8014bb83588d5fc58 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_0/index_tensor([1319, 1319])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up meticulously depicts a single bottle of red wine, its deep green glass framed by glistening droplets of condensation. The bottle stands perfectly still on the perfectly polished, reflective surface of a dark mahogany table. A single, warm, low-key light source from the side outlines the bottle, creating sharp highlights in the droplets and casting long, deep shadows, creating an atmosphere of tense silence. Suddenly, without cause, the bottle topples sideways. It falls onto the tabletop with a sharp crack, instantly shattering in a dramatic explosion of movement. As the crimson wine spurts out, dark, glistening shards of glass scatter outward in a spectacular, viscous splash, its droplets arcing through the air before crashing and rapidly spreading across the polished wood, staining its perfect surface and chaotically reflecting warm light in dark, sprawling pools.", + "seed": 27710, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_0/index_1319_seed_27710_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": -0.061328813806176184, + "vlm_logprob": -0.061328813806176184, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.061328813806176184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18797507882118225 + ], + "opposite_logprob_tokens": [ + -1.8129750490188599 + ], + "logprob_reward": -0.18797507882118225, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18797507882118225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1309211105108261, + 0.09878066182136536, + 0.13392925262451172, + 0.2421579509973526, + 0.13147588074207306, + 0.11153660714626312, + 0.053574562072753906, + 0.03268823027610779, + 0.02475002594292164, + 0.04018569365143776 + ], + "credit_map": [ + 0.02313350699841976, + 0.010179580189287663, + 0.018161937594413757, + 0.015190903097391129, + 0.011754336766898632, + 0.010709377937018871, + 0.012266695499420166, + 0.029524775221943855, + 0.009303771890699863, + 0.013336175121366978, + 0.012920350767672062, + 0.011713952757418156, + 0.011931304819881916, + 0.013095217756927013, + 0.016538599506020546, + 0.009941289201378822, + 0.010323937982320786, + 0.012984167784452438, + 0.014890201389789581, + 0.010744615457952023, + 0.02246180921792984, + 0.029241198673844337, + 0.018850307911634445, + 0.01443302258849144, + 0.01880989409983158, + 0.02998148836195469, + 0.028377564623951912, + 0.04867533966898918, + 0.02763875387609005, + 0.03240269795060158, + 0.03436703607439995, + 0.021905168890953064, + 0.010419189929962158, + 0.01721632108092308, + 0.013975542970001698, + 0.013730382546782494, + 0.01490748766809702, + 0.009637061506509781, + 0.039928022772073746, + 0.011661874130368233, + 0.009312791749835014, + 0.019640430808067322, + 0.01388454344123602, + 0.021630464121699333, + 0.014345319010317326, + 0.009649469517171383, + 0.014064515009522438, + 0.009009077213704586, + 0.005876658018678427, + 0.007220948114991188, + 0.0068800561130046844, + 0.005107433535158634, + 0.009022326208651066, + 0.0070327650755643845, + 0.005113917402923107, + 0.00732045853510499, + 0.0027847483288496733, + 0.004512113984674215, + 0.00591318728402257, + 0.0045148879289627075, + 0.006609498057514429, + 0.002729240572080016, + 0.0027455738745629787, + 0.0028789807111024857, + 0.003869422944262624, + 0.003196343779563904, + 0.0035288541112095118, + 0.004155328962951899, + 0.0028110421262681484, + 0.002826911397278309, + 0.002018776722252369, + 0.002343345433473587, + 0.007107060868293047, + 0.005161780398339033, + 0.003963775932788849, + 0.0035537653602659702, + 0.0053405254147946835, + 0.004006757866591215, + 0.005536348093301058, + 0.005515677854418755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.18797507882118225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908133953809738 + ], + "opposite_logprob_tokens": [ + -2.464081287384033 + ], + "logprob_reward": -0.08908133953809738, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908133953809738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21747377514839172, + 0.21920689940452576, + 0.11474301666021347, + 0.07997595518827438, + 0.053544316440820694, + 0.15937304496765137, + 0.052694354206323624, + 0.02783784456551075, + 0.04479188844561577, + 0.03035888448357582 + ], + "credit_map": [ + 0.020824093371629715, + 0.019048327580094337, + 0.026167990639805794, + 0.028283091261982918, + 0.04694151133298874, + 0.02138562873005867, + 0.028674567118287086, + 0.02614857442677021, + 0.020623624324798584, + 0.01731601171195507, + 0.0285037849098444, + 0.017572686076164246, + 0.02935320883989334, + 0.053049806505441666, + 0.026373758912086487, + 0.0264140535145998, + 0.009648486040532589, + 0.013170071877539158, + 0.01064794510602951, + 0.013655357994139194, + 0.017465583980083466, + 0.021990317851305008, + 0.016309168189764023, + 0.011856098659336567, + 0.010140389204025269, + 0.0103985033929348, + 0.008867804892361164, + 0.016059134155511856, + 0.008627024479210377, + 0.009615116752684116, + 0.008132501505315304, + 0.008135482668876648, + 0.004050432704389095, + 0.008837482891976833, + 0.0059225051663815975, + 0.007792574819177389, + 0.006726395338773727, + 0.005633316934108734, + 0.009311385452747345, + 0.005270225461572409, + 0.014201391488313675, + 0.02027803100645542, + 0.02024885080754757, + 0.02120138145983219, + 0.019152676686644554, + 0.02040814608335495, + 0.028349777683615685, + 0.01553281769156456, + 0.005948830861598253, + 0.007543068379163742, + 0.006264513358473778, + 0.003847207175567746, + 0.012932335957884789, + 0.006152438465505838, + 0.003776978701353073, + 0.006228981539607048, + 0.003565756604075432, + 0.00396279850974679, + 0.004279026761651039, + 0.0025097811594605446, + 0.003984491806477308, + 0.004109448753297329, + 0.0029477551579475403, + 0.002478791167959571, + 0.006701724138110876, + 0.00706108333542943, + 0.00553539814427495, + 0.0038275904953479767, + 0.007727541495114565, + 0.005736175458878279, + 0.0035958990920335054, + 0.004606477916240692, + 0.0028775541577488184, + 0.004100365098565817, + 0.004061865620315075, + 0.0028378264978528023, + 0.0047101532109081745, + 0.004287317860871553, + 0.004174225032329559, + 0.0033095800317823887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.08908133953809738 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210343360900879 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014210343360900879, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210343360900879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11323867738246918, + 0.14953523874282837, + 0.117213174700737, + 0.2558083236217499, + 0.14639736711978912, + 0.0627121701836586, + 0.04068917781114578, + 0.02891455590724945, + 0.03030102141201496, + 0.055190254002809525 + ], + "credit_map": [ + 0.014920571818947792, + 0.011630664579570293, + 0.017234820872545242, + 0.01284810807555914, + 0.017374012619256973, + 0.011886083520948887, + 0.01264867652207613, + 0.014695736579596996, + 0.015731172636151314, + 0.020646775141358376, + 0.01917804218828678, + 0.01841946505010128, + 0.01680276356637478, + 0.016297323629260063, + 0.02466776594519615, + 0.01779193989932537, + 0.01080284547060728, + 0.014210470952093601, + 0.01281302236020565, + 0.025685608386993408, + 0.012082337401807308, + 0.012992321513593197, + 0.015247426927089691, + 0.013379138894379139, + 0.025515705347061157, + 0.049172088503837585, + 0.04068896174430847, + 0.04183177649974823, + 0.02314755879342556, + 0.029245810583233833, + 0.026081452146172523, + 0.020124975591897964, + 0.01380001101642847, + 0.01635877974331379, + 0.01651262491941452, + 0.017601357772946358, + 0.01416575163602829, + 0.01583082601428032, + 0.03870289772748947, + 0.013425130397081375, + 0.007960797287523746, + 0.015552133321762085, + 0.00743332551792264, + 0.008797590620815754, + 0.008931756019592285, + 0.005117278546094894, + 0.0051814476028084755, + 0.0037378394044935703, + 0.007395267020910978, + 0.0068955994211137295, + 0.005073918029665947, + 0.00308340135961771, + 0.007110478822141886, + 0.0037836465053260326, + 0.003633350133895874, + 0.0037135190796107054, + 0.002722567180171609, + 0.004075174685567617, + 0.004508147481828928, + 0.0031932529527693987, + 0.005981415510177612, + 0.003929965663701296, + 0.002532585058361292, + 0.001971449935808778, + 0.003472693031653762, + 0.004410897381603718, + 0.0057375188916921616, + 0.003072228981181979, + 0.004794692620635033, + 0.002594251651316881, + 0.0027964406181126833, + 0.0034222980029881, + 0.004904107190668583, + 0.007082131691277027, + 0.0073152887634932995, + 0.00571640208363533, + 0.00862038228660822, + 0.01066032238304615, + 0.006068225018680096, + 0.004823398310691118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.014210343360900879 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738557480275631 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738557480275631, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738557480275631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062211278825998306, + 0.07444053888320923, + 0.20349077880382538, + 0.09462130069732666, + 0.2683534026145935, + 0.16817671060562134, + 0.04183744266629219, + 0.015027453191578388, + 0.04844984784722328, + 0.02339131385087967 + ], + "credit_map": [ + 0.007851498201489449, + 0.004435098264366388, + 0.010629681870341301, + 0.0054428596049547195, + 0.007909209467470646, + 0.010562038980424404, + 0.005286829546093941, + 0.010094059631228447, + 0.010169620625674725, + 0.012248354963958263, + 0.009560654871165752, + 0.006877602078020573, + 0.008005241863429546, + 0.00870796013623476, + 0.00934804417192936, + 0.009523062035441399, + 0.016890618950128555, + 0.024625523015856743, + 0.018847648054361343, + 0.024675410240888596, + 0.03278122842311859, + 0.036026060581207275, + 0.027142370119690895, + 0.022501908242702484, + 0.007382107432931662, + 0.01071134302765131, + 0.012958589009940624, + 0.013370441272854805, + 0.011632824316620827, + 0.015248432755470276, + 0.013102716766297817, + 0.01021484937518835, + 0.022400150075554848, + 0.029648898169398308, + 0.03217611461877823, + 0.026664135977625847, + 0.024173973128199577, + 0.01764124073088169, + 0.09247734397649765, + 0.023171527311205864, + 0.015538748353719711, + 0.027645139023661613, + 0.020369678735733032, + 0.03164830431342125, + 0.017451468855142593, + 0.015003892593085766, + 0.02721552550792694, + 0.013303954154253006, + 0.006014730781316757, + 0.0060844700783491135, + 0.008151443675160408, + 0.0037733835633844137, + 0.0074289520271122456, + 0.0031624999828636646, + 0.003607042133808136, + 0.003614919725805521, + 0.0016488991677761078, + 0.0019245878793299198, + 0.002495056949555874, + 0.0015441972063854337, + 0.0029764524661004543, + 0.0018010061467066407, + 0.001242640195414424, + 0.0013946129474788904, + 0.007151671219617128, + 0.007338964380323887, + 0.005861093755811453, + 0.003635470289736986, + 0.009789744392037392, + 0.005874428432434797, + 0.004108772613108158, + 0.004689699038863182, + 0.003421423491090536, + 0.0025301435962319374, + 0.0028003256302326918, + 0.0033414135687053204, + 0.0028902264311909676, + 0.003262112382799387, + 0.0026305096689611673, + 0.002515158848837018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.006738557480275631 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638749830424786 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638749830424786, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638749830424786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03644683584570885, + 0.020393148064613342, + 0.05446567013859749, + 0.08044130355119705, + 0.2589666247367859, + 0.2869170606136322, + 0.07515408098697662, + 0.034832488745450974, + 0.096043661236763, + 0.056339140981435776 + ], + "credit_map": [ + 0.003978426102548838, + 0.003131567034870386, + 0.004300480708479881, + 0.005098595749586821, + 0.006563163362443447, + 0.0038685258477926254, + 0.004456100519746542, + 0.005049979779869318, + 0.0026363155338913202, + 0.002337715355679393, + 0.0025299889966845512, + 0.0021986952051520348, + 0.0018652675207704306, + 0.0026017366908490658, + 0.002252256963402033, + 0.00397117156535387, + 0.005794494412839413, + 0.0077416179701685905, + 0.006655963137745857, + 0.011396982707083225, + 0.0050484552048146725, + 0.0057890512980520725, + 0.00661539426073432, + 0.005423711612820625, + 0.0074185640551149845, + 0.012208192609250546, + 0.010242682881653309, + 0.019146453589200974, + 0.006508172024041414, + 0.007397768087685108, + 0.009614208713173866, + 0.00790525320917368, + 0.022562799975275993, + 0.03931029140949249, + 0.029834125190973282, + 0.026104293763637543, + 0.028097737580537796, + 0.0147045673802495, + 0.08038999140262604, + 0.01796281337738037, + 0.029371244832873344, + 0.051733069121837616, + 0.04137660190463066, + 0.030640993267297745, + 0.026426641270518303, + 0.020111287012696266, + 0.05420044809579849, + 0.03305678442120552, + 0.013548742048442364, + 0.010478253476321697, + 0.010817455127835274, + 0.006842195987701416, + 0.012846237048506737, + 0.0056193554773926735, + 0.008064120076596737, + 0.006937725003808737, + 0.004341539926826954, + 0.004975152667611837, + 0.005965965799987316, + 0.004653871525079012, + 0.006661524996161461, + 0.0027640245389193296, + 0.002818831941112876, + 0.0026515780482441187, + 0.01621105521917343, + 0.013279017992317677, + 0.015797054395079613, + 0.014236952178180218, + 0.012884795665740967, + 0.006682068109512329, + 0.008964650332927704, + 0.007988081313669682, + 0.00865987129509449, + 0.006068216636776924, + 0.007176853250712156, + 0.008065853267908096, + 0.006210619118064642, + 0.008290117606520653, + 0.006250767037272453, + 0.005616836715489626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.008638749830424786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up meticulously depicts a single bottle of red wine, its deep green glass framed by glistening droplets of condensation. The bottle stands perfectly still on the perfectly polished, reflective surface of a dark mahogany table. A single, warm, low-key light source from the side outlines the bottle, creating sharp highlights in the droplets and casting long, deep shadows, creating an atmosphere of tense silence. Suddenly, without cause, the bottle topples sideways. It falls onto the tabletop with a sharp crack, instantly shattering in a dramatic explosion of movement. As the crimson wine spurts out, dark, glistening shards of glass scatter outward in a spectacular, viscous splash, its droplets arcing through the air before crashing and rapidly spreading across the polished wood, staining its perfect surface and chaotically reflecting warm light in dark, sprawling pools.", + "seed": 27711, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_0/index_1319_seed_27711_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": -0.02899448312819004, + "vlm_logprob": -0.02899448312819004, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02899448312819004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09487434476613998 + ], + "opposite_logprob_tokens": [ + -2.469874382019043 + ], + "logprob_reward": -0.09487434476613998, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09487434476613998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10641546547412872, + 0.4274919033050537, + 0.08165174722671509, + 0.10909467190504074, + 0.02883315272629261, + 0.1422724723815918, + 0.024620700627565384, + 0.02338838018476963, + 0.030388187617063522, + 0.025843283161520958 + ], + "credit_map": [ + 0.009809385985136032, + 0.009543695487082005, + 0.009233486838638783, + 0.00964381080120802, + 0.02194829098880291, + 0.01874028891324997, + 0.013143153861165047, + 0.014353365637362003, + 0.08762383460998535, + 0.021345393732190132, + 0.03454126417636871, + 0.02146824635565281, + 0.020204588770866394, + 0.05903104692697525, + 0.03885328397154808, + 0.14442425966262817, + 0.010626383125782013, + 0.0068288519978523254, + 0.024399371817708015, + 0.00666039576753974, + 0.004696200136095285, + 0.00855565071105957, + 0.005445418879389763, + 0.014439478516578674, + 0.0193618256598711, + 0.016819199547171593, + 0.01523729506880045, + 0.014197374694049358, + 0.01095262449234724, + 0.010257784277200699, + 0.009089503437280655, + 0.01317907590419054, + 0.0048688119277358055, + 0.003249244298785925, + 0.0032806855160743, + 0.002631782554090023, + 0.006213856860995293, + 0.0022661196999251842, + 0.0036399748641997576, + 0.0026826770044863224, + 0.02916991524398327, + 0.035823699086904526, + 0.008113620802760124, + 0.010890158824622631, + 0.012284555472433567, + 0.01994161866605282, + 0.010124675929546356, + 0.0159242395311594, + 0.0020175690297037363, + 0.0015393104404211044, + 0.003954194486141205, + 0.0028533670119941235, + 0.0030421284027397633, + 0.0033045662567019463, + 0.0036656789015978575, + 0.0042438870295882225, + 0.0014681934844702482, + 0.0013885616790503263, + 0.005027179140597582, + 0.003020713571459055, + 0.0017747365636751056, + 0.0029842311050742865, + 0.003395383246243, + 0.004329382907599211, + 0.0019639157690107822, + 0.003431065706536174, + 0.0024275658652186394, + 0.002024702960625291, + 0.002524143550544977, + 0.0057561458088457584, + 0.004392394796013832, + 0.00786825455725193, + 0.0022331613581627607, + 0.0016236481023952365, + 0.0042459359392523766, + 0.00267476262524724, + 0.00216672639362514, + 0.003989950753748417, + 0.004786061588674784, + 0.004123037215322256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.09487434476613998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01423149835318327 + ], + "opposite_logprob_tokens": [ + -4.2642316818237305 + ], + "logprob_reward": -0.01423149835318327, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01423149835318327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07708709686994553, + 0.7141295075416565, + 0.08150868117809296, + 0.02943943254649639, + 0.030165042728185654, + 0.026629911735653877, + 0.011003551073372364, + 0.007167179137468338, + 0.01834668591618538, + 0.004522922448813915 + ], + "credit_map": [ + 0.0056654601357877254, + 0.006910415831953287, + 0.007972965948283672, + 0.00811996590346098, + 0.014053585939109325, + 0.012990588322281837, + 0.010415853001177311, + 0.01095825619995594, + 0.14978930354118347, + 0.034281834959983826, + 0.05736541375517845, + 0.028858212754130363, + 0.022978294640779495, + 0.10027001053094864, + 0.06535268574953079, + 0.2552337944507599, + 0.018612723797559738, + 0.011468745768070221, + 0.0065469564869999886, + 0.005469298921525478, + 0.004296533297747374, + 0.01117080170661211, + 0.00651122909039259, + 0.01743238978087902, + 0.008225174620747566, + 0.007952168583869934, + 0.002368172397837043, + 0.0025287114549428225, + 0.0024273048620671034, + 0.0019489842234179378, + 0.0015584074426442385, + 0.002430505584925413, + 0.0023722536861896515, + 0.002224001567810774, + 0.004577342886477709, + 0.002966489875689149, + 0.004431137349456549, + 0.0033848495222628117, + 0.003962599206715822, + 0.006246369797736406, + 0.005516051314771175, + 0.006229447666555643, + 0.002434452995657921, + 0.002089043380692601, + 0.002150384010747075, + 0.0029535675421357155, + 0.002128231106325984, + 0.0031287348829209805, + 0.0012701655505225062, + 0.0011823269305750728, + 0.001386177260428667, + 0.001799233490601182, + 0.0014003547839820385, + 0.001421406865119934, + 0.0012391111813485622, + 0.0013047748943790793, + 0.0007112048915587366, + 0.0006411958020180464, + 0.0012871571816504002, + 0.0008089825278148055, + 0.0007822454790584743, + 0.0011513193603605032, + 0.0008940757834352553, + 0.0008909982279874384, + 0.0015429070917889476, + 0.00271091191098094, + 0.0014424700057134032, + 0.001505212509073317, + 0.001854436588473618, + 0.0037369467318058014, + 0.0018972971010953188, + 0.0036565037444233894, + 0.00039245616062544286, + 0.00034869738738052547, + 0.0008289027027785778, + 0.0004946322878822684, + 0.0005008020671084523, + 0.000630434020422399, + 0.0007765635964460671, + 0.0005504338187165558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.01423149835318327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060169354081154 + ], + "opposite_logprob_tokens": [ + -3.895601749420166 + ], + "logprob_reward": -0.02060169354081154, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060169354081154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12649193406105042, + 0.12680745124816895, + 0.09420573711395264, + 0.14717398583889008, + 0.0723547637462616, + 0.28277647495269775, + 0.04218936339020729, + 0.021622030064463615, + 0.06788738816976547, + 0.018490897491574287 + ], + "credit_map": [ + 0.013194640167057514, + 0.009679111652076244, + 0.015117344446480274, + 0.00963837094604969, + 0.02232731319963932, + 0.02326297014951706, + 0.017426934093236923, + 0.015845248475670815, + 0.021175911650061607, + 0.006538833025842905, + 0.01262610498815775, + 0.00780370132997632, + 0.018897296860814095, + 0.01752249151468277, + 0.011046258732676506, + 0.031196855008602142, + 0.020998438820242882, + 0.014619596302509308, + 0.007177163381129503, + 0.010950782336294651, + 0.0044014714658260345, + 0.011613243259489536, + 0.006483364384621382, + 0.017961673438549042, + 0.016166139394044876, + 0.014341931790113449, + 0.01517731137573719, + 0.014437107369303703, + 0.01267290860414505, + 0.02325044944882393, + 0.013721704483032227, + 0.037406422197818756, + 0.009263437241315842, + 0.007458273321390152, + 0.007987553253769875, + 0.0061636012978851795, + 0.013288874179124832, + 0.007842272520065308, + 0.010393655858933926, + 0.00995709653943777, + 0.0551714189350605, + 0.06688303500413895, + 0.019288456067442894, + 0.02336989715695381, + 0.026834947988390923, + 0.04230606555938721, + 0.019351987168192863, + 0.029570680111646652, + 0.0033140787854790688, + 0.0031281032133847475, + 0.005656789056956768, + 0.0056497398763895035, + 0.005038885399699211, + 0.006328146904706955, + 0.007054968737065792, + 0.006018651649355888, + 0.001676613581366837, + 0.0019278691615909338, + 0.0032765651121735573, + 0.0027055440004915, + 0.0020744255743920803, + 0.003924875520169735, + 0.0030676915775984526, + 0.0029684456530958414, + 0.0057595218531787395, + 0.011456149630248547, + 0.004554068669676781, + 0.00536816893145442, + 0.007052951492369175, + 0.014942198991775513, + 0.007117376197129488, + 0.011636953800916672, + 0.001988820033147931, + 0.0023853557649999857, + 0.002168581821024418, + 0.0018701126100495458, + 0.0015290804440155625, + 0.0026523652486503124, + 0.003175470046699047, + 0.002721113385632634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.02060169354081154 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632143795490265 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632143795490265, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632143795490265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12341264635324478, + 0.0705733597278595, + 0.23526860773563385, + 0.1092635914683342, + 0.08366623520851135, + 0.17239615321159363, + 0.10261209309101105, + 0.03840767964720726, + 0.04953627660870552, + 0.014863280579447746 + ], + "credit_map": [ + 0.011167332530021667, + 0.008962764404714108, + 0.01569177769124508, + 0.01103587169200182, + 0.023334605619311333, + 0.0200364850461483, + 0.017688587307929993, + 0.015495246276259422, + 0.01234052050858736, + 0.003800180507823825, + 0.00845574401319027, + 0.007363274227827787, + 0.005525246262550354, + 0.008543692529201508, + 0.007549352943897247, + 0.01699535734951496, + 0.0795740857720375, + 0.0356193445622921, + 0.031195197254419327, + 0.01233665645122528, + 0.008589794859290123, + 0.017879027873277664, + 0.011166291311383247, + 0.03890823945403099, + 0.010650175623595715, + 0.008351754397153854, + 0.009548021480441093, + 0.010198875330388546, + 0.01252555288374424, + 0.019941158592700958, + 0.011338043957948685, + 0.0267100278288126, + 0.007718980778008699, + 0.006536552682518959, + 0.01198109146207571, + 0.007727883290499449, + 0.014272656291723251, + 0.010869907215237617, + 0.010537528432905674, + 0.014021639712154865, + 0.017172036692500114, + 0.021743224933743477, + 0.01642502471804619, + 0.018685059621930122, + 0.02559647709131241, + 0.03541889041662216, + 0.016057783737778664, + 0.021297672763466835, + 0.008970694616436958, + 0.005992743652313948, + 0.017043519765138626, + 0.012269005179405212, + 0.012781588360667229, + 0.014621098525822163, + 0.016431108117103577, + 0.014502347446978092, + 0.003420812077820301, + 0.0022363041061908007, + 0.00595058361068368, + 0.004879998508840799, + 0.0035531988833099604, + 0.005550611764192581, + 0.004881895147264004, + 0.007934280671179295, + 0.004470770712941885, + 0.007869899272918701, + 0.004127924796193838, + 0.004218701273202896, + 0.004780552815645933, + 0.009468638338148594, + 0.005612489767372608, + 0.0089873056858778, + 0.001344788004644215, + 0.0012065540067851543, + 0.002602354856207967, + 0.0015916046686470509, + 0.0017080025281757116, + 0.0021551991812884808, + 0.0021845432929694653, + 0.0020702348556369543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.007632143795490265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632735185325146 + ], + "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": [ + 0.10650309175252914, + 0.04712527245283127, + 0.23240713775157928, + 0.05637994036078453, + 0.151155486702919, + 0.11700336635112762, + 0.11674007028341293, + 0.047004830092191696, + 0.08910631388425827, + 0.03657453879714012 + ], + "credit_map": [ + 0.006705224048346281, + 0.007073758635669947, + 0.010672706179320812, + 0.01108185388147831, + 0.02098175883293152, + 0.01923999935388565, + 0.019987307488918304, + 0.010760488919913769, + 0.008135329000651836, + 0.002089395420625806, + 0.004037051927298307, + 0.004000949207693338, + 0.0020747531671077013, + 0.006737734191119671, + 0.004690810572355986, + 0.015359248965978622, + 0.07829821854829788, + 0.03868540748953819, + 0.02380201779305935, + 0.013711459934711456, + 0.00735149160027504, + 0.017883582040667534, + 0.009503182955086231, + 0.043171778321266174, + 0.004225308541208506, + 0.004657127428799868, + 0.007458005100488663, + 0.00834262277930975, + 0.007225329987704754, + 0.007092151790857315, + 0.005065221805125475, + 0.01231417153030634, + 0.013703751377761364, + 0.011618485674262047, + 0.01578894816339016, + 0.011021433398127556, + 0.02899102121591568, + 0.019853204488754272, + 0.022326523438096046, + 0.02785211242735386, + 0.007907958701252937, + 0.00865630991756916, + 0.015421644784510136, + 0.012512271292507648, + 0.017507188022136688, + 0.02620375156402588, + 0.012419288046658039, + 0.016374951228499413, + 0.010983708314597607, + 0.007606753148138523, + 0.019891543313860893, + 0.013179085217416286, + 0.01370956189930439, + 0.01677636429667473, + 0.01488788053393364, + 0.019705170765519142, + 0.0036452854983508587, + 0.0027680303901433945, + 0.011488066986203194, + 0.005657614674419165, + 0.004388840403407812, + 0.005958694498986006, + 0.00554727902635932, + 0.00755101814866066, + 0.006475352682173252, + 0.01420456450432539, + 0.005639852490276098, + 0.006144339218735695, + 0.00837485771626234, + 0.022076543420553207, + 0.008401301689445972, + 0.017789499834179878, + 0.003932490944862366, + 0.004842991009354591, + 0.00530253117904067, + 0.004166044294834137, + 0.002855750499293208, + 0.00555375125259161, + 0.004439240787178278, + 0.0054817344062030315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.007632735185325146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up meticulously depicts a single bottle of red wine, its deep green glass framed by glistening droplets of condensation. The bottle stands perfectly still on the perfectly polished, reflective surface of a dark mahogany table. A single, warm, low-key light source from the side outlines the bottle, creating sharp highlights in the droplets and casting long, deep shadows, creating an atmosphere of tense silence. Suddenly, without cause, the bottle topples sideways. It falls onto the tabletop with a sharp crack, instantly shattering in a dramatic explosion of movement. As the crimson wine spurts out, dark, glistening shards of glass scatter outward in a spectacular, viscous splash, its droplets arcing through the air before crashing and rapidly spreading across the polished wood, staining its perfect surface and chaotically reflecting warm light in dark, sprawling pools.", + "seed": 27712, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_0/index_1319_seed_27712_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": -0.11714077815413475, + "vlm_logprob": -0.11714077815413475, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.11714077815413475, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4805816411972046 + ], + "opposite_logprob_tokens": [ + -0.9805816411972046 + ], + "logprob_reward": -0.4805816411972046, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4805816411972046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1348172426223755, + 0.07693358510732651, + 0.13047076761722565, + 0.08662057667970657, + 0.13191421329975128, + 0.07840994745492935, + 0.10132361203432083, + 0.10104432702064514, + 0.07908646762371063, + 0.07937929034233093 + ], + "credit_map": [ + 0.012810824438929558, + 0.016006121411919594, + 0.021187132224440575, + 0.021468833088874817, + 0.015283828601241112, + 0.019920876249670982, + 0.013220187276601791, + 0.014919448643922806, + 0.008277267217636108, + 0.010222558863461018, + 0.01083318516612053, + 0.00936738308519125, + 0.008297519758343697, + 0.00749034620821476, + 0.011079016141593456, + 0.01136630680412054, + 0.010004865005612373, + 0.012298853136599064, + 0.021743474528193474, + 0.032665859907865524, + 0.009912646375596523, + 0.017979687079787254, + 0.012213765643537045, + 0.013651615008711815, + 0.007365883328020573, + 0.008113664574921131, + 0.01123394537717104, + 0.017713677138090134, + 0.00807272270321846, + 0.011578865349292755, + 0.009280181489884853, + 0.013261640444397926, + 0.014241646975278854, + 0.0057079182006418705, + 0.025226520374417305, + 0.03485117480158806, + 0.004637355916202068, + 0.01388869620859623, + 0.02035861648619175, + 0.013002278283238411, + 0.009357614442706108, + 0.0048128049820661545, + 0.011117973364889622, + 0.010203372687101364, + 0.008004732429981232, + 0.011859588325023651, + 0.012352374382317066, + 0.010701491497457027, + 0.010393358767032623, + 0.010828032158315182, + 0.013119221664965153, + 0.014485879801213741, + 0.011672752909362316, + 0.012198450975120068, + 0.012695186771452427, + 0.01593072898685932, + 0.008232029154896736, + 0.008473694324493408, + 0.009150687605142593, + 0.01129260752350092, + 0.008390354923903942, + 0.016448985785245895, + 0.012149813584983349, + 0.026906155049800873, + 0.007144802715629339, + 0.013717381283640862, + 0.008454330265522003, + 0.008038907311856747, + 0.01209659781306982, + 0.00924435630440712, + 0.012519692070782185, + 0.007870391942560673, + 0.00671976525336504, + 0.010348082520067692, + 0.008417874574661255, + 0.006524404510855675, + 0.01646467298269272, + 0.009586266241967678, + 0.014532371424138546, + 0.006785844452679157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.4805816411972046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055010486394166946 + ], + "opposite_logprob_tokens": [ + -2.9300105571746826 + ], + "logprob_reward": -0.055010486394166946, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055010486394166946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06769947707653046, + 0.07439825683832169, + 0.08116631209850311, + 0.12260040640830994, + 0.08060932159423828, + 0.062194209545850754, + 0.05122256278991699, + 0.34505748748779297, + 0.0885641872882843, + 0.02648785337805748 + ], + "credit_map": [ + 0.006460282020270824, + 0.008365783840417862, + 0.010424301959574223, + 0.009749505668878555, + 0.0058191255666315556, + 0.010973852127790451, + 0.009675746783614159, + 0.006230877712368965, + 0.005610250402241945, + 0.01615847274661064, + 0.009709183126688004, + 0.008945995941758156, + 0.008101743645966053, + 0.007298543583601713, + 0.010069461539387703, + 0.008504610508680344, + 0.00594376353546977, + 0.012191087007522583, + 0.01020522229373455, + 0.014228075742721558, + 0.010880942456424236, + 0.010440331883728504, + 0.009066857397556305, + 0.008210030384361744, + 0.010173523798584938, + 0.009873916395008564, + 0.01649370975792408, + 0.022398972883820534, + 0.011460424400866032, + 0.01949557475745678, + 0.01384233683347702, + 0.018861941993236542, + 0.010816809721291065, + 0.004060974810272455, + 0.018346650525927544, + 0.01583907939493656, + 0.004117123782634735, + 0.008220590651035309, + 0.010509476996958256, + 0.008698610588908195, + 0.00675754901021719, + 0.004443809390068054, + 0.007427203468978405, + 0.009594623930752277, + 0.0061372979544103146, + 0.012748604640364647, + 0.009316536597907543, + 0.005768583621829748, + 0.005280644167214632, + 0.004159234464168549, + 0.007776075974106789, + 0.007405108772218227, + 0.007408935576677322, + 0.005272045731544495, + 0.005981072783470154, + 0.007939445786178112, + 0.010913397185504436, + 0.00936833955347538, + 0.010089743882417679, + 0.014559844508767128, + 0.015784814953804016, + 0.1319168508052826, + 0.020530544221401215, + 0.13189393281936646, + 0.007815923541784286, + 0.01347216684371233, + 0.013034990057349205, + 0.011378257535398006, + 0.011379270814359188, + 0.010400594212114811, + 0.00983211025595665, + 0.011250872164964676, + 0.002081226324662566, + 0.0018784389831125736, + 0.0037588917184621096, + 0.003949571400880814, + 0.0028676053043454885, + 0.0030514858663082123, + 0.005077109206467867, + 0.003823523875325918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.055010486394166946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812059387564659 + ], + "opposite_logprob_tokens": [ + -3.2881205081939697 + ], + "logprob_reward": -0.03812059387564659, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812059387564659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1527380347251892, + 0.14489522576332092, + 0.06847039610147476, + 0.063226617872715, + 0.12901200354099274, + 0.07931669056415558, + 0.05282358080148697, + 0.1985551118850708, + 0.08448503166437149, + 0.026477333158254623 + ], + "credit_map": [ + 0.010955236852169037, + 0.018913080915808678, + 0.018862999975681305, + 0.027701200917363167, + 0.0162659864872694, + 0.016573913395404816, + 0.022042378783226013, + 0.021423235535621643, + 0.009866282343864441, + 0.02158195897936821, + 0.01712857559323311, + 0.020034633576869965, + 0.014598537236452103, + 0.014456753619015217, + 0.023725146427750587, + 0.023503342643380165, + 0.007911372929811478, + 0.005169764161109924, + 0.017057988792657852, + 0.009041315875947475, + 0.004022497683763504, + 0.01089178305119276, + 0.004776111338287592, + 0.009599562734365463, + 0.0069623226299881935, + 0.005543679930269718, + 0.010037917643785477, + 0.011932156048715115, + 0.00583652313798666, + 0.006952949799597263, + 0.0064359502866864204, + 0.009525121189653873, + 0.011568954214453697, + 0.007742896676063538, + 0.00977842602878809, + 0.03274093568325043, + 0.009914162568747997, + 0.024026094004511833, + 0.0196702741086483, + 0.013570250943303108, + 0.008658921346068382, + 0.004201375879347324, + 0.012151946313679218, + 0.011684012599289417, + 0.005774171557277441, + 0.010678731836378574, + 0.015572112053632736, + 0.0105954110622406, + 0.004620302002876997, + 0.004091204144060612, + 0.007283744402229786, + 0.008487792685627937, + 0.006425459869205952, + 0.00633781123906374, + 0.006654897704720497, + 0.00892237015068531, + 0.0059784953482449055, + 0.008038921281695366, + 0.0069474768824875355, + 0.00845783669501543, + 0.008401526138186455, + 0.07229599356651306, + 0.01258243340998888, + 0.07585245370864868, + 0.009520003572106361, + 0.011140579357743263, + 0.013978756032884121, + 0.00868156272917986, + 0.011753955855965614, + 0.010206309147179127, + 0.008816897869110107, + 0.01038697361946106, + 0.0015779677778482437, + 0.0019285870948806405, + 0.004509639926254749, + 0.002689455868676305, + 0.0036943848244845867, + 0.003149644238874316, + 0.004819562658667564, + 0.004108089953660965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.03812059387564659 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254743155092001 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254743155092001, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254743155092001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10389751940965652, + 0.100899338722229, + 0.1924206167459488, + 0.12379109859466553, + 0.12672822177410126, + 0.125162735581398, + 0.04785879701375961, + 0.10256863385438919, + 0.049099408090114594, + 0.027573667466640472 + ], + "credit_map": [ + 0.008539345115423203, + 0.013327808119356632, + 0.016027672216296196, + 0.01417683344334364, + 0.015155116096138954, + 0.01306219957768917, + 0.010275527834892273, + 0.013333016075193882, + 0.009905501268804073, + 0.013337834738194942, + 0.01375862117856741, + 0.0108322249725461, + 0.008637028746306896, + 0.00886994507163763, + 0.016975602135062218, + 0.01858258806169033, + 0.01241325680166483, + 0.020164866000413895, + 0.018893998116254807, + 0.04978729039430618, + 0.02224692329764366, + 0.026402687653899193, + 0.019802721217274666, + 0.022708885371685028, + 0.010883139446377754, + 0.010722336359322071, + 0.018244396895170212, + 0.02220148779451847, + 0.01088608056306839, + 0.017698850482702255, + 0.014055679552257061, + 0.019099127501249313, + 0.015229703858494759, + 0.010190603323280811, + 0.020389962941408157, + 0.01967480778694153, + 0.007656200788915157, + 0.01865420863032341, + 0.012238978408277035, + 0.022693756967782974, + 0.00993837509304285, + 0.007637444417923689, + 0.013861782848834991, + 0.02170981653034687, + 0.010826905257999897, + 0.01714036799967289, + 0.022655576467514038, + 0.021392464637756348, + 0.003971697762608528, + 0.0032912527676671743, + 0.005958166904747486, + 0.008386868052184582, + 0.004977320786565542, + 0.005543690174818039, + 0.006044203881174326, + 0.009685596451163292, + 0.003930749371647835, + 0.008335838094353676, + 0.007258908357471228, + 0.006271468475461006, + 0.006494282744824886, + 0.03139447793364525, + 0.0068129017017781734, + 0.032069992274045944, + 0.0044248769991099834, + 0.0065514640882611275, + 0.00966752227395773, + 0.00630349013954401, + 0.005589130334556103, + 0.005582785699516535, + 0.0036697634495794773, + 0.007310372311621904, + 0.0021496445406228304, + 0.0026561894919723272, + 0.0031634632032364607, + 0.002977135591208935, + 0.005128893535584211, + 0.0031107235699892044, + 0.005916661582887173, + 0.002470956416800618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.005254743155092001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736426148563623 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736426148563623, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736426148563623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04437152296304703, + 0.09017002582550049, + 0.08517806977033615, + 0.14975936710834503, + 0.3201742172241211, + 0.07152058929204941, + 0.05203236639499664, + 0.05621419474482536, + 0.05370914936065674, + 0.07687050104141235 + ], + "credit_map": [ + 0.003371953498572111, + 0.004477208945900202, + 0.006465323735028505, + 0.00752860214561224, + 0.0030386403668671846, + 0.0047199358232319355, + 0.00788324885070324, + 0.006886606570333242, + 0.005745782051235437, + 0.012014229781925678, + 0.012233675457537174, + 0.01316675916314125, + 0.005555650219321251, + 0.008570674806833267, + 0.01653774455189705, + 0.016345512121915817, + 0.007328818552196026, + 0.006739612203091383, + 0.015515636652708054, + 0.021242931485176086, + 0.006365880370140076, + 0.011716639623045921, + 0.007663588039577007, + 0.008604954928159714, + 0.014252835884690285, + 0.015385240316390991, + 0.018230335786938667, + 0.026356911286711693, + 0.013680176809430122, + 0.02225368097424507, + 0.015791622921824455, + 0.023808563128113747, + 0.03493773564696312, + 0.012298732995986938, + 0.06523953378200531, + 0.10079275816679001, + 0.007462288718670607, + 0.02295040898025036, + 0.05748588591814041, + 0.019006863236427307, + 0.008266329765319824, + 0.005557109601795673, + 0.011016401462256908, + 0.010234865359961987, + 0.005346022546291351, + 0.010208243504166603, + 0.0111371586099267, + 0.00975445844233036, + 0.005852972622960806, + 0.005283629056066275, + 0.006386160850524902, + 0.007137923501431942, + 0.006686750333756208, + 0.007100333925336599, + 0.006404642015695572, + 0.007179953623563051, + 0.008498298935592175, + 0.005193799268454313, + 0.004806270357221365, + 0.004566781222820282, + 0.0048027075827121735, + 0.009113552048802376, + 0.008690034039318562, + 0.010542754083871841, + 0.005802565719932318, + 0.00708807585760951, + 0.009037408977746964, + 0.004878892097622156, + 0.006457380019128323, + 0.0071883779019117355, + 0.005644391290843487, + 0.007612057030200958, + 0.005847449880093336, + 0.008234684355556965, + 0.00752375228330493, + 0.006044781766831875, + 0.014542898163199425, + 0.009313415735960007, + 0.017519375309348106, + 0.007844138890504837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.006736426148563623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up meticulously depicts a single bottle of red wine, its deep green glass framed by glistening droplets of condensation. The bottle stands perfectly still on the perfectly polished, reflective surface of a dark mahogany table. A single, warm, low-key light source from the side outlines the bottle, creating sharp highlights in the droplets and casting long, deep shadows, creating an atmosphere of tense silence. Suddenly, without cause, the bottle topples sideways. It falls onto the tabletop with a sharp crack, instantly shattering in a dramatic explosion of movement. As the crimson wine spurts out, dark, glistening shards of glass scatter outward in a spectacular, viscous splash, its droplets arcing through the air before crashing and rapidly spreading across the polished wood, staining its perfect surface and chaotically reflecting warm light in dark, sprawling pools.", + "seed": 27713, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_0/index_1319_seed_27713_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": -0.018382193613797428, + "vlm_logprob": -0.018382193613797428, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.018382193613797428, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027968091890215874 + ], + "opposite_logprob_tokens": [ + -3.777968168258667 + ], + "logprob_reward": -0.027968091890215874, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027968091890215874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06781870871782303, + 0.13622130453586578, + 0.06854787468910217, + 0.15019315481185913, + 0.145566925406456, + 0.0529373399913311, + 0.08556213974952698, + 0.1815018504858017, + 0.051398053765296936, + 0.06025268882513046 + ], + "credit_map": [ + 0.008731826208531857, + 0.00943476427346468, + 0.008781886659562588, + 0.0070562902837991714, + 0.010293318890035152, + 0.007867414504289627, + 0.006294765044003725, + 0.009358445182442665, + 0.018252423033118248, + 0.010913149453699589, + 0.019876597449183464, + 0.02354777231812477, + 0.023312849923968315, + 0.012552511878311634, + 0.013264447450637817, + 0.014501551166176796, + 0.00927730742841959, + 0.0050515420734882355, + 0.01145376916974783, + 0.011026956140995026, + 0.013041486032307148, + 0.006067479494959116, + 0.007789862807840109, + 0.0048394715413451195, + 0.013441910035908222, + 0.010761706158518791, + 0.016534941270947456, + 0.017602713778614998, + 0.010016274638473988, + 0.03597230464220047, + 0.02544121816754341, + 0.020422084257006645, + 0.012368917465209961, + 0.011469298973679543, + 0.04021405056118965, + 0.015158917754888535, + 0.007358820177614689, + 0.009189953096210957, + 0.018977290019392967, + 0.030829671770334244, + 0.0042793783359229565, + 0.006192070432007313, + 0.007489004638046026, + 0.007973921485245228, + 0.008209857158362865, + 0.004006032831966877, + 0.009939874522387981, + 0.0048472038470208645, + 0.008502688258886337, + 0.00816700141876936, + 0.011484244838356972, + 0.011038349010050297, + 0.014090237207710743, + 0.011264811269938946, + 0.010416132397949696, + 0.010598677210509777, + 0.020467106252908707, + 0.028900479897856712, + 0.028339983895421028, + 0.029723554849624634, + 0.018818262964487076, + 0.012861436232924461, + 0.026462169364094734, + 0.015928860753774643, + 0.005814048927277327, + 0.00774559797719121, + 0.006782792042940855, + 0.006459792144596577, + 0.008761914446949959, + 0.004888275172561407, + 0.006233544088900089, + 0.004712086636573076, + 0.007436665706336498, + 0.012123458087444305, + 0.007695099804550409, + 0.006865558680146933, + 0.007075333036482334, + 0.005674006417393684, + 0.0065575772896409035, + 0.006824991200119257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.027968091890215874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018228041008114815 + ], + "opposite_logprob_tokens": [ + -4.018228054046631 + ], + "logprob_reward": -0.018228041008114815, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018228041008114815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19019271433353424, + 0.07363060116767883, + 0.02963193506002426, + 0.08933750540018082, + 0.25849950313568115, + 0.04068243131041527, + 0.14180636405944824, + 0.06018773838877678, + 0.055188775062561035, + 0.06084243580698967 + ], + "credit_map": [ + 0.026091119274497032, + 0.023418359458446503, + 0.024365128949284554, + 0.02143956534564495, + 0.03381175920367241, + 0.02235269360244274, + 0.015521213412284851, + 0.0231928750872612, + 0.011667000129818916, + 0.005878990050405264, + 0.012094184756278992, + 0.015587251633405685, + 0.009276042692363262, + 0.00488138385117054, + 0.00647113798186183, + 0.007774604484438896, + 0.003515847958624363, + 0.0020493273623287678, + 0.005927292164415121, + 0.0030318177305161953, + 0.005945999640971422, + 0.0030786744318902493, + 0.003535240888595581, + 0.0025477332528680563, + 0.008327757008373737, + 0.006875292863696814, + 0.014699475839734077, + 0.00964024756103754, + 0.006246699020266533, + 0.022141005843877792, + 0.012399809435009956, + 0.009007217362523079, + 0.023901749402284622, + 0.018052268773317337, + 0.08329234272241592, + 0.020626232028007507, + 0.011505154892802238, + 0.021378878504037857, + 0.018436552956700325, + 0.061306342482566833, + 0.005641222465783358, + 0.00447063846513629, + 0.0059570553712546825, + 0.0059062703512609005, + 0.005630774889141321, + 0.003203050699084997, + 0.005833309609442949, + 0.004040108993649483, + 0.015731187537312508, + 0.014539854601025581, + 0.020169267430901527, + 0.025424033403396606, + 0.02010539546608925, + 0.011591694317758083, + 0.021753625944256783, + 0.012491312809288502, + 0.0057610138319432735, + 0.007960432209074497, + 0.008788622915744781, + 0.009466639719903469, + 0.00680350698530674, + 0.005104809999465942, + 0.008029687218368053, + 0.008273025043308735, + 0.007290315814316273, + 0.007961143739521503, + 0.007114095147699118, + 0.007414794992655516, + 0.006594337522983551, + 0.006311924196779728, + 0.007447652984410524, + 0.005054512061178684, + 0.009059065952897072, + 0.008385127410292625, + 0.008172967471182346, + 0.007554035168141127, + 0.006390845403075218, + 0.005924917291849852, + 0.007385496515780687, + 0.007969978265464306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.018228041008114815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372011333703995 + ], + "opposite_logprob_tokens": [ + -3.408720016479492 + ], + "logprob_reward": -0.03372011333703995, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372011333703995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05075230076909065, + 0.041703835129737854, + 0.05482093244791031, + 0.10139603167772293, + 0.34457725286483765, + 0.0516151562333107, + 0.1899937391281128, + 0.06336700171232224, + 0.04209273308515549, + 0.059681106358766556 + ], + "credit_map": [ + 0.004580168053507805, + 0.00681554339826107, + 0.007381661329418421, + 0.004298854153603315, + 0.010700319893658161, + 0.005410272162407637, + 0.004239944741129875, + 0.007325530517846346, + 0.0037994866725057364, + 0.003681887174025178, + 0.004827727098017931, + 0.008340936154127121, + 0.006883854977786541, + 0.005204224027693272, + 0.004743325058370829, + 0.004222395364195108, + 0.010990537703037262, + 0.006132563576102257, + 0.009588298387825489, + 0.008413981646299362, + 0.006073962897062302, + 0.003340022638440132, + 0.005993281491100788, + 0.0042882817797362804, + 0.013183798640966415, + 0.00786054227501154, + 0.0214031133800745, + 0.007988913916051388, + 0.00547861447557807, + 0.02053077705204487, + 0.014502517879009247, + 0.010447749868035316, + 0.031956858932971954, + 0.02348698303103447, + 0.10768930613994598, + 0.02868405543267727, + 0.02006923034787178, + 0.031586870551109314, + 0.02205144800245762, + 0.07905246317386627, + 0.004856922198086977, + 0.00575121957808733, + 0.007153565064072609, + 0.010596591979265213, + 0.006123895291239023, + 0.004194850567728281, + 0.005504278931766748, + 0.007433828432112932, + 0.018934348598122597, + 0.018650295212864876, + 0.02709539793431759, + 0.028231902047991753, + 0.03472250699996948, + 0.024840161204338074, + 0.02236800454556942, + 0.015151106752455235, + 0.006275552790611982, + 0.011715096421539783, + 0.010172827169299126, + 0.009436705149710178, + 0.009951556101441383, + 0.004063994158059359, + 0.006648263894021511, + 0.005102997645735741, + 0.0046707745641469955, + 0.005408140365034342, + 0.005117691587656736, + 0.0054631466045975685, + 0.0065567283891141415, + 0.004598041530698538, + 0.005511192604899406, + 0.004767012316733599, + 0.007800797000527382, + 0.012609370984137058, + 0.0071503822691738605, + 0.007319209165871143, + 0.0067690396681427956, + 0.006629429291933775, + 0.005989696364849806, + 0.005413180682808161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.03372011333703995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739859934896231 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006739859934896231, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739859934896231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07411754131317139, + 0.033086344599723816, + 0.03834526985883713, + 0.10625911504030228, + 0.13881798088550568, + 0.06739012151956558, + 0.2031734138727188, + 0.06182872876524925, + 0.13358032703399658, + 0.14340122044086456 + ], + "credit_map": [ + 0.010678467340767384, + 0.008625883609056473, + 0.01112446654587984, + 0.006102060899138451, + 0.01325282733887434, + 0.00868358463048935, + 0.005731228739023209, + 0.009919017553329468, + 0.003975181374698877, + 0.002634436823427677, + 0.003506200388073921, + 0.005373913794755936, + 0.006516959983855486, + 0.004182690288871527, + 0.0034553336445242167, + 0.003441627137362957, + 0.008461149409413338, + 0.005178310442715883, + 0.007405501324683428, + 0.003230586415156722, + 0.004601212218403816, + 0.0023891585879027843, + 0.0035957980435341597, + 0.0034835534170269966, + 0.013138809241354465, + 0.008840247988700867, + 0.022894354537129402, + 0.013599288649857044, + 0.00789108406752348, + 0.010864794254302979, + 0.01661149226129055, + 0.012419032864272594, + 0.016443658620119095, + 0.01230941154062748, + 0.03309907391667366, + 0.017846107482910156, + 0.008965279906988144, + 0.010783709585666656, + 0.013992086984217167, + 0.025378653779625893, + 0.006822300609201193, + 0.008179182186722755, + 0.008318750187754631, + 0.010802694596350193, + 0.01088711991906166, + 0.0054444195702672005, + 0.010505306534469128, + 0.0064303395338356495, + 0.02141587622463703, + 0.019764890894293785, + 0.021804627031087875, + 0.034607287496328354, + 0.0373813696205616, + 0.017582934349775314, + 0.030182896181941032, + 0.020433522760868073, + 0.005960890091955662, + 0.011451832950115204, + 0.009213702753186226, + 0.00822911225259304, + 0.007288269232958555, + 0.00560410600155592, + 0.007877705618739128, + 0.00620310427621007, + 0.01542388554662466, + 0.017861677333712578, + 0.019970983266830444, + 0.018904725089669228, + 0.013637865893542767, + 0.01633521169424057, + 0.01592109352350235, + 0.015524886548519135, + 0.019785264506936073, + 0.02771596983075142, + 0.01989683508872986, + 0.01953408308327198, + 0.012875489890575409, + 0.013523705303668976, + 0.015156850218772888, + 0.014913013204932213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.006739859934896231 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level close-up of a single bottle of red wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of deep green glass with glistening condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle stand still on a dark mahogany table before suddenly toppling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle shatter into dark, glistening shards with crimson wine spurting out in a splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the spilled wine droplets arc through the air and spread across the table, staining the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052548618987202644 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.0052548618987202644, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052548618987202644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05962081626057625, + 0.020309703424572945, + 0.029930906370282173, + 0.041262514889240265, + 0.2134883552789688, + 0.0424986369907856, + 0.06870649755001068, + 0.08894795179367065, + 0.17728275060653687, + 0.25795188546180725 + ], + "credit_map": [ + 0.006377066019922495, + 0.006428311578929424, + 0.007943917997181416, + 0.010061489418148994, + 0.01139762345701456, + 0.004744706675410271, + 0.005672350991517305, + 0.006995347794145346, + 0.002409360371530056, + 0.0019265190931037068, + 0.0017503012204542756, + 0.004728725180029869, + 0.0035003316588699818, + 0.0018070571823045611, + 0.002026280155405402, + 0.002161128679290414, + 0.007431648205965757, + 0.0045362235978245735, + 0.004972024355083704, + 0.002420160686597228, + 0.003979656379669905, + 0.0018452868098393083, + 0.0025666356086730957, + 0.002179270377382636, + 0.0038812498096376657, + 0.0023485322017222643, + 0.0052445619367063046, + 0.005031748674809933, + 0.002901247702538967, + 0.00934048742055893, + 0.006911982782185078, + 0.005602704826742411, + 0.021412063390016556, + 0.010465974919497967, + 0.08142201602458954, + 0.01851111277937889, + 0.005457211285829544, + 0.007277397438883781, + 0.009871507994830608, + 0.05907106772065163, + 0.005068373400717974, + 0.004810885060578585, + 0.006101534236222506, + 0.00825593899935484, + 0.004800362978130579, + 0.0030900745186954737, + 0.005881147924810648, + 0.004490318242460489, + 0.0064364210702478886, + 0.007646297104656696, + 0.009502422995865345, + 0.011603494174778461, + 0.007797168102115393, + 0.007184817921370268, + 0.01113039068877697, + 0.007405484095215797, + 0.007587198633700609, + 0.012352990917861462, + 0.015832548961043358, + 0.013143678195774555, + 0.00985736958682537, + 0.005563603714108467, + 0.014995680190622807, + 0.009614885784685612, + 0.021435102447867393, + 0.021294280886650085, + 0.024850407615303993, + 0.02314937114715576, + 0.019158432260155678, + 0.020246323198080063, + 0.024973925203084946, + 0.022174907848238945, + 0.031463585793972015, + 0.05931651592254639, + 0.03662455454468727, + 0.029574571177363396, + 0.025145074352622032, + 0.028297865763306618, + 0.022882936522364616, + 0.024646762758493423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.0052548618987202644 + } + ], + "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/0000025/group_1/index_tensor([6214, 6214])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_1/index_tensor([6214, 6214])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..90ba78776178012ef17bc2a0dc83f3bd47ba64d7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_1/index_tensor([6214, 6214])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, shot from a shaky shoulder-mounted camera at a tense, tilted angle, locks the viewer's gaze squarely on a pale finger. Illuminated by a beam of cool white light from above, the pores and subtle texture of the skin are clearly visible. At the center of the fingertip, in a tiny, needle-like wound, a plump, crystalline drop of bright red blood slowly gathers and trembles, its surface reflecting a dazzling light. As the camera gently pans, a steady yet slightly trembling hand carefully brings the end of a white plastic test strip close to the falling drop. At the moment when the drop reaches its limit under its own weight, the tip of the strip, with its yellow capillary zone, precisely touches the edge, preparing to draw it in. The entire scene, imbued with a raw, documentary quality, completely fixates the viewer's attention on this crucial and suspenseful moment of medical testing, creating a breathtaking atmosphere of tension.", + "seed": 6183, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_1/index_6214_seed_6183_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": -0.15366472194436936, + "vlm_logprob": -0.15366472194436936, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.15366472194436936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013716607354581356 + ], + "opposite_logprob_tokens": [ + -6.763716697692871 + ], + "logprob_reward": -0.013716607354581356, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.013716607354581356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07851752638816833, + 0.12278049439191818, + 0.03902085870504379, + 0.030852695927023888, + 0.03539225086569786, + 0.0530327744781971, + 0.07337344437837601, + 0.0777498111128807, + 0.11253122240304947, + 0.3767489790916443 + ], + "credit_map": [ + 0.007829458452761173, + 0.00819968618452549, + 0.0075953658670187, + 0.008763061836361885, + 0.010661606676876545, + 0.01006742101162672, + 0.010820853523910046, + 0.0145800756290555, + 0.004856862593442202, + 0.007299634162336588, + 0.004268460441380739, + 0.010059542022645473, + 0.006204354111105204, + 0.041064001619815826, + 0.009919297881424427, + 0.039108335971832275, + 0.0033000491093844175, + 0.007108163554221392, + 0.0034120359923690557, + 0.007131198886781931, + 0.0037460699677467346, + 0.004658445715904236, + 0.004183137323707342, + 0.00548175722360611, + 0.002530456520617008, + 0.00431758351624012, + 0.0024686502292752266, + 0.004733334295451641, + 0.0031229620799422264, + 0.004089165944606066, + 0.005661710165441036, + 0.003928830847144127, + 0.0026875382754951715, + 0.006059923209249973, + 0.004251490812748671, + 0.006486420053988695, + 0.003506286768242717, + 0.004153333604335785, + 0.0036815539933741093, + 0.004565703682601452, + 0.005067506339401007, + 0.004236581269651651, + 0.008959680795669556, + 0.007452460005879402, + 0.0062630027532577515, + 0.005428472068160772, + 0.0099787013605237, + 0.005646369885653257, + 0.007528099697083235, + 0.007645520847290754, + 0.012603804469108582, + 0.0065429918467998505, + 0.012448745779693127, + 0.005427535157650709, + 0.012843169271945953, + 0.00833357498049736, + 0.00786912627518177, + 0.006637888494879007, + 0.010412251576781273, + 0.008315959013998508, + 0.010503921657800674, + 0.011560875922441483, + 0.013639622367918491, + 0.008810167200863361, + 0.009407668374478817, + 0.005594077054411173, + 0.00773665402084589, + 0.012512898072600365, + 0.005135132931172848, + 0.043729815632104874, + 0.008264873176813126, + 0.02015010640025139, + 0.016721859574317932, + 0.008966575376689434, + 0.008317109197378159, + 0.055048633366823196, + 0.00775166554376483, + 0.16479966044425964, + 0.024299461394548416, + 0.09084401279687881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.013716607354581356 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762065052986145 + ], + "opposite_logprob_tokens": [ + -0.8262065052986145 + ], + "logprob_reward": -0.5762065052986145, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762065052986145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11918719857931137, + 0.10269784927368164, + 0.032314009964466095, + 0.019039088860154152, + 0.031225915998220444, + 0.03287385776638985, + 0.010232186876237392, + 0.06425049155950546, + 0.2020963430404663, + 0.38608303666114807 + ], + "credit_map": [ + 0.010762394405901432, + 0.01075852382928133, + 0.010888963006436825, + 0.012213321402668953, + 0.015279094688594341, + 0.018841451033949852, + 0.01466295775026083, + 0.025780491530895233, + 0.0038737573195248842, + 0.008236042223870754, + 0.0038829331751912832, + 0.007305361796170473, + 0.00573930935934186, + 0.03346360847353935, + 0.007540826685726643, + 0.03265601024031639, + 0.0026913306210190058, + 0.0029848723206669092, + 0.0027964694891124964, + 0.004092688672244549, + 0.0038249408826231956, + 0.004580172710120678, + 0.0073980726301670074, + 0.003945461008697748, + 0.0016549549764022231, + 0.0029569563921540976, + 0.0022561969235539436, + 0.003079524962231517, + 0.0018346593715250492, + 0.00196415139362216, + 0.0030861797276884317, + 0.0022064652293920517, + 0.0032029952853918076, + 0.004136807285249233, + 0.003912327811121941, + 0.004296652507036924, + 0.002801966853439808, + 0.002349300542846322, + 0.007104543037712574, + 0.003421324072405696, + 0.002214105101302266, + 0.0037851298693567514, + 0.004173842258751392, + 0.0033087257761508226, + 0.006273397710174322, + 0.004564977250993252, + 0.005016091745346785, + 0.0035375880543142557, + 0.0012149761896580458, + 0.00086799735436216, + 0.0013985083205625415, + 0.0010662067215889692, + 0.0012434907257556915, + 0.0009674821048974991, + 0.002155338181182742, + 0.001318186754360795, + 0.011870337650179863, + 0.0070124720223248005, + 0.005808739457279444, + 0.005078530870378017, + 0.0029840562492609024, + 0.020162148401141167, + 0.0061881705187261105, + 0.005146035924553871, + 0.01836620457470417, + 0.010873951017856598, + 0.02947888895869255, + 0.07537204772233963, + 0.0058752321638166904, + 0.02159946598112583, + 0.01752655766904354, + 0.02300400100648403, + 0.027704065665602684, + 0.012302464805543423, + 0.010074368678033352, + 0.03643101453781128, + 0.009309336543083191, + 0.16991488635540009, + 0.027098398655653, + 0.09324850887060165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.5762065052986145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019519339548423886 + ], + "opposite_logprob_tokens": [ + -6.251952171325684 + ], + "logprob_reward": -0.0019519339548423886, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019519339548423886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03674907237291336, + 0.17402653396129608, + 0.059834133833646774, + 0.09626095741987228, + 0.06112166494131088, + 0.04545402154326439, + 0.03277474269270897, + 0.06740405410528183, + 0.07773001492023468, + 0.34864482283592224 + ], + "credit_map": [ + 0.003384753130376339, + 0.005017174873501062, + 0.003079487243667245, + 0.006111427675932646, + 0.004796851892024279, + 0.004906062036752701, + 0.004831862170249224, + 0.004621448926627636, + 0.009349002502858639, + 0.02351769246160984, + 0.009352508001029491, + 0.018375791609287262, + 0.012164128944277763, + 0.04327116906642914, + 0.015035778284072876, + 0.04296044260263443, + 0.003972150385379791, + 0.006245063152164221, + 0.005085291340947151, + 0.007189895026385784, + 0.005077796522527933, + 0.009805445559322834, + 0.012504800222814083, + 0.009953681379556656, + 0.010349873453378677, + 0.013592605479061604, + 0.007919023744761944, + 0.013984078541398048, + 0.007993001490831375, + 0.011414650827646255, + 0.019362177699804306, + 0.011645538732409477, + 0.004579512402415276, + 0.009815357625484467, + 0.0066749947145581245, + 0.010353406891226768, + 0.0050322567112743855, + 0.006448502652347088, + 0.010607320815324783, + 0.0076103052124381065, + 0.0030918028205633163, + 0.007127256598323584, + 0.003616554196923971, + 0.0073461951687932014, + 0.010088515467941761, + 0.005485961679369211, + 0.004448921885341406, + 0.004248807672411203, + 0.003322612727060914, + 0.0033153179101645947, + 0.004539635963737965, + 0.0029441313818097115, + 0.005765847861766815, + 0.0029088943265378475, + 0.006237504538148642, + 0.003740795888006687, + 0.014117095619440079, + 0.0051963794976472855, + 0.014466547407209873, + 0.005406459793448448, + 0.0037621629890054464, + 0.012401043437421322, + 0.006560856010764837, + 0.005493502132594585, + 0.005362150724977255, + 0.0048244115896523, + 0.00410598237067461, + 0.02790236286818981, + 0.003517738776281476, + 0.019740363582968712, + 0.006673048250377178, + 0.005603954195976257, + 0.02101992629468441, + 0.009081507101655006, + 0.00853611808270216, + 0.046345219016075134, + 0.007680085953325033, + 0.15463101863861084, + 0.02298295684158802, + 0.07836797088384628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.0019519339548423886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071418300271034 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071418300271034, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071418300271034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037539344280958176, + 0.08892304450273514, + 0.01762472093105316, + 0.03223260119557381, + 0.03895166143774986, + 0.07486709952354431, + 0.05709731578826904, + 0.10308197885751724, + 0.24150484800338745, + 0.3081774115562439 + ], + "credit_map": [ + 0.003376528387889266, + 0.004291611723601818, + 0.005023278295993805, + 0.0048605902120471, + 0.005199007224291563, + 0.004366629756987095, + 0.0048308647237718105, + 0.005590830463916063, + 0.002928949659690261, + 0.005350752733647823, + 0.006050314754247665, + 0.004546007607132196, + 0.0045842318795621395, + 0.029635531827807426, + 0.005874317139387131, + 0.029952937737107277, + 0.0011621172307059169, + 0.0021459232084453106, + 0.001289463136345148, + 0.001965534407645464, + 0.0017330952687188983, + 0.0025493805296719074, + 0.004238917957991362, + 0.0025402901228517294, + 0.0021435359958559275, + 0.0057022301480174065, + 0.0031699545215815306, + 0.005394656676799059, + 0.003886475693434477, + 0.0030150366947054863, + 0.005562122445553541, + 0.0033585892524570227, + 0.003819195320829749, + 0.005269585642963648, + 0.004353709053248167, + 0.006111879833042622, + 0.0036809209268540144, + 0.004921936430037022, + 0.006370687857270241, + 0.004423750098794699, + 0.007665260694921017, + 0.009063352830708027, + 0.0074820248410105705, + 0.009253246709704399, + 0.01064754743129015, + 0.01673217862844467, + 0.006900538224726915, + 0.007122947368770838, + 0.006021848879754543, + 0.005904488265514374, + 0.009035193361341953, + 0.004066467750817537, + 0.009265932254493237, + 0.004948912188410759, + 0.008648552931845188, + 0.009205919690430164, + 0.020592499524354935, + 0.011509032920002937, + 0.010906554758548737, + 0.00684366375207901, + 0.006291228346526623, + 0.030972948297858238, + 0.0076936762779951096, + 0.008272373117506504, + 0.025268523022532463, + 0.01355962734669447, + 0.032539624720811844, + 0.03584258258342743, + 0.007680166512727737, + 0.06672697514295578, + 0.018290888518095016, + 0.04159644618630409, + 0.014585472643375397, + 0.007619290146976709, + 0.00807198602706194, + 0.04317326843738556, + 0.007388574071228504, + 0.13301251828670502, + 0.0204776544123888, + 0.07384864240884781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.016071418300271034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603771448135376 + ], + "opposite_logprob_tokens": [ + -1.9103771448135376 + ], + "logprob_reward": -0.1603771448135376, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603771448135376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033607758581638336, + 0.11909212172031403, + 0.02986009605228901, + 0.06155422702431679, + 0.027001768350601196, + 0.03481870889663696, + 0.030171940103173256, + 0.06006008759140968, + 0.06435167044401169, + 0.5394815802574158 + ], + "credit_map": [ + 0.0024259949568659067, + 0.004222468473017216, + 0.003207919653505087, + 0.004711040295660496, + 0.004782982636243105, + 0.004191630054265261, + 0.003506593406200409, + 0.006559127941727638, + 0.003915309440344572, + 0.004947648849338293, + 0.006229660008102655, + 0.004816235043108463, + 0.004740760196000338, + 0.04357942193746567, + 0.007627047598361969, + 0.043236032128334045, + 0.0023882754612714052, + 0.003300715470686555, + 0.0031875993590801954, + 0.004312263801693916, + 0.0029979562386870384, + 0.00420939177274704, + 0.00531367352232337, + 0.00415021600201726, + 0.005565608851611614, + 0.008761463686823845, + 0.0075426409021019936, + 0.00889639463275671, + 0.00567297637462616, + 0.008778352290391922, + 0.007902492769062519, + 0.008434298448264599, + 0.0025495646987110376, + 0.004115702584385872, + 0.004059218801558018, + 0.0032893444877117872, + 0.0032074679620563984, + 0.0029060333035886288, + 0.0039454237557947636, + 0.0029290143866091967, + 0.0024148772936314344, + 0.00452395947650075, + 0.0039566936902701855, + 0.00445530703291297, + 0.008035334758460522, + 0.003649355610832572, + 0.004633495118469, + 0.0031496838200837374, + 0.0033060507848858833, + 0.002791895065456629, + 0.0037486678920686245, + 0.002818159991875291, + 0.004945007152855396, + 0.0031594627071172, + 0.005434196908026934, + 0.003968498203903437, + 0.007759125903248787, + 0.006080825813114643, + 0.006021772511303425, + 0.008006986230611801, + 0.004435246344655752, + 0.015249853953719139, + 0.00788931641727686, + 0.004616961348801851, + 0.004694669973105192, + 0.004008184652775526, + 0.0038345607463270426, + 0.010764865204691887, + 0.0031421161256730556, + 0.02434609830379486, + 0.005680133122950792, + 0.007881037890911102, + 0.029832476750016212, + 0.013450101017951965, + 0.014253510162234306, + 0.06705713272094727, + 0.012834347784519196, + 0.23629719018936157, + 0.035706281661987305, + 0.13005058467388153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.1603771448135376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot from a shaky shoulder-mounted camera at a tense, tilted angle, locks the viewer's gaze squarely on a pale finger. Illuminated by a beam of cool white light from above, the pores and subtle texture of the skin are clearly visible. At the center of the fingertip, in a tiny, needle-like wound, a plump, crystalline drop of bright red blood slowly gathers and trembles, its surface reflecting a dazzling light. As the camera gently pans, a steady yet slightly trembling hand carefully brings the end of a white plastic test strip close to the falling drop. At the moment when the drop reaches its limit under its own weight, the tip of the strip, with its yellow capillary zone, precisely touches the edge, preparing to draw it in. The entire scene, imbued with a raw, documentary quality, completely fixates the viewer's attention on this crucial and suspenseful moment of medical testing, creating a breathtaking atmosphere of tension.", + "seed": 6184, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_1/index_6214_seed_6184_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": -0.12070640597958118, + "vlm_logprob": -0.12070640597958118, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.12070640597958118, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0053537567146122456 + ], + "opposite_logprob_tokens": [ + -7.255353927612305 + ], + "logprob_reward": -0.0053537567146122456, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0053537567146122456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10663039982318878, + 0.05558658391237259, + 0.038230475038290024, + 0.06627807766199112, + 0.10874926298856735, + 0.044721659272909164, + 0.08514382690191269, + 0.04450305923819542, + 0.35647591948509216, + 0.09368076920509338 + ], + "credit_map": [ + 0.011905787512660027, + 0.01586906611919403, + 0.011228266172111034, + 0.01701272837817669, + 0.015159191563725471, + 0.012608037330210209, + 0.011178450658917427, + 0.011668878607451916, + 0.0051746368408203125, + 0.006518821232020855, + 0.004612118937075138, + 0.00971239898353815, + 0.005102275405079126, + 0.0051969666965305805, + 0.012487796135246754, + 0.006781567353755236, + 0.003593655303120613, + 0.0046631330624222755, + 0.003483614418655634, + 0.007138102315366268, + 0.003377753309905529, + 0.003979790024459362, + 0.006898033898323774, + 0.005096387583762407, + 0.00820209551602602, + 0.008259032852947712, + 0.00672103138640523, + 0.011086895130574703, + 0.006085973232984543, + 0.006432192400097847, + 0.01005350798368454, + 0.00943734496831894, + 0.017865119501948357, + 0.017390424385666847, + 0.011737667955458164, + 0.012639456428587437, + 0.01222591195255518, + 0.010286589153110981, + 0.01177710760384798, + 0.014826989732682705, + 0.006326423957943916, + 0.005460201296955347, + 0.004184631630778313, + 0.005750615149736404, + 0.0049527655355632305, + 0.004646355751901865, + 0.006132006179541349, + 0.007268661633133888, + 0.010949334129691124, + 0.015363438986241817, + 0.010369098745286465, + 0.01021223608404398, + 0.007631487678736448, + 0.008702420629560947, + 0.011167400516569614, + 0.01074841246008873, + 0.006307440809905529, + 0.0061179762706160545, + 0.004657225217670202, + 0.00717579759657383, + 0.005410623271018267, + 0.0026755845174193382, + 0.0060354918241500854, + 0.006122917402535677, + 0.04154524952173233, + 0.04225184768438339, + 0.052611589431762695, + 0.046035852283239365, + 0.03657305985689163, + 0.01690894551575184, + 0.0647372230887413, + 0.05581212416291237, + 0.011132299900054932, + 0.006920948624610901, + 0.02283572033047676, + 0.01388509664684534, + 0.005269506946206093, + 0.005538409110158682, + 0.00970735214650631, + 0.018391432240605354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.0053537567146122456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289512634277344 + ], + "opposite_logprob_tokens": [ + -1.0539512634277344 + ], + "logprob_reward": -0.4289512634277344, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289512634277344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11844900250434875, + 0.04359916225075722, + 0.03653565049171448, + 0.05141911655664444, + 0.11574970185756683, + 0.021469498053193092, + 0.025959935039281845, + 0.026873478665947914, + 0.07136832177639008, + 0.48857611417770386 + ], + "credit_map": [ + 0.016287095844745636, + 0.015967581421136856, + 0.01356090884655714, + 0.021183406934142113, + 0.014977974817156792, + 0.012228110805153847, + 0.010928869247436523, + 0.013315059244632721, + 0.005297885742038488, + 0.005256182514131069, + 0.004888491239398718, + 0.007258560974150896, + 0.0046562692150473595, + 0.00357267283834517, + 0.00765568669885397, + 0.0050134179182350636, + 0.005244693718850613, + 0.0041120038367807865, + 0.0032200473360717297, + 0.00642765685915947, + 0.0028079308103770018, + 0.0029658451676368713, + 0.008048041723668575, + 0.003709434298798442, + 0.008067946881055832, + 0.0071783652529120445, + 0.006528479512780905, + 0.008474068716168404, + 0.0034633995965123177, + 0.005058889277279377, + 0.006035838276147842, + 0.00661213556304574, + 0.015959607437253, + 0.011943669058382511, + 0.01671949401497841, + 0.016245661303400993, + 0.009174806997179985, + 0.011576824821531773, + 0.016687719151377678, + 0.01744193211197853, + 0.0024598983582109213, + 0.00310100382193923, + 0.002089300425723195, + 0.003430957905948162, + 0.0020736351143568754, + 0.0021551824174821377, + 0.003132469253614545, + 0.0030270505230873823, + 0.0027071565855294466, + 0.0031399207655340433, + 0.0034311471972614527, + 0.0018851030617952347, + 0.002345198532566428, + 0.006433263886719942, + 0.002577375154942274, + 0.0034407726489007473, + 0.0034149419516324997, + 0.002917235717177391, + 0.0029967373702675104, + 0.004711237736046314, + 0.0039304099045693874, + 0.002498478163033724, + 0.002535960404202342, + 0.0038684795144945383, + 0.010121251456439495, + 0.00943661481142044, + 0.007140284404158592, + 0.008321079425513744, + 0.009323653765022755, + 0.0060517149977386, + 0.01176280528306961, + 0.009210919961333275, + 0.059500306844711304, + 0.037447527050971985, + 0.12363442778587341, + 0.0680534839630127, + 0.026217138394713402, + 0.028267769142985344, + 0.038197051733732224, + 0.10725843161344528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.4289512634277344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011847150744870305 + ], + "opposite_logprob_tokens": [ + -6.751184940338135 + ], + "logprob_reward": -0.0011847150744870305, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011847150744870305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06789208948612213, + 0.061425771564245224, + 0.05459965765476227, + 0.04582063853740692, + 0.09573433548212051, + 0.05714566633105278, + 0.08815402537584305, + 0.0654829815030098, + 0.21416139602661133, + 0.2495834231376648 + ], + "credit_map": [ + 0.010355794802308083, + 0.008727616630494595, + 0.0077672735787928104, + 0.010110635310411453, + 0.00703335041180253, + 0.006401508115231991, + 0.010131786577403545, + 0.0073641217313706875, + 0.007598501164466143, + 0.007076186127960682, + 0.005891609471291304, + 0.011614550836384296, + 0.006265971343964338, + 0.005067311692982912, + 0.010711785405874252, + 0.007199855986982584, + 0.005496291909366846, + 0.0057486966252326965, + 0.0053544361144304276, + 0.009406724944710732, + 0.006507586222141981, + 0.005207419395446777, + 0.011372149921953678, + 0.0055063520558178425, + 0.005296807270497084, + 0.005437243729829788, + 0.005290753673762083, + 0.009579920209944248, + 0.0036454207729548216, + 0.004097234457731247, + 0.006644233129918575, + 0.005829025059938431, + 0.015685154125094414, + 0.014927766285836697, + 0.009952856227755547, + 0.009966133162379265, + 0.010023951530456543, + 0.012675284408032894, + 0.009408949874341488, + 0.013094247318804264, + 0.005986486561596394, + 0.005830995738506317, + 0.005478215869516134, + 0.012287238612771034, + 0.0049254861660301685, + 0.006274793297052383, + 0.00884908065199852, + 0.007513369899243116, + 0.012939638458192348, + 0.012614667415618896, + 0.012869725935161114, + 0.009202822111546993, + 0.007802886888384819, + 0.007630000822246075, + 0.009462558664381504, + 0.015631720423698425, + 0.007178073283284903, + 0.006501591298729181, + 0.007285014726221561, + 0.008217024616897106, + 0.009578162804245949, + 0.006317634601145983, + 0.010830722749233246, + 0.009574765339493752, + 0.020998181775212288, + 0.024942521005868912, + 0.022237643599510193, + 0.029671678319573402, + 0.02037237398326397, + 0.014439327642321587, + 0.05020276829600334, + 0.031296905130147934, + 0.045838989317417145, + 0.03156222030520439, + 0.036003030836582184, + 0.04097381606698036, + 0.021658537909388542, + 0.014318595640361309, + 0.028181595727801323, + 0.031046636402606964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.0011847150744870305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890553817152977 + ], + "opposite_logprob_tokens": [ + -2.464055299758911 + ], + "logprob_reward": -0.0890553817152977, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890553817152977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055030904710292816, + 0.04661950841546059, + 0.0345383957028389, + 0.04787493869662285, + 0.07378505170345306, + 0.03521313518285751, + 0.09421569108963013, + 0.03467004746198654, + 0.2530542314052582, + 0.324998140335083 + ], + "credit_map": [ + 0.005642860196530819, + 0.0065708584152162075, + 0.005889290012419224, + 0.008127198554575443, + 0.008632068522274494, + 0.007173922844231129, + 0.006635279860347509, + 0.006359424442052841, + 0.004384487401694059, + 0.005132667720317841, + 0.003840215504169464, + 0.007478294428437948, + 0.004198740236461163, + 0.004486625548452139, + 0.010954601690173149, + 0.006143876351416111, + 0.003225521417334676, + 0.00392231484875083, + 0.0030142609030008316, + 0.00618163775652647, + 0.0030108685605227947, + 0.0031340753193944693, + 0.008276354521512985, + 0.003773361910134554, + 0.006558767519891262, + 0.006536252796649933, + 0.005581152159720659, + 0.007104580290615559, + 0.0038943712133914232, + 0.004645409993827343, + 0.007243622560054064, + 0.006310782860964537, + 0.010604996234178543, + 0.010617755353450775, + 0.007647762540727854, + 0.011070668697357178, + 0.006944671738892794, + 0.006627827882766724, + 0.010046599432826042, + 0.010224769823253155, + 0.0028654506895691156, + 0.0029705516062676907, + 0.0036107541527599096, + 0.0066145360469818115, + 0.0033645317889750004, + 0.005136047024279833, + 0.00513977138325572, + 0.005511490162461996, + 0.011761689558625221, + 0.01677044667303562, + 0.011897112242877483, + 0.012744113802909851, + 0.007656638976186514, + 0.01034289225935936, + 0.011544584296643734, + 0.011498206295073032, + 0.004797385539859533, + 0.004100210964679718, + 0.004124384373426437, + 0.0040276232175529, + 0.005017891060560942, + 0.002337817568331957, + 0.004541259258985519, + 0.005723473150283098, + 0.027353806421160698, + 0.029944196343421936, + 0.029104797169566154, + 0.03511839359998703, + 0.022322185337543488, + 0.017139414325356483, + 0.058178022503852844, + 0.03389342129230499, + 0.05638772249221802, + 0.04077274724841118, + 0.0503900982439518, + 0.05144106224179268, + 0.025810178369283676, + 0.019949428737163544, + 0.03458476811647415, + 0.04566213861107826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.0890553817152977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898691296577454 + ], + "opposite_logprob_tokens": [ + -2.578986883163452 + ], + "logprob_reward": -0.07898691296577454, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898691296577454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09623628109693527, + 0.058369748294353485, + 0.03905266150832176, + 0.09342016279697418, + 0.16110646724700928, + 0.04247943311929703, + 0.08789055049419403, + 0.06111082807183266, + 0.21249142289161682, + 0.14784245193004608 + ], + "credit_map": [ + 0.01121132168918848, + 0.01046122144907713, + 0.014737588353455067, + 0.017416980117559433, + 0.010440239682793617, + 0.008920359425246716, + 0.011030809953808784, + 0.012017766013741493, + 0.0052364724688231945, + 0.007348574232310057, + 0.00476812943816185, + 0.009339628741145134, + 0.006016384344547987, + 0.007265258580446243, + 0.010431277565658092, + 0.00796402059495449, + 0.004663130734115839, + 0.004522779956459999, + 0.0040010325610637665, + 0.005685172509402037, + 0.004062458872795105, + 0.003875866997987032, + 0.006727438885718584, + 0.005514778662472963, + 0.013644540682435036, + 0.01199035719037056, + 0.010980181396007538, + 0.014580408111214638, + 0.006583401467651129, + 0.009924734011292458, + 0.012727340683341026, + 0.01298918854445219, + 0.029399976134300232, + 0.028615567833185196, + 0.021023642271757126, + 0.016063282266259193, + 0.01172585692256689, + 0.020266661420464516, + 0.0121329165995121, + 0.02187856286764145, + 0.003873825306072831, + 0.0042217872105538845, + 0.004147945903241634, + 0.008442260324954987, + 0.003197268582880497, + 0.004129043314605951, + 0.007686407770961523, + 0.006780894938856363, + 0.010180884040892124, + 0.011450526304543018, + 0.01177128404378891, + 0.014130528084933758, + 0.005752929486334324, + 0.011851883493363857, + 0.011821435764431953, + 0.010931075550615788, + 0.007807108573615551, + 0.006229343358427286, + 0.00712176226079464, + 0.011372766457498074, + 0.006490375380963087, + 0.0041482895612716675, + 0.00991683267056942, + 0.008024349808692932, + 0.02647777833044529, + 0.02148757316172123, + 0.03794063627719879, + 0.02729089744389057, + 0.014705033041536808, + 0.012299512512981892, + 0.028654199093580246, + 0.043635815382003784, + 0.02648242935538292, + 0.016407174989581108, + 0.020450009033083916, + 0.027453893795609474, + 0.010585365816950798, + 0.007147247437387705, + 0.021950429305434227, + 0.01736590266227722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.07898691296577454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot from a shaky shoulder-mounted camera at a tense, tilted angle, locks the viewer's gaze squarely on a pale finger. Illuminated by a beam of cool white light from above, the pores and subtle texture of the skin are clearly visible. At the center of the fingertip, in a tiny, needle-like wound, a plump, crystalline drop of bright red blood slowly gathers and trembles, its surface reflecting a dazzling light. As the camera gently pans, a steady yet slightly trembling hand carefully brings the end of a white plastic test strip close to the falling drop. At the moment when the drop reaches its limit under its own weight, the tip of the strip, with its yellow capillary zone, precisely touches the edge, preparing to draw it in. The entire scene, imbued with a raw, documentary quality, completely fixates the viewer's attention on this crucial and suspenseful moment of medical testing, creating a breathtaking atmosphere of tension.", + "seed": 6185, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_1/index_6214_seed_6185_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": -0.12509660185314714, + "vlm_logprob": -0.12509660185314714, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.12509660185314714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006591245997697115 + ], + "opposite_logprob_tokens": [ + -6.631591320037842 + ], + "logprob_reward": -0.006591245997697115, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.006591245997697115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13431890308856964, + 0.05237919092178345, + 0.0622776634991169, + 0.09206505119800568, + 0.17528128623962402, + 0.09902109205722809, + 0.10651487857103348, + 0.09667856246232986, + 0.07620301097631454, + 0.10526036471128464 + ], + "credit_map": [ + 0.013203083537518978, + 0.014148766174912453, + 0.012579784728586674, + 0.0102087939158082, + 0.017391255125403404, + 0.02244897000491619, + 0.01811479590833187, + 0.026223456487059593, + 0.0052230507135391235, + 0.004891270771622658, + 0.00742553872987628, + 0.005001760087907314, + 0.0066283815540373325, + 0.006077678874135017, + 0.008953996933996677, + 0.008177514187991619, + 0.007044691126793623, + 0.006056523881852627, + 0.005328213330358267, + 0.009853558614850044, + 0.009381364099681377, + 0.007424289360642433, + 0.009752104990184307, + 0.007436915300786495, + 0.006384473294019699, + 0.007776267826557159, + 0.01050558127462864, + 0.009634374640882015, + 0.010077426210045815, + 0.018206842243671417, + 0.008509084582328796, + 0.02097100391983986, + 0.01423779595643282, + 0.044937342405319214, + 0.02258833684027195, + 0.010419068858027458, + 0.033608727157115936, + 0.014877130277454853, + 0.016528230160474777, + 0.01808464713394642, + 0.007489075884222984, + 0.009454559534788132, + 0.017170876264572144, + 0.006114553194493055, + 0.013605776242911816, + 0.013495604507625103, + 0.01851711794734001, + 0.01317352894693613, + 0.011958532966673374, + 0.019485050812363625, + 0.016228968277573586, + 0.008069708943367004, + 0.012422190979123116, + 0.010389844886958599, + 0.01650177873671055, + 0.011458800174295902, + 0.009120555594563484, + 0.010329791344702244, + 0.014495719224214554, + 0.009565101936459541, + 0.006956491153687239, + 0.012635355815291405, + 0.01489689014852047, + 0.018678657710552216, + 0.009467300958931446, + 0.01108783483505249, + 0.011515023186802864, + 0.0065179308876395226, + 0.010059362277388573, + 0.0118853859603405, + 0.006851280573755503, + 0.008818889036774635, + 0.0108655896037817, + 0.015107993967831135, + 0.01661793328821659, + 0.007607140578329563, + 0.01119233202189207, + 0.015838416293263435, + 0.009803907945752144, + 0.018227042630314827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.006591245997697115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234020948410034 + ], + "opposite_logprob_tokens": [ + -0.8984020948410034 + ], + "logprob_reward": -0.5234020948410034, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234020948410034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11126211285591125, + 0.061651501804590225, + 0.08014488965272903, + 0.0869293063879013, + 0.23306579887866974, + 0.12346458435058594, + 0.09201329201459885, + 0.0807318240404129, + 0.05718613043427467, + 0.0735505074262619 + ], + "credit_map": [ + 0.0114767299965024, + 0.01272770669311285, + 0.01552426815032959, + 0.01044902391731739, + 0.016834814101457596, + 0.014628870412707329, + 0.014062773436307907, + 0.015557932667434216, + 0.006601134315133095, + 0.0077073960565030575, + 0.00803051982074976, + 0.0056303152814507484, + 0.005429350771009922, + 0.007630922831594944, + 0.010046330280601978, + 0.010575535707175732, + 0.009972484782338142, + 0.010655157268047333, + 0.010322785936295986, + 0.007100194692611694, + 0.009882248938083649, + 0.010008616372942924, + 0.007884522899985313, + 0.014318886213004589, + 0.006520240567624569, + 0.00929263886064291, + 0.013387016952037811, + 0.006217740476131439, + 0.011390332132577896, + 0.014136498793959618, + 0.009191717952489853, + 0.016793126240372658, + 0.019477659836411476, + 0.025931352749466896, + 0.03861840069293976, + 0.013491974212229252, + 0.045792121440172195, + 0.03245590254664421, + 0.026309864595532417, + 0.030988531187176704, + 0.008523997850716114, + 0.018531132489442825, + 0.01967652514576912, + 0.007979714311659336, + 0.018269101157784462, + 0.018530048429965973, + 0.013365361839532852, + 0.018588704988360405, + 0.009384417906403542, + 0.019751913845539093, + 0.012050524353981018, + 0.007590930908918381, + 0.010307630524039268, + 0.01140538603067398, + 0.010577037930488586, + 0.010945457965135574, + 0.0075835539028048515, + 0.007998918183147907, + 0.01262359693646431, + 0.007514785975217819, + 0.010279113426804543, + 0.007517473306506872, + 0.014067973010241985, + 0.013146411627531052, + 0.007242141757160425, + 0.007445775903761387, + 0.0077781095169484615, + 0.004165969789028168, + 0.006690968759357929, + 0.007500522304326296, + 0.005505595356225967, + 0.010857049375772476, + 0.009937537834048271, + 0.011781858280301094, + 0.012117496691644192, + 0.005405157804489136, + 0.005754937417805195, + 0.013007973320782185, + 0.004580565728247166, + 0.010964985936880112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.5234020948410034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036258448380976915 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.0036258448380976915, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036258448380976915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12557317316532135, + 0.13616371154785156, + 0.06625180691480637, + 0.04866451025009155, + 0.199827641248703, + 0.11420981585979462, + 0.06849810481071472, + 0.11141481250524521, + 0.06135359779000282, + 0.0680427998304367 + ], + "credit_map": [ + 0.013256462290883064, + 0.014129583723843098, + 0.01917022466659546, + 0.01269090548157692, + 0.016615936532616615, + 0.013840622268617153, + 0.020377980545163155, + 0.015491455793380737, + 0.014820625074207783, + 0.01477382518351078, + 0.01614811085164547, + 0.012948398478329182, + 0.015955401584506035, + 0.018257224932312965, + 0.01751643419265747, + 0.025743698701262474, + 0.006487607024610043, + 0.009321188554167747, + 0.009024146944284439, + 0.00485825352370739, + 0.007705667521804571, + 0.008340422064065933, + 0.010340907610952854, + 0.01017361506819725, + 0.0040960428304970264, + 0.0047100805677473545, + 0.006824609357863665, + 0.0036415369249880314, + 0.005095429718494415, + 0.008868426084518433, + 0.005964004900306463, + 0.009464381262660027, + 0.013840173371136189, + 0.02525157295167446, + 0.04134025424718857, + 0.01305070798844099, + 0.024959834292531013, + 0.024606408551335335, + 0.026859425008296967, + 0.029919281601905823, + 0.009570938535034657, + 0.01704687997698784, + 0.014227712526917458, + 0.008674053475260735, + 0.018905246630311012, + 0.01924276351928711, + 0.012465957552194595, + 0.014076266437768936, + 0.0055706072598695755, + 0.006035085767507553, + 0.010993542149662971, + 0.004308468196541071, + 0.006941143423318863, + 0.006505511235445738, + 0.01592160016298294, + 0.012222148478031158, + 0.011374874040484428, + 0.01303846575319767, + 0.01824697107076645, + 0.011218957602977753, + 0.014880307950079441, + 0.008977371267974377, + 0.02095845714211464, + 0.012719405815005302, + 0.0036865021102130413, + 0.00448539899662137, + 0.009105423465371132, + 0.004657992627471685, + 0.007708298042416573, + 0.008236520923674107, + 0.006696814671158791, + 0.01677664928138256, + 0.005118633154779673, + 0.007339873351156712, + 0.011053468100726604, + 0.0036783195100724697, + 0.009331800043582916, + 0.01242775097489357, + 0.010217572562396526, + 0.008875380270183086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0036258448380976915 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028177103959023952 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028177103959023952, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028177103959023952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06829557567834854, + 0.04252628982067108, + 0.042243048548698425, + 0.03920403867959976, + 0.3583492934703827, + 0.13219217956066132, + 0.06901545822620392, + 0.08320484310388565, + 0.105246402323246, + 0.05972282961010933 + ], + "credit_map": [ + 0.006516287103295326, + 0.010144653730094433, + 0.008251136168837547, + 0.005482296459376812, + 0.010217452421784401, + 0.008521015755832195, + 0.010165795683860779, + 0.008996943011879921, + 0.004830338526517153, + 0.0034672378096729517, + 0.006084314547479153, + 0.0054666148498654366, + 0.004046432673931122, + 0.004180909600108862, + 0.009104501456022263, + 0.005345940589904785, + 0.005547800567001104, + 0.006489886902272701, + 0.006113783922046423, + 0.0033458624966442585, + 0.0037490054965019226, + 0.0046162912622094154, + 0.005086029879748821, + 0.007294387090951204, + 0.003256703494116664, + 0.003297839779406786, + 0.004949278198182583, + 0.005875767208635807, + 0.004935988690704107, + 0.006324744783341885, + 0.004373850300908089, + 0.006189869716763496, + 0.020604204386472702, + 0.02864430658519268, + 0.09424465149641037, + 0.015898026525974274, + 0.05937515199184418, + 0.03211403638124466, + 0.030859261751174927, + 0.07660970091819763, + 0.01234551053494215, + 0.017412368208169937, + 0.01727813482284546, + 0.013686535879969597, + 0.017724361270666122, + 0.021814720705151558, + 0.015282975509762764, + 0.016647566109895706, + 0.00599120557308197, + 0.010005955584347248, + 0.010063222609460354, + 0.006918989587575197, + 0.008780527859926224, + 0.008549218066036701, + 0.011284147389233112, + 0.00742219528183341, + 0.0065611302852630615, + 0.007445184979587793, + 0.012016115710139275, + 0.0069929687306284904, + 0.012368842959403992, + 0.011341419070959091, + 0.010208677500486374, + 0.016270514577627182, + 0.008233577944338322, + 0.011944551952183247, + 0.017859311774373055, + 0.008546444587409496, + 0.0074664512649178505, + 0.015032531693577766, + 0.009371609427034855, + 0.026791928336024284, + 0.006112574599683285, + 0.008037950843572617, + 0.007606540806591511, + 0.00459471857175231, + 0.008344929665327072, + 0.007501669693738222, + 0.009774118661880493, + 0.007750332355499268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.0028177103959023952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904611319303513 + ], + "opposite_logprob_tokens": [ + -2.464046001434326 + ], + "logprob_reward": -0.08904611319303513, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904611319303513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06999966502189636, + 0.06652010977268219, + 0.04324246197938919, + 0.03671230003237724, + 0.38079842925071716, + 0.13847871124744415, + 0.10693762451410294, + 0.0898304283618927, + 0.023484554141759872, + 0.043995678424835205 + ], + "credit_map": [ + 0.009886073879897594, + 0.007275840733200312, + 0.009394223801791668, + 0.005228761583566666, + 0.006808557081967592, + 0.0090797059237957, + 0.01057540625333786, + 0.011751104146242142, + 0.009185106493532658, + 0.009659722447395325, + 0.009041135199368, + 0.004586986731737852, + 0.010131476446986198, + 0.006575900129973888, + 0.009038922376930714, + 0.008300865069031715, + 0.004423779901117086, + 0.004275804851204157, + 0.004557033069431782, + 0.0038933081086724997, + 0.00779548566788435, + 0.004910665564239025, + 0.006167460698634386, + 0.007218922488391399, + 0.003179151564836502, + 0.002817007480189204, + 0.004866814240813255, + 0.00195208506193012, + 0.0040962351486086845, + 0.00945964828133583, + 0.0034587986301630735, + 0.006882561836391687, + 0.021850988268852234, + 0.029552359133958817, + 0.11127064377069473, + 0.014230387285351753, + 0.04689425602555275, + 0.031550366431474686, + 0.02808043733239174, + 0.0973690077662468, + 0.011419904418289661, + 0.021398033946752548, + 0.02067101188004017, + 0.011755884625017643, + 0.01700361631810665, + 0.023342834785580635, + 0.015047200955450535, + 0.017840225249528885, + 0.010668097995221615, + 0.01637721247971058, + 0.015592342242598534, + 0.012111212126910686, + 0.014333713799715042, + 0.012296433560550213, + 0.012370252050459385, + 0.013188362121582031, + 0.009848976507782936, + 0.009156813845038414, + 0.01323844213038683, + 0.007300953380763531, + 0.01176880020648241, + 0.014531935565173626, + 0.008207086473703384, + 0.01577741652727127, + 0.002241931390017271, + 0.00290433457121253, + 0.004017014056444168, + 0.0017232745885849, + 0.0033594383858144283, + 0.003531579626724124, + 0.0030905508901923895, + 0.002616431098431349, + 0.0034824558533728123, + 0.006374786142259836, + 0.006886596791446209, + 0.003364432603120804, + 0.004293810110539198, + 0.009959405288100243, + 0.003842503996565938, + 0.005791687872260809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.08904611319303513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot from a shaky shoulder-mounted camera at a tense, tilted angle, locks the viewer's gaze squarely on a pale finger. Illuminated by a beam of cool white light from above, the pores and subtle texture of the skin are clearly visible. At the center of the fingertip, in a tiny, needle-like wound, a plump, crystalline drop of bright red blood slowly gathers and trembles, its surface reflecting a dazzling light. As the camera gently pans, a steady yet slightly trembling hand carefully brings the end of a white plastic test strip close to the falling drop. At the moment when the drop reaches its limit under its own weight, the tip of the strip, with its yellow capillary zone, precisely touches the edge, preparing to draw it in. The entire scene, imbued with a raw, documentary quality, completely fixates the viewer's attention on this crucial and suspenseful moment of medical testing, creating a breathtaking atmosphere of tension.", + "seed": 6186, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_1/index_6214_seed_6186_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": -0.19322539679706097, + "vlm_logprob": -0.19322539679706097, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.19322539679706097, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a pale finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009363308548927307 + ], + "opposite_logprob_tokens": [ + -5.509363174438477 + ], + "logprob_reward": -0.009363308548927307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.009363308548927307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35550457239151, + 0.10940814763307571, + 0.06581560522317886, + 0.07196808606386185, + 0.060517072677612305, + 0.061392780393362045, + 0.02578696981072426, + 0.14914476871490479, + 0.04769168049097061, + 0.05277033895254135 + ], + "credit_map": [ + 0.021685389801859856, + 0.03822096809744835, + 0.02796984650194645, + 0.04535902291536331, + 0.04294367879629135, + 0.04988207295536995, + 0.10015794634819031, + 0.029285622760653496, + 0.010151161812245846, + 0.02022094652056694, + 0.01194843277335167, + 0.011617440730333328, + 0.01179162971675396, + 0.014649717137217522, + 0.020295588299632072, + 0.008733227849006653, + 0.0045875245705246925, + 0.013765446841716766, + 0.005414738319814205, + 0.00653580529615283, + 0.012978965416550636, + 0.011641759425401688, + 0.007971422746777534, + 0.002919941209256649, + 0.009182636626064777, + 0.01021572295576334, + 0.009569158777594566, + 0.007740245666354895, + 0.009601026773452759, + 0.010035797022283077, + 0.009076888673007488, + 0.00654660165309906, + 0.005671375431120396, + 0.005824003368616104, + 0.007472602650523186, + 0.009261448867619038, + 0.0063357301987707615, + 0.008065292611718178, + 0.011171894147992134, + 0.006714718881994486, + 0.007160683628171682, + 0.005023414734750986, + 0.008218164555728436, + 0.013240293599665165, + 0.007225331384688616, + 0.006721660029143095, + 0.008145450614392757, + 0.005657777655869722, + 0.002842262852936983, + 0.0026750119868665934, + 0.0030963358003646135, + 0.0031596794724464417, + 0.004635742399841547, + 0.003521781647577882, + 0.002357438439503312, + 0.003498713718727231, + 0.014414399862289429, + 0.018033357337117195, + 0.017213057726621628, + 0.01980414055287838, + 0.0196634903550148, + 0.02109326422214508, + 0.01808428391814232, + 0.020838769152760506, + 0.0044972579926252365, + 0.004860029090195894, + 0.004464284982532263, + 0.007279596291482449, + 0.00560626620426774, + 0.007245006039738655, + 0.005688031669706106, + 0.008051203563809395, + 0.006101810373365879, + 0.00878919567912817, + 0.007305982522666454, + 0.009999383240938187, + 0.00445789098739624, + 0.006390964146703482, + 0.005544195882976055, + 0.004180910065770149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.009363308548927307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578366994857788 + ], + "opposite_logprob_tokens": [ + -0.6328366994857788 + ], + "logprob_reward": -0.7578366994857788, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578366994857788, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11805647611618042, + 0.05621335282921791, + 0.1284797340631485, + 0.044343702495098114, + 0.05519882217049599, + 0.042602237313985825, + 0.05456257238984108, + 0.08820275217294693, + 0.06475386023521423, + 0.34758642315864563 + ], + "credit_map": [ + 0.01272533554583788, + 0.015519409440457821, + 0.014962502755224705, + 0.017964523285627365, + 0.01544574648141861, + 0.013322742655873299, + 0.019686182960867882, + 0.008430047892034054, + 0.00906627718359232, + 0.008922184817492962, + 0.007811450399458408, + 0.006743133999407291, + 0.004902547225356102, + 0.006873500999063253, + 0.007058537565171719, + 0.004835725296288729, + 0.0050234729424119, + 0.006711618509143591, + 0.005275006406009197, + 0.007253297138959169, + 0.04425857961177826, + 0.04508613422513008, + 0.008902167901396751, + 0.005969461984932423, + 0.00412625540047884, + 0.00523495115339756, + 0.004159323871135712, + 0.008894044905900955, + 0.004169114865362644, + 0.006454612128436565, + 0.006255744025111198, + 0.005049662198871374, + 0.008549719117581844, + 0.005485598463565111, + 0.009040847420692444, + 0.009853124618530273, + 0.0040540737099945545, + 0.005688264034688473, + 0.00712617626413703, + 0.0054010250605642796, + 0.004580921959131956, + 0.005463323090225458, + 0.005636603571474552, + 0.00905914418399334, + 0.004420795477926731, + 0.005281320307403803, + 0.004527856595814228, + 0.0036322770174592733, + 0.006923293694853783, + 0.005206821486353874, + 0.007288692053407431, + 0.011328251101076603, + 0.006221149582415819, + 0.005922175478190184, + 0.005511867348104715, + 0.006160329561680555, + 0.008319753222167492, + 0.00888979434967041, + 0.008230566047132015, + 0.013579709455370903, + 0.009942935779690742, + 0.013526921160519123, + 0.01264920737594366, + 0.013063873164355755, + 0.006878006737679243, + 0.006454237271100283, + 0.008942628279328346, + 0.013058427721261978, + 0.005342195276170969, + 0.007447968702763319, + 0.010655173100531101, + 0.005975234322249889, + 0.06828868389129639, + 0.07859114557504654, + 0.11080837994813919, + 0.024826325476169586, + 0.019823651760816574, + 0.01731322892010212, + 0.016691047698259354, + 0.01124400645494461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.7578366994857788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206396967172623 + ], + "opposite_logprob_tokens": [ + -2.8120639324188232 + ], + "logprob_reward": -0.06206396967172623, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206396967172623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29124870896339417, + 0.06796631217002869, + 0.10515246540307999, + 0.07524197548627853, + 0.052002403885126114, + 0.07130274176597595, + 0.035019125789403915, + 0.14316153526306152, + 0.10120681673288345, + 0.05769796669483185 + ], + "credit_map": [ + 0.01752885803580284, + 0.031044600531458855, + 0.021527284756302834, + 0.040916040539741516, + 0.03867649286985397, + 0.04125669598579407, + 0.07182233035564423, + 0.028476383537054062, + 0.004990384913980961, + 0.00816231407225132, + 0.006008441094309092, + 0.008913245052099228, + 0.00859803892672062, + 0.00974365696310997, + 0.015212351456284523, + 0.006337879691272974, + 0.0030634363647550344, + 0.007520137820392847, + 0.0032766014337539673, + 0.005672698374837637, + 0.035532593727111816, + 0.03581230342388153, + 0.009959228336811066, + 0.0043154628947377205, + 0.004642743617296219, + 0.006890284363180399, + 0.005143401678651571, + 0.010440501384437084, + 0.0083363251760602, + 0.015177791938185692, + 0.01490024197846651, + 0.00971067976206541, + 0.006513727828860283, + 0.00761828850954771, + 0.008397671394050121, + 0.006970887538045645, + 0.005654702428728342, + 0.006628008559346199, + 0.006314554717391729, + 0.0039045652374625206, + 0.006422324571758509, + 0.008501495234668255, + 0.006805231794714928, + 0.010650092735886574, + 0.015250612050294876, + 0.008633465506136417, + 0.007925109937787056, + 0.007114410866051912, + 0.003664565272629261, + 0.00358373299241066, + 0.0033012358471751213, + 0.004364949185401201, + 0.006488960236310959, + 0.0038881259970366955, + 0.005816678982228041, + 0.003910879138857126, + 0.010958799161016941, + 0.021508926525712013, + 0.013449088670313358, + 0.018600255250930786, + 0.020517118275165558, + 0.02085600420832634, + 0.01878279820084572, + 0.018488548696041107, + 0.008160272613167763, + 0.013769702054560184, + 0.010021616704761982, + 0.015598051249980927, + 0.011791210621595383, + 0.014813251793384552, + 0.013498006388545036, + 0.013554701581597328, + 0.007532373070716858, + 0.00829565990716219, + 0.009400387294590473, + 0.006322181783616543, + 0.007908182218670845, + 0.007148873060941696, + 0.007083836942911148, + 0.00400647334754467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.06206396967172623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978943519294262 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.00978943519294262, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978943519294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12135322391986847, + 0.06710419803857803, + 0.06176053360104561, + 0.140402689576149, + 0.15472157299518585, + 0.06523550301790237, + 0.04890430346131325, + 0.16049538552761078, + 0.055331751704216, + 0.12469086796045303 + ], + "credit_map": [ + 0.008362607099115849, + 0.012529205530881882, + 0.011442096903920174, + 0.016285499557852745, + 0.015582540072500706, + 0.018197551369667053, + 0.027425922453403473, + 0.011527803726494312, + 0.007172536104917526, + 0.010798643343150616, + 0.009730568155646324, + 0.006272222381085157, + 0.008383924141526222, + 0.009485834278166294, + 0.010897689498960972, + 0.004362781066447496, + 0.003212353680282831, + 0.005098797380924225, + 0.003097740700468421, + 0.006813683547079563, + 0.013684500940144062, + 0.013417230919003487, + 0.010824589990079403, + 0.005611637141555548, + 0.011205440387129784, + 0.019472714513540268, + 0.01054979395121336, + 0.015445142053067684, + 0.019155537709593773, + 0.0291921254247427, + 0.018540402874350548, + 0.016841530799865723, + 0.012523607350885868, + 0.017786303535103798, + 0.013738512992858887, + 0.02017269656062126, + 0.02709358185529709, + 0.029707269743084908, + 0.018467942252755165, + 0.015231658704578876, + 0.006731363479048014, + 0.0072432043962180614, + 0.006112325470894575, + 0.011948498897254467, + 0.011588530614972115, + 0.007896875031292439, + 0.008262432180345058, + 0.0054522729478776455, + 0.005356936249881983, + 0.006220020819455385, + 0.005411325488239527, + 0.005715612787753344, + 0.010828424245119095, + 0.005469867028295994, + 0.0052914805710315704, + 0.004610635340213776, + 0.014393462799489498, + 0.016777323558926582, + 0.01609504036605358, + 0.02298094891011715, + 0.020075401291251183, + 0.02429732121527195, + 0.020597927272319794, + 0.025277961045503616, + 0.004326588939875364, + 0.008256223052740097, + 0.005847153719514608, + 0.006428402848541737, + 0.01146953459829092, + 0.0069702863693237305, + 0.007356869522482157, + 0.004676695913076401, + 0.010778255760669708, + 0.027896059677004814, + 0.020638376474380493, + 0.014974309131503105, + 0.01781608909368515, + 0.010762580670416355, + 0.01308037806302309, + 0.008744814433157444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.00978943519294262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot shaky, tilted, and from a shoulder-mounted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a plump, crystalline drop of bright red blood visible gathering in a tiny wound on the fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hand bring a white plastic test strip with a yellow capillary zone close to the blood drop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan while the scene is illuminated by a beam of cool white light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707357108592987 + ], + "opposite_logprob_tokens": [ + -2.1270735263824463 + ], + "logprob_reward": -0.12707357108592987, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707357108592987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07408510148525238, + 0.06327846646308899, + 0.045355383306741714, + 0.057446252554655075, + 0.10763148963451385, + 0.02361014299094677, + 0.016693005338311195, + 0.04327923059463501, + 0.0810433104634285, + 0.4875776171684265 + ], + "credit_map": [ + 0.008584599941968918, + 0.016729731112718582, + 0.00924260076135397, + 0.007616470102220774, + 0.008811135776340961, + 0.010809829458594322, + 0.007503678556531668, + 0.004787059500813484, + 0.006123700644820929, + 0.011352363973855972, + 0.0075345104560256, + 0.010105539113283157, + 0.0063555557280778885, + 0.007032386958599091, + 0.00904148444533348, + 0.005732927471399307, + 0.0032277617137879133, + 0.005807000678032637, + 0.002976021496579051, + 0.0032149050384759903, + 0.01234111376106739, + 0.01323349867016077, + 0.0026271422393620014, + 0.0019279414555057883, + 0.008513329550623894, + 0.011334753595292568, + 0.00995099451392889, + 0.006749630440026522, + 0.007862808182835579, + 0.0047056651674211025, + 0.004347128327935934, + 0.003981940448284149, + 0.01286403089761734, + 0.009835485368967056, + 0.013021310791373253, + 0.015978960320353508, + 0.008725608699023724, + 0.013880900107324123, + 0.019861802458763123, + 0.013463390059769154, + 0.0020434448961168528, + 0.002901955973356962, + 0.002289256313815713, + 0.004383303690701723, + 0.003930308390408754, + 0.0034399880096316338, + 0.0022442874033004045, + 0.0023775994777679443, + 0.002284299349412322, + 0.0020938741508871317, + 0.002064679516479373, + 0.002613790798932314, + 0.002172385575249791, + 0.0020371864084154367, + 0.0017323081847280264, + 0.0016944811213761568, + 0.00407222006469965, + 0.004032798111438751, + 0.004752857610583305, + 0.008115111850202084, + 0.004258219618350267, + 0.005980819463729858, + 0.006933445110917091, + 0.005133758764714003, + 0.007591737434267998, + 0.008503022603690624, + 0.007541127502918243, + 0.011976169422268867, + 0.009538857266306877, + 0.011287146247923374, + 0.011725224554538727, + 0.012880025431513786, + 0.08375971764326096, + 0.1417180746793747, + 0.16560585796833038, + 0.03588972985744476, + 0.01727263256907463, + 0.020315570756793022, + 0.011030000634491444, + 0.011986039578914642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.12707357108592987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_tensor([9816, 9816])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_tensor([9816, 9816])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ca8fe0551d149d5f0a7df17b2ad2e48d9cbabb8f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_tensor([9816, 9816])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "disorienting low-angle tracking shot rotates on its axis in a continuous barrel roll, keeping pace with the cyclists moving at high speed. The subject, clad in a fitted, waterproof black technical jacket that gleams beneath a steady drizzle, slumps over the handlebars of their road bike. Their face is a mask of intense concentration, partially obscured by sleek, aerodynamic glasses with transparent lenses studded with tiny water droplets. At night, they pedal frantically, their legs a powerful, rhythmic blur propelling them forward, swiftly and decisively overtaking a line of slow-moving cars on the rain-slicked city streets. The black asphalt, like a dark, liquid mirror, reflects a chaotic swirl of red taillights and the vibrant electric glow of neon storefront signs. As the camera rotates, towering buildings and streetlights merge into a dazzling vortex of light and shadow. Water sprays from the rear wheel in an exquisite arc of illumination, capturing the multicolored light and adding to the scene's dynamic quality, creating a visceral and tense atmosphere of exhilarating speed and urban danger.", + "seed": 5930, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_9816_seed_5930_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": -0.08849599384702742, + "vlm_logprob": -0.08849599384702742, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.08849599384702742, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.01227126270532608 + ], + "opposite_logprob_tokens": [ + -4.887271404266357 + ], + "logprob_reward": -0.01227126270532608, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01227126270532608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2299855798482895, + 0.06052074581384659, + 0.07893171906471252, + 0.025292646139860153, + 0.08521557599306107, + 0.10691560059785843, + 0.052541334182024, + 0.23948422074317932, + 0.056822795420885086, + 0.06428978592157364 + ], + "credit_map": [ + 0.011981161311268806, + 0.014316658489406109, + 0.07013778388500214, + 0.024522924795746803, + 0.0632663145661354, + 0.014383978210389614, + 0.013663961552083492, + 0.0177127867937088, + 0.005871750880032778, + 0.0051977308467030525, + 0.011980476789176464, + 0.010761935263872147, + 0.006770723965018988, + 0.00540328910574317, + 0.007467062212526798, + 0.007067776285111904, + 0.009312395937740803, + 0.008060555905103683, + 0.01549569796770811, + 0.014699524268507957, + 0.009198885411024094, + 0.007064211182296276, + 0.007345233578234911, + 0.007755211088806391, + 0.002526377560570836, + 0.0027504521422088146, + 0.004623549524694681, + 0.004297441802918911, + 0.003943374380469322, + 0.0016267112223431468, + 0.003023297293111682, + 0.0025014427956193686, + 0.006269225385040045, + 0.007325301878154278, + 0.01581367291510105, + 0.013582168146967888, + 0.016663817688822746, + 0.006548650097101927, + 0.010035278275609016, + 0.008977456018328667, + 0.007920965552330017, + 0.019934065639972687, + 0.011254874989390373, + 0.017404524609446526, + 0.01572800613939762, + 0.018111813813447952, + 0.00893455371260643, + 0.0076267956756055355, + 0.005163290537893772, + 0.005140065681189299, + 0.0041501447558403015, + 0.006729835644364357, + 0.011624103412032127, + 0.006797865964472294, + 0.005503501743078232, + 0.0074325259774923325, + 0.03514375910162926, + 0.016281645745038986, + 0.0636061504483223, + 0.01571933925151825, + 0.02289760299026966, + 0.026151984930038452, + 0.027397306635975838, + 0.03228643909096718, + 0.008289581164717674, + 0.005123853217810392, + 0.009337452240288258, + 0.00466226926073432, + 0.004890474025160074, + 0.012665431015193462, + 0.004465209320187569, + 0.0073885247111320496, + 0.008306720294058323, + 0.0082650575786829, + 0.010256793349981308, + 0.007105063647031784, + 0.007619478739798069, + 0.011710402555763721, + 0.0042158192954957485, + 0.006810449529439211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.01227126270532608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.018206970766186714 + ], + "opposite_logprob_tokens": [ + -4.01820707321167 + ], + "logprob_reward": -0.018206970766186714, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018206970766186714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19273172318935394, + 0.08650217205286026, + 0.0882401391863823, + 0.053822316229343414, + 0.08020976185798645, + 0.09706936776638031, + 0.03384269028902054, + 0.10068228095769882, + 0.19205163419246674, + 0.07484791427850723 + ], + "credit_map": [ + 0.010262374766170979, + 0.011855892837047577, + 0.06399159878492355, + 0.02023484744131565, + 0.052273742854595184, + 0.01296216994524002, + 0.007047600112855434, + 0.01410349365323782, + 0.008624088950455189, + 0.0061728511936962605, + 0.018273940309882164, + 0.013783322647213936, + 0.010579396970570087, + 0.007993766106665134, + 0.012254094704985619, + 0.008820708841085434, + 0.010911785066127777, + 0.010337380692362785, + 0.01416002307087183, + 0.013943793252110481, + 0.011060308665037155, + 0.007390348706394434, + 0.009108276106417179, + 0.01132822223007679, + 0.007570062298327684, + 0.0053518787026405334, + 0.01053264457732439, + 0.009521198458969593, + 0.008775044232606888, + 0.0034271294716745615, + 0.004515892826020718, + 0.0041284626349806786, + 0.008408544585108757, + 0.009351372718811035, + 0.013295717537403107, + 0.011917718686163425, + 0.010775923728942871, + 0.006484381388872862, + 0.00834535900503397, + 0.011630750261247158, + 0.008982885628938675, + 0.009109769947826862, + 0.007564747240394354, + 0.016233427450060844, + 0.017365440726280212, + 0.018867524340748787, + 0.011231763288378716, + 0.007713808212429285, + 0.0037553170695900917, + 0.0030849112663418055, + 0.0025334800593554974, + 0.00452263792976737, + 0.006951407063752413, + 0.002950876485556364, + 0.0038675772957503796, + 0.006176482420414686, + 0.011581453494727612, + 0.00728128245100379, + 0.017519094049930573, + 0.005920614115893841, + 0.012427839450538158, + 0.014697013422846794, + 0.018412070348858833, + 0.012842917814850807, + 0.028571026399731636, + 0.015951750800013542, + 0.033424973487854004, + 0.015426672995090485, + 0.02013411372900009, + 0.02468642219901085, + 0.030943527817726135, + 0.022913146764039993, + 0.009304383769631386, + 0.007764223963022232, + 0.011312870308756828, + 0.00906775239855051, + 0.009197517298161983, + 0.014827108010649681, + 0.005783826112747192, + 0.007590229157358408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.018206970766186714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.12707975506782532 + ], + "opposite_logprob_tokens": [ + -2.127079725265503 + ], + "logprob_reward": -0.12707975506782532, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707975506782532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.251057505607605, + 0.10608325898647308, + 0.11803454905748367, + 0.1200793907046318, + 0.06643274426460266, + 0.05237937718629837, + 0.041683487594127655, + 0.12189675867557526, + 0.06136239692568779, + 0.06099054217338562 + ], + "credit_map": [ + 0.010396759025752544, + 0.009088702499866486, + 0.09427112340927124, + 0.02384926937520504, + 0.07776894420385361, + 0.011535652913153172, + 0.006861865520477295, + 0.017285192385315895, + 0.008355625905096531, + 0.008662143722176552, + 0.02002074383199215, + 0.01642324961721897, + 0.015101193450391293, + 0.009929656982421875, + 0.01945706084370613, + 0.008133584633469582, + 0.016214800998568535, + 0.014552773907780647, + 0.019575154408812523, + 0.020596878603100777, + 0.013419748283922672, + 0.009389645420014858, + 0.012867016717791557, + 0.011418528854846954, + 0.01123876217752695, + 0.012848436832427979, + 0.028304029256105423, + 0.0285804346203804, + 0.017905283719301224, + 0.006199544761329889, + 0.008726471103727818, + 0.006276426836848259, + 0.005720603279769421, + 0.015032188035547733, + 0.01183396577835083, + 0.01073467917740345, + 0.008954429998993874, + 0.004215863533318043, + 0.0051336768083274364, + 0.0048073395155370235, + 0.0066011263988912106, + 0.008072668686509132, + 0.004934906028211117, + 0.00925778690725565, + 0.007304700557142496, + 0.006290067452937365, + 0.006042396649718285, + 0.003875722410157323, + 0.0038354583084583282, + 0.004123356658965349, + 0.0036715238820761442, + 0.006218104623258114, + 0.005796068347990513, + 0.005183351691812277, + 0.006031942553818226, + 0.006823680363595486, + 0.011776212602853775, + 0.007045231759548187, + 0.015006371773779392, + 0.008660675026476383, + 0.024694999679923058, + 0.014393032528460026, + 0.027199141681194305, + 0.013121085241436958, + 0.009103994816541672, + 0.005500872619450092, + 0.007977934554219246, + 0.005462128669023514, + 0.005911532789468765, + 0.011978277005255222, + 0.008345525711774826, + 0.007082127500325441, + 0.006502637639641762, + 0.010652899742126465, + 0.009808813221752644, + 0.008526328951120377, + 0.007905883714556694, + 0.004359394311904907, + 0.0066602290607988834, + 0.006574355531483889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.12707975506782532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.28130480647087097 + ], + "opposite_logprob_tokens": [ + -1.4063048362731934 + ], + "logprob_reward": -0.28130480647087097, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28130480647087097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1967112421989441, + 0.08202957361936569, + 0.06802177429199219, + 0.22500813007354736, + 0.10441287606954575, + 0.15581397712230682, + 0.03720438480377197, + 0.03740570694208145, + 0.048637207597494125, + 0.044755127280950546 + ], + "credit_map": [ + 0.0119694285094738, + 0.014920759946107864, + 0.04957299306988716, + 0.02129046618938446, + 0.047826118767261505, + 0.02015370875597, + 0.014122705906629562, + 0.016855066642165184, + 0.008411713875830173, + 0.005912753753364086, + 0.017328809946775436, + 0.016368087381124496, + 0.012279557064175606, + 0.006331063807010651, + 0.00929933786392212, + 0.0060982489958405495, + 0.008989916183054447, + 0.00809283833950758, + 0.008852600120007992, + 0.01054154522716999, + 0.011004000902175903, + 0.004639059770852327, + 0.007588824722915888, + 0.008312985301017761, + 0.02792242541909218, + 0.02213515155017376, + 0.04730156436562538, + 0.042101625353097916, + 0.03890138491988182, + 0.012987159192562103, + 0.01929384656250477, + 0.014364988543093204, + 0.0089808264747262, + 0.018019303679466248, + 0.019638262689113617, + 0.017038516700267792, + 0.016838865354657173, + 0.006257752887904644, + 0.009389303624629974, + 0.00825005117803812, + 0.015367208048701286, + 0.014937584288418293, + 0.011370660737156868, + 0.025465471670031548, + 0.030161375179886818, + 0.020938778296113014, + 0.021217331290245056, + 0.01635555550456047, + 0.0032741131726652384, + 0.00483662448823452, + 0.0028968174010515213, + 0.005505561828613281, + 0.007185356225818396, + 0.0038476898334920406, + 0.004215977154672146, + 0.005442243069410324, + 0.005461628083139658, + 0.002856828970834613, + 0.0069315372966229916, + 0.004074633587151766, + 0.004429722670465708, + 0.00467940978705883, + 0.003950604237616062, + 0.005021342076361179, + 0.005417513661086559, + 0.0050854189321398735, + 0.0083727752789855, + 0.003936129156500101, + 0.00494835851714015, + 0.011249490082263947, + 0.00421555433422327, + 0.005411967169493437, + 0.005034199915826321, + 0.00480179488658905, + 0.007189730182290077, + 0.004688782151788473, + 0.007821625098586082, + 0.005473271012306213, + 0.004933600779622793, + 0.004812125116586685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.28130480647087097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.0036171742249280214 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036171742249280214, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036171742249280214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17554417252540588, + 0.04098689183592796, + 0.07925795763731003, + 0.0689159482717514, + 0.09587932378053665, + 0.0622822642326355, + 0.03664184734225273, + 0.24104008078575134, + 0.10064216703176498, + 0.0988093763589859 + ], + "credit_map": [ + 0.011799047701060772, + 0.009368541650474072, + 0.0496729277074337, + 0.015678295865654945, + 0.050935205072164536, + 0.01441759429872036, + 0.009473885409533978, + 0.014198677614331245, + 0.004307680297642946, + 0.0038089589215815067, + 0.007086047902703285, + 0.005987402983009815, + 0.005437631160020828, + 0.0041183303110301495, + 0.005931732710450888, + 0.004309108946472406, + 0.010014806874096394, + 0.008736851625144482, + 0.016318144276738167, + 0.012245675548911095, + 0.010262659750878811, + 0.006618119310587645, + 0.007233534008264542, + 0.00782817043364048, + 0.007942316122353077, + 0.007402368355542421, + 0.01584869995713234, + 0.013354107737541199, + 0.009175889194011688, + 0.00431566359475255, + 0.006285376846790314, + 0.004591523203998804, + 0.008277120999991894, + 0.025094524025917053, + 0.012179992161691189, + 0.017359362915158272, + 0.011765162460505962, + 0.004167091101408005, + 0.008001483976840973, + 0.009034575894474983, + 0.005229435861110687, + 0.012552713975310326, + 0.00523478165268898, + 0.011811698786914349, + 0.0062597449868917465, + 0.007809883449226618, + 0.006787202786654234, + 0.006596799474209547, + 0.0038937763310968876, + 0.00426502525806427, + 0.0036637757439166307, + 0.004613094963133335, + 0.007517877500504255, + 0.004045382607728243, + 0.004433945752680302, + 0.004208970814943314, + 0.03268739953637123, + 0.015312070026993752, + 0.06879720836877823, + 0.01554263848811388, + 0.02323818765580654, + 0.02702373079955578, + 0.027419444173574448, + 0.031019384041428566, + 0.012823255732655525, + 0.007875991053879261, + 0.016725560650229454, + 0.0069601577706635, + 0.008105576038360596, + 0.027314642444252968, + 0.008299657143652439, + 0.012537320144474506, + 0.011810258962213993, + 0.008447874337434769, + 0.01383909024298191, + 0.008519369177520275, + 0.01347881555557251, + 0.026031462475657463, + 0.005554957780987024, + 0.011127544566988945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0036171742249280214 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "disorienting low-angle tracking shot rotates on its axis in a continuous barrel roll, keeping pace with the cyclists moving at high speed. The subject, clad in a fitted, waterproof black technical jacket that gleams beneath a steady drizzle, slumps over the handlebars of their road bike. Their face is a mask of intense concentration, partially obscured by sleek, aerodynamic glasses with transparent lenses studded with tiny water droplets. At night, they pedal frantically, their legs a powerful, rhythmic blur propelling them forward, swiftly and decisively overtaking a line of slow-moving cars on the rain-slicked city streets. The black asphalt, like a dark, liquid mirror, reflects a chaotic swirl of red taillights and the vibrant electric glow of neon storefront signs. As the camera rotates, towering buildings and streetlights merge into a dazzling vortex of light and shadow. Water sprays from the rear wheel in an exquisite arc of illumination, capturing the multicolored light and adding to the scene's dynamic quality, creating a visceral and tense atmosphere of exhilarating speed and urban danger.", + "seed": 5931, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_9816_seed_5931_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": -0.17973413318395615, + "vlm_logprob": -0.17973413318395615, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.17973413318395615, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.440225213766098 + ], + "opposite_logprob_tokens": [ + -1.0652252435684204 + ], + "logprob_reward": -0.440225213766098, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.440225213766098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056345224380493164, + 0.03939278796315193, + 0.39956599473953247, + 0.08224916458129883, + 0.01994529366493225, + 0.08562101423740387, + 0.186155766248703, + 0.04082874953746796, + 0.05414117872714996, + 0.035754844546318054 + ], + "credit_map": [ + 0.0059646419249475, + 0.004139112774282694, + 0.005920866038650274, + 0.0027811219915747643, + 0.014458263292908669, + 0.0059311105869710445, + 0.0038435799069702625, + 0.013306532055139542, + 0.00461183488368988, + 0.004008383955806494, + 0.00651967478916049, + 0.006160594988614321, + 0.0038187168538570404, + 0.004753928165882826, + 0.003432831959798932, + 0.006086823996156454, + 0.03534901514649391, + 0.03979452699422836, + 0.06860030442476273, + 0.05062118172645569, + 0.05083737522363663, + 0.05413873866200447, + 0.07989691197872162, + 0.02032792940735817, + 0.007457280065864325, + 0.008827758952975273, + 0.012496039271354675, + 0.01145131140947342, + 0.00826228316873312, + 0.010403186082839966, + 0.010573282837867737, + 0.01277802512049675, + 0.0025634458288550377, + 0.003938010428100824, + 0.0025255184154957533, + 0.0022790913935750723, + 0.0032322080805897713, + 0.001676735351793468, + 0.0016306336037814617, + 0.0020996492821723223, + 0.00619399081915617, + 0.006837493274360895, + 0.0039689973928034306, + 0.011835647746920586, + 0.022044170647859573, + 0.006100417114794254, + 0.0040466198697686195, + 0.02459368109703064, + 0.028652537614107132, + 0.023575004190206528, + 0.019672460854053497, + 0.022529710084199905, + 0.02125619538128376, + 0.02056129463016987, + 0.03401833400130272, + 0.015890231356024742, + 0.006613898556679487, + 0.0058464547619223595, + 0.004630540497601032, + 0.007144056260585785, + 0.004037783946841955, + 0.0044367690570652485, + 0.0044005983509123325, + 0.0037186488043516874, + 0.007557520642876625, + 0.008294274099171162, + 0.010598178021609783, + 0.0071246144361793995, + 0.0055096386931836605, + 0.0034035476855933666, + 0.004979875870049, + 0.0066735283471643925, + 0.005191213916987181, + 0.004019617568701506, + 0.005782517604529858, + 0.0058188545517623425, + 0.0031583530362695456, + 0.004510597325861454, + 0.003864326048642397, + 0.003409366821870208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.440225213766098 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.016144733875989914 + ], + "opposite_logprob_tokens": [ + -4.141144752502441 + ], + "logprob_reward": -0.016144733875989914, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016144733875989914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18434958159923553, + 0.0716480240225792, + 0.08730174601078033, + 0.28298819065093994, + 0.06049873307347298, + 0.04464038088917732, + 0.08824767917394638, + 0.0509188175201416, + 0.06286489218473434, + 0.06654195487499237 + ], + "credit_map": [ + 0.014275438152253628, + 0.012000782415270805, + 0.014408588409423828, + 0.010221239179372787, + 0.05245401710271835, + 0.012623419053852558, + 0.016038397327065468, + 0.05232769623398781, + 0.01073317602276802, + 0.009454464539885521, + 0.011503535322844982, + 0.010333165526390076, + 0.007737764157354832, + 0.009754601866006851, + 0.007080103270709515, + 0.005051213316619396, + 0.008246329613029957, + 0.009322412312030792, + 0.015421954914927483, + 0.010196401737630367, + 0.010734672658145428, + 0.010182585567235947, + 0.01826605573296547, + 0.004931334406137466, + 0.02927251346409321, + 0.029613278806209564, + 0.04402347281575203, + 0.03920869156718254, + 0.027421949431300163, + 0.03151915594935417, + 0.038603901863098145, + 0.04332524538040161, + 0.00789090245962143, + 0.007478632498532534, + 0.009498421102762222, + 0.00516736414283514, + 0.005720851477235556, + 0.006062118336558342, + 0.003991207107901573, + 0.014689233154058456, + 0.003447504248470068, + 0.0048643676564097404, + 0.0032652728259563446, + 0.006210168823599815, + 0.00945287849754095, + 0.004474081099033356, + 0.003907956648617983, + 0.00901815015822649, + 0.012766594998538494, + 0.016757948324084282, + 0.011535638011991978, + 0.008149937726557255, + 0.006305083632469177, + 0.012117373757064342, + 0.009939480572938919, + 0.010675614699721336, + 0.007199794985353947, + 0.00964223500341177, + 0.0048475125804543495, + 0.00502576120197773, + 0.00442853756248951, + 0.005228215828537941, + 0.007357817143201828, + 0.007188939489424229, + 0.0062198154628276825, + 0.007515075150877237, + 0.00921641755849123, + 0.0060504223220050335, + 0.007897506467998028, + 0.01091584749519825, + 0.0068921977654099464, + 0.008157609030604362, + 0.006166457664221525, + 0.005583386868238449, + 0.007619921583682299, + 0.00848607812076807, + 0.009661135263741016, + 0.011470019817352295, + 0.008804557844996452, + 0.00875039678066969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.016144733875989914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.16039785742759705 + ], + "opposite_logprob_tokens": [ + -1.9103978872299194 + ], + "logprob_reward": -0.16039785742759705, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039785742759705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15158621966838837, + 0.074094757437706, + 0.05970895290374756, + 0.3906269967556, + 0.07046487927436829, + 0.03546783700585365, + 0.06863096356391907, + 0.05988841503858566, + 0.04713950678706169, + 0.042391471564769745 + ], + "credit_map": [ + 0.013188950717449188, + 0.013694331049919128, + 0.015709886327385902, + 0.01052870787680149, + 0.03387146443128586, + 0.015530250035226345, + 0.00921272486448288, + 0.039849914610385895, + 0.009218893013894558, + 0.006050036288797855, + 0.009569121524691582, + 0.011127189733088017, + 0.008570985868573189, + 0.009938626550137997, + 0.00717944884672761, + 0.012440462596714497, + 0.004081788472831249, + 0.009048203006386757, + 0.012510094791650772, + 0.013531452044844627, + 0.006707482505589724, + 0.003358864691108465, + 0.006686687469482422, + 0.0037843785248696804, + 0.04180004447698593, + 0.03522905707359314, + 0.05736861377954483, + 0.048821330070495605, + 0.03789935261011124, + 0.04556510969996452, + 0.04381740838289261, + 0.08012609928846359, + 0.010759556666016579, + 0.013721699826419353, + 0.01011043880134821, + 0.007119893096387386, + 0.008686546236276627, + 0.006435684394091368, + 0.005717230960726738, + 0.007913821376860142, + 0.004343277774751186, + 0.0037375111132860184, + 0.0023861532099545, + 0.004093565046787262, + 0.006732983514666557, + 0.004236938431859016, + 0.002364045474678278, + 0.007573361974209547, + 0.012647178955376148, + 0.01588921621441841, + 0.008219492621719837, + 0.00680349487811327, + 0.004955526441335678, + 0.0058272238820791245, + 0.007858931086957455, + 0.006429901346564293, + 0.007123903837054968, + 0.008173812180757523, + 0.005477722734212875, + 0.00864487886428833, + 0.005862138699740171, + 0.008419392630457878, + 0.009070416912436485, + 0.007116150576621294, + 0.004508430603891611, + 0.0058517795987427235, + 0.008109834045171738, + 0.004797609522938728, + 0.006129121407866478, + 0.003899072762578726, + 0.00777994841337204, + 0.006063710432499647, + 0.00577399879693985, + 0.0037542530335485935, + 0.007238913793116808, + 0.005953321233391762, + 0.005066317040473223, + 0.004215200897306204, + 0.00532393716275692, + 0.005065528210252523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.16039785742759705 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.2520933151245117 + ], + "opposite_logprob_tokens": [ + -1.5020933151245117 + ], + "logprob_reward": -0.2520933151245117, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520933151245117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09668462723493576, + 0.07488438487052917, + 0.08920378237962723, + 0.12624181807041168, + 0.16881012916564941, + 0.1532071828842163, + 0.0910712406039238, + 0.09013056755065918, + 0.06657874584197998, + 0.043187472969293594 + ], + "credit_map": [ + 0.006614805664867163, + 0.00407137768343091, + 0.009114349260926247, + 0.005907125771045685, + 0.027852652594447136, + 0.007780164014548063, + 0.005922409240156412, + 0.029421741142868996, + 0.010217010974884033, + 0.0075032152235507965, + 0.010848142206668854, + 0.009477590210735798, + 0.007906569167971611, + 0.013084063306450844, + 0.005138906184583902, + 0.01070888340473175, + 0.006722711957991123, + 0.012076059356331825, + 0.01589146815240383, + 0.013161214999854565, + 0.011242659762501717, + 0.00615254882723093, + 0.01728593185544014, + 0.006671187002211809, + 0.01408951822668314, + 0.011216138489544392, + 0.018433190882205963, + 0.0180625319480896, + 0.010288040153682232, + 0.012865277007222176, + 0.012834036722779274, + 0.028453076258301735, + 0.01621921733021736, + 0.019266989082098007, + 0.03909074887633324, + 0.013435288332402706, + 0.037351083010435104, + 0.009894431568682194, + 0.010412847623229027, + 0.023139523342251778, + 0.01667078398168087, + 0.01086419727653265, + 0.011471654288470745, + 0.020439818501472473, + 0.034030668437480927, + 0.014963255263864994, + 0.008076369762420654, + 0.03669043630361557, + 0.020656654611229897, + 0.020837610587477684, + 0.009977303445339203, + 0.007767203729599714, + 0.007115515414625406, + 0.008218727074563503, + 0.009341566823422909, + 0.007156663108617067, + 0.009156312793493271, + 0.012190598994493484, + 0.009023498743772507, + 0.007856611162424088, + 0.011843382380902767, + 0.021685969084501266, + 0.008827625773847103, + 0.009546568617224693, + 0.007155560422688723, + 0.008183336816728115, + 0.008870469406247139, + 0.006723048631101847, + 0.010333721525967121, + 0.010038096457719803, + 0.0056534805335104465, + 0.009621035307645798, + 0.008026238530874252, + 0.005039148963987827, + 0.009166487492620945, + 0.0059869298711419106, + 0.0031209809239953756, + 0.0036673869471997023, + 0.003891182132065296, + 0.004289120435714722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.2520933151245117 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.02980954572558403 + ], + "opposite_logprob_tokens": [ + -3.5298094749450684 + ], + "logprob_reward": -0.02980954572558403, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980954572558403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08582647144794464, + 0.0500975176692009, + 0.17253120243549347, + 0.11824027448892593, + 0.05075522139668465, + 0.0523330383002758, + 0.20805886387825012, + 0.10576597601175308, + 0.0879141315817833, + 0.06847736239433289 + ], + "credit_map": [ + 0.007027987856417894, + 0.006635242607444525, + 0.015725400298833847, + 0.008590065874159336, + 0.013746094889938831, + 0.015092460438609123, + 0.008992619812488556, + 0.010016587562859058, + 0.006735774222761393, + 0.004062266554683447, + 0.0073639689944684505, + 0.006560249254107475, + 0.006397401914000511, + 0.007731801830232143, + 0.0034673702903091908, + 0.007778679486364126, + 0.020077306777238846, + 0.017594410106539726, + 0.0268903486430645, + 0.020169248804450035, + 0.020200269296765327, + 0.03194310516119003, + 0.02452700026333332, + 0.011129499413073063, + 0.010238425806164742, + 0.011469360440969467, + 0.018314430490136147, + 0.01567312330007553, + 0.012754663825035095, + 0.01633215881884098, + 0.015405731275677681, + 0.018052365630865097, + 0.00494417967274785, + 0.0064509473741054535, + 0.008680065162479877, + 0.004437985830008984, + 0.004915312863886356, + 0.004485831130295992, + 0.003208535723388195, + 0.013632358983159065, + 0.005556653253734112, + 0.005460564978420734, + 0.004434820264577866, + 0.0066094473004341125, + 0.010413378477096558, + 0.0044154576025903225, + 0.0030288943089544773, + 0.012413816526532173, + 0.04932956397533417, + 0.0460624136030674, + 0.02019818127155304, + 0.021753301844000816, + 0.019266488030552864, + 0.01262081228196621, + 0.026672501116991043, + 0.012155581265687943, + 0.016498738899827003, + 0.016401775181293488, + 0.010037627071142197, + 0.0200804453343153, + 0.008858120068907738, + 0.01412119623273611, + 0.011349868029356003, + 0.008418199606239796, + 0.010656364262104034, + 0.012691762298345566, + 0.01580055058002472, + 0.011103004217147827, + 0.01026900578290224, + 0.005656998138874769, + 0.010867205448448658, + 0.010869225487112999, + 0.011039293371140957, + 0.006886683404445648, + 0.01269436813890934, + 0.014634011313319206, + 0.005495503079146147, + 0.005025043152272701, + 0.006583033595234156, + 0.006119417957961559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.02980954572558403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "disorienting low-angle tracking shot rotates on its axis in a continuous barrel roll, keeping pace with the cyclists moving at high speed. The subject, clad in a fitted, waterproof black technical jacket that gleams beneath a steady drizzle, slumps over the handlebars of their road bike. Their face is a mask of intense concentration, partially obscured by sleek, aerodynamic glasses with transparent lenses studded with tiny water droplets. At night, they pedal frantically, their legs a powerful, rhythmic blur propelling them forward, swiftly and decisively overtaking a line of slow-moving cars on the rain-slicked city streets. The black asphalt, like a dark, liquid mirror, reflects a chaotic swirl of red taillights and the vibrant electric glow of neon storefront signs. As the camera rotates, towering buildings and streetlights merge into a dazzling vortex of light and shadow. Water sprays from the rear wheel in an exquisite arc of illumination, capturing the multicolored light and adding to the scene's dynamic quality, creating a visceral and tense atmosphere of exhilarating speed and urban danger.", + "seed": 5932, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_9816_seed_5932_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": -0.09405759256333113, + "vlm_logprob": -0.09405759256333113, + "vlm_yes_no_margin": 2.9749999761581423, + "vlm_token_credit": -0.09405759256333113, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.06580623239278793 + ], + "opposite_logprob_tokens": [ + -2.9408061504364014 + ], + "logprob_reward": -0.06580623239278793, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06580623239278793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2023402601480484, + 0.13883814215660095, + 0.06992834061384201, + 0.052854299545288086, + 0.03667507693171501, + 0.09559604525566101, + 0.08038411289453506, + 0.09011983871459961, + 0.13674995303153992, + 0.09651397913694382 + ], + "credit_map": [ + 0.030071895569562912, + 0.042857736349105835, + 0.02270624414086342, + 0.031579677015542984, + 0.024061696603894234, + 0.025189116597175598, + 0.010477767325937748, + 0.015396123751997948, + 0.01504551898688078, + 0.01965988613665104, + 0.032576169818639755, + 0.017042802646756172, + 0.015875697135925293, + 0.021605990827083588, + 0.00741583202034235, + 0.009616225026547909, + 0.006487286649644375, + 0.012663130648434162, + 0.01018888782709837, + 0.008621606975793839, + 0.009326686151325703, + 0.011715199798345566, + 0.0057103452272713184, + 0.005215199198573828, + 0.00787388440221548, + 0.007791783194988966, + 0.005889260675758123, + 0.006572933401912451, + 0.007329404819756746, + 0.005373334512114525, + 0.0061973147094249725, + 0.005826386157423258, + 0.003617583541199565, + 0.0036671296693384647, + 0.00599546916782856, + 0.007362459320574999, + 0.0038360487669706345, + 0.0028344844467937946, + 0.006963414140045643, + 0.002398486016318202, + 0.008005985990166664, + 0.007596051320433617, + 0.007828116416931152, + 0.011114044114947319, + 0.011549025774002075, + 0.024094052612781525, + 0.014421922154724598, + 0.01098684873431921, + 0.013837868347764015, + 0.012000669725239277, + 0.009499795734882355, + 0.01153777539730072, + 0.0068900734186172485, + 0.009792444296181202, + 0.007585044950246811, + 0.009240440092980862, + 0.007556977216154337, + 0.013205486349761486, + 0.010115151293575764, + 0.015723448246717453, + 0.014727250672876835, + 0.013149451464414597, + 0.007654561195522547, + 0.007987512275576591, + 0.013086509890854359, + 0.01330228429287672, + 0.03050418756902218, + 0.013883308507502079, + 0.009968324564397335, + 0.02408275380730629, + 0.013034723699092865, + 0.018887856975197792, + 0.008061232976615429, + 0.008300984278321266, + 0.012508448213338852, + 0.010166400112211704, + 0.007585494313389063, + 0.02073233760893345, + 0.012074632570147514, + 0.017084451392292976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.06580623239278793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.014239607378840446 + ], + "opposite_logprob_tokens": [ + -4.26423978805542 + ], + "logprob_reward": -0.014239607378840446, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014239607378840446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06895387917757034, + 0.052461158484220505, + 0.08309106528759003, + 0.4066166579723358, + 0.03635421022772789, + 0.03800955042243004, + 0.06962708383798599, + 0.08006829768419266, + 0.09894128143787384, + 0.06587686389684677 + ], + "credit_map": [ + 0.008908571675419807, + 0.010825734585523605, + 0.009406674653291702, + 0.009548674337565899, + 0.008792285807430744, + 0.00808549765497446, + 0.005478897597640753, + 0.007907532155513763, + 0.0058016483671963215, + 0.008596603758633137, + 0.01297010388225317, + 0.00805472768843174, + 0.006434818264096975, + 0.005172638222575188, + 0.0029675974510610104, + 0.002463012933731079, + 0.01884515769779682, + 0.024212270975112915, + 0.004954805132001638, + 0.008554751053452492, + 0.009317059069871902, + 0.00608833646401763, + 0.006174506153911352, + 0.004944168496876955, + 0.08912158012390137, + 0.029099274426698685, + 0.017947005107998848, + 0.03954073414206505, + 0.033546365797519684, + 0.02418070286512375, + 0.019831418991088867, + 0.15334953367710114, + 0.004707265645265579, + 0.003266950137913227, + 0.004183446057140827, + 0.005254254676401615, + 0.0041642761789262295, + 0.003754283068701625, + 0.007139559369534254, + 0.00388417043723166, + 0.005251628812402487, + 0.003274782095104456, + 0.004081335384398699, + 0.004193291533738375, + 0.005667808465659618, + 0.006556432228535414, + 0.00522522022947669, + 0.0037590477149933577, + 0.00966036319732666, + 0.00862251315265894, + 0.007953135296702385, + 0.010078860446810722, + 0.009310416877269745, + 0.007174258586019278, + 0.009960882365703583, + 0.006866645533591509, + 0.008856401778757572, + 0.013603274710476398, + 0.009020966477692127, + 0.014476248063147068, + 0.011167977005243301, + 0.008424712345004082, + 0.0076248603872954845, + 0.0068938541226089, + 0.01450107991695404, + 0.010775125585496426, + 0.013480359688401222, + 0.014521865174174309, + 0.008032790385186672, + 0.010889510624110699, + 0.01358688436448574, + 0.013153661973774433, + 0.007710565347224474, + 0.006895565893501043, + 0.007184438873082399, + 0.006915247067809105, + 0.010235919617116451, + 0.009777029976248741, + 0.008001795038580894, + 0.009156296029686928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.014239607378840446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.1270836442708969 + ], + "opposite_logprob_tokens": [ + -2.1270835399627686 + ], + "logprob_reward": -0.1270836442708969, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270836442708969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13986755907535553, + 0.0738760307431221, + 0.12755855917930603, + 0.12908488512039185, + 0.07467925548553467, + 0.02943294309079647, + 0.09250480681657791, + 0.18677151203155518, + 0.07024971395730972, + 0.0759747102856636 + ], + "credit_map": [ + 0.01589847542345524, + 0.021444067358970642, + 0.021117104217410088, + 0.016247212886810303, + 0.019788822159171104, + 0.0176169965416193, + 0.011572809889912605, + 0.016182057559490204, + 0.008094768039882183, + 0.006989079061895609, + 0.009839151985943317, + 0.013279475271701813, + 0.009605992585420609, + 0.014002532698214054, + 0.006243810057640076, + 0.005821222439408302, + 0.030167190358042717, + 0.03607147932052612, + 0.00795978493988514, + 0.011491512879729271, + 0.01389649510383606, + 0.012300906702876091, + 0.008763034828007221, + 0.00690815644338727, + 0.023105408996343613, + 0.00809114333242178, + 0.006210403051227331, + 0.012768900021910667, + 0.012692935764789581, + 0.008090891875326633, + 0.011454340070486069, + 0.04667087271809578, + 0.00760508282110095, + 0.005508828442543745, + 0.007464678958058357, + 0.021715322509407997, + 0.0052040270529687405, + 0.0042226193472743034, + 0.0192145723849535, + 0.00374412490054965, + 0.002740174066275358, + 0.0031189918518066406, + 0.004763445816934109, + 0.003312859684228897, + 0.004994658753275871, + 0.002974344650283456, + 0.00398813933134079, + 0.00354032963514328, + 0.010118886828422546, + 0.005179672967642546, + 0.006801808252930641, + 0.01670469343662262, + 0.012864447198808193, + 0.01894231326878071, + 0.01286743301898241, + 0.009025555104017258, + 0.015061150304973125, + 0.027463626116514206, + 0.021570125594735146, + 0.025949867442250252, + 0.026318958029150963, + 0.031058691442012787, + 0.025721900165081024, + 0.0136271882802248, + 0.005159312393516302, + 0.007149157114326954, + 0.011633655987679958, + 0.009157349355518818, + 0.00543275848031044, + 0.010254395194351673, + 0.010039949789643288, + 0.01142313051968813, + 0.007075180299580097, + 0.006140805780887604, + 0.009767577983438969, + 0.005253922194242477, + 0.006343214772641659, + 0.018920965492725372, + 0.007073704618960619, + 0.01539934054017067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.1270836442708969 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.2520773708820343 + ], + "opposite_logprob_tokens": [ + -1.502077341079712 + ], + "logprob_reward": -0.2520773708820343, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520773708820343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07914629578590393, + 0.12571510672569275, + 0.17610031366348267, + 0.08924360573291779, + 0.08194473385810852, + 0.0710621103644371, + 0.11514472216367722, + 0.12628977000713348, + 0.05019837245345116, + 0.08515502512454987 + ], + "credit_map": [ + 0.010153274983167648, + 0.01117207296192646, + 0.009899559430778027, + 0.00734757911413908, + 0.007078825030475855, + 0.017457976937294006, + 0.006605122704058886, + 0.00943188276141882, + 0.013477056287229061, + 0.016169466078281403, + 0.03487944230437279, + 0.019923236221075058, + 0.010546978563070297, + 0.011833571828901768, + 0.007196510210633278, + 0.011688842438161373, + 0.05221978574991226, + 0.06273067742586136, + 0.007883932441473007, + 0.015116654336452484, + 0.013248845934867859, + 0.011145766824483871, + 0.007269640918821096, + 0.006484999321401119, + 0.010271682403981686, + 0.006936841178685427, + 0.004953069146722555, + 0.011294068768620491, + 0.012284928932785988, + 0.007033701986074448, + 0.009431128390133381, + 0.027038175612688065, + 0.009224645793437958, + 0.010064947418868542, + 0.008222030475735664, + 0.011855658143758774, + 0.013594569638371468, + 0.007877386175096035, + 0.014731599017977715, + 0.006373901851475239, + 0.0065000019967556, + 0.007299618795514107, + 0.010788459330797195, + 0.008139313198626041, + 0.011998907662928104, + 0.011526405811309814, + 0.008862048387527466, + 0.005947355646640062, + 0.009808181785047054, + 0.009523320011794567, + 0.008321939036250114, + 0.018751174211502075, + 0.018312787637114525, + 0.01958729699254036, + 0.014685549773275852, + 0.016154468059539795, + 0.010523607023060322, + 0.019525323063135147, + 0.017815401777625084, + 0.018061649054288864, + 0.019771696999669075, + 0.01576615869998932, + 0.01658170484006405, + 0.00824424996972084, + 0.008810371160507202, + 0.005894798319786787, + 0.01086688507348299, + 0.005351391155272722, + 0.0038588999304920435, + 0.004078221041709185, + 0.005692982580512762, + 0.005644819233566523, + 0.005780723877251148, + 0.006578588858246803, + 0.004065524321049452, + 0.00809785071760416, + 0.008723514154553413, + 0.021430248394608498, + 0.011678371578454971, + 0.018800200894474983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.2520773708820343 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.011081107892096043 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081107892096043, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081107892096043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18063484132289886, + 0.11242328584194183, + 0.06634733825922012, + 0.1667197048664093, + 0.06007139012217522, + 0.08655635267496109, + 0.04423515498638153, + 0.09828157722949982, + 0.08196645975112915, + 0.10276392102241516 + ], + "credit_map": [ + 0.024018293246626854, + 0.034770842641592026, + 0.0211993046104908, + 0.032323189079761505, + 0.020458703860640526, + 0.022526815533638, + 0.010216140188276768, + 0.015121557749807835, + 0.0157000795006752, + 0.024281218647956848, + 0.01737215928733349, + 0.010680115781724453, + 0.013377171941101551, + 0.016247188672423363, + 0.007026392035186291, + 0.007738955784589052, + 0.00847953837364912, + 0.016634004190564156, + 0.004970853216946125, + 0.007897340692579746, + 0.008420145139098167, + 0.006475140806287527, + 0.008632693439722061, + 0.004837624728679657, + 0.03835396468639374, + 0.01192169077694416, + 0.008249850943684578, + 0.015647202730178833, + 0.013306221924722195, + 0.010538140311837196, + 0.008726091124117374, + 0.05997654050588608, + 0.008190752007067204, + 0.006951980292797089, + 0.009497077204287052, + 0.008316264487802982, + 0.007564969826489687, + 0.006441445555537939, + 0.007950353436172009, + 0.00515854824334383, + 0.007085748482495546, + 0.006701712496578693, + 0.00963215809315443, + 0.008346300572156906, + 0.009048442356288433, + 0.02430765889585018, + 0.010897339321672916, + 0.010537000373005867, + 0.006864805240184069, + 0.006184044759720564, + 0.006433325354009867, + 0.006155031267553568, + 0.006050596479326487, + 0.00498013524338603, + 0.004404198843985796, + 0.003163016401231289, + 0.006782684940844774, + 0.014683444052934647, + 0.010915577411651611, + 0.018117472529411316, + 0.017044242471456528, + 0.014611180871725082, + 0.009739450179040432, + 0.006387525238096714, + 0.009307010099291801, + 0.007744588423520327, + 0.008371847681701183, + 0.008549858815968037, + 0.0072012185119092464, + 0.018098605796694756, + 0.008623696863651276, + 0.01406963262706995, + 0.009803053922951221, + 0.008418312296271324, + 0.01284001488238573, + 0.008566413074731827, + 0.0086138304322958, + 0.02629118598997593, + 0.007648929953575134, + 0.020582186058163643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.011081107892096043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "disorienting low-angle tracking shot rotates on its axis in a continuous barrel roll, keeping pace with the cyclists moving at high speed. The subject, clad in a fitted, waterproof black technical jacket that gleams beneath a steady drizzle, slumps over the handlebars of their road bike. Their face is a mask of intense concentration, partially obscured by sleek, aerodynamic glasses with transparent lenses studded with tiny water droplets. At night, they pedal frantically, their legs a powerful, rhythmic blur propelling them forward, swiftly and decisively overtaking a line of slow-moving cars on the rain-slicked city streets. The black asphalt, like a dark, liquid mirror, reflects a chaotic swirl of red taillights and the vibrant electric glow of neon storefront signs. As the camera rotates, towering buildings and streetlights merge into a dazzling vortex of light and shadow. Water sprays from the rear wheel in an exquisite arc of illumination, capturing the multicolored light and adding to the scene's dynamic quality, creating a visceral and tense atmosphere of exhilarating speed and urban danger.", + "seed": 5933, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_9816_seed_5933_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": -0.0565308315679431, + "vlm_logprob": -0.0565308315679431, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.0565308315679431, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.02690150961279869 + ], + "opposite_logprob_tokens": [ + -3.7769014835357666 + ], + "logprob_reward": -0.02690150961279869, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02690150961279869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18823248147964478, + 0.04348929971456528, + 0.07038844376802444, + 0.0826670378446579, + 0.09399507194757462, + 0.03425430878996849, + 0.10579374432563782, + 0.06133395805954933, + 0.11549418419599533, + 0.20435145497322083 + ], + "credit_map": [ + 0.018567794933915138, + 0.011508522555232048, + 0.046584755182266235, + 0.013153749518096447, + 0.016807900741696358, + 0.024400003254413605, + 0.010862698778510094, + 0.04634704813361168, + 0.00407017907127738, + 0.0045613995753228664, + 0.00705315824598074, + 0.0043281917460262775, + 0.004967496730387211, + 0.008222591131925583, + 0.003251640824601054, + 0.00703464075922966, + 0.006259089335799217, + 0.009413596242666245, + 0.006095512770116329, + 0.007083894219249487, + 0.01428927667438984, + 0.008745091035962105, + 0.007642804645001888, + 0.010859181173145771, + 0.009635819122195244, + 0.00932851992547512, + 0.010985901579260826, + 0.013685300014913082, + 0.008763102814555168, + 0.005586881190538406, + 0.010357236489653587, + 0.014324277639389038, + 0.007174041587859392, + 0.015348180197179317, + 0.009278065524995327, + 0.014250166714191437, + 0.014685011468827724, + 0.011590732261538506, + 0.01200846303254366, + 0.009660406038165092, + 0.002490621292963624, + 0.002908809343352914, + 0.0026224383618682623, + 0.004069298971444368, + 0.004254571627825499, + 0.004856426734477282, + 0.00610928563401103, + 0.006942853331565857, + 0.02173583395779133, + 0.015632323920726776, + 0.011743542738258839, + 0.013285067863762379, + 0.010724739171564579, + 0.007388676516711712, + 0.009430885314941406, + 0.0158526711165905, + 0.011071449145674706, + 0.00962673407047987, + 0.01347135379910469, + 0.006917930208146572, + 0.005201106425374746, + 0.006954600568860769, + 0.004208792001008987, + 0.003881991608068347, + 0.023441094905138016, + 0.009994381107389927, + 0.020280886441469193, + 0.023347033187747, + 0.00822994764894247, + 0.015828056260943413, + 0.005397797096520662, + 0.008974989876151085, + 0.01697380840778351, + 0.016024034470319748, + 0.015778930857777596, + 0.056849390268325806, + 0.00828789547085762, + 0.07045964151620865, + 0.0070691402070224285, + 0.012908616103231907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.02690150961279869 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.011102918535470963 + ], + "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": [ + 0.1184079572558403, + 0.052692778408527374, + 0.0620403029024601, + 0.10744569450616837, + 0.11453695595264435, + 0.10749392956495285, + 0.09797364473342896, + 0.03891899064183235, + 0.10527976602315903, + 0.19520996510982513 + ], + "credit_map": [ + 0.011257333680987358, + 0.011368934065103531, + 0.024171646684408188, + 0.0076336027123034, + 0.00815271120518446, + 0.01670614629983902, + 0.01111025270074606, + 0.02800733596086502, + 0.006701047532260418, + 0.007591527421027422, + 0.009142250753939152, + 0.005668262485414743, + 0.005472998134791851, + 0.0061242287047207355, + 0.005101282149553299, + 0.006891180295497179, + 0.008463814854621887, + 0.013921869918704033, + 0.006899488624185324, + 0.006104319356381893, + 0.006661644671112299, + 0.004934841766953468, + 0.005984695628285408, + 0.00906963087618351, + 0.012663407251238823, + 0.012002392672002316, + 0.01426244992762804, + 0.017914026975631714, + 0.011437280103564262, + 0.00759462034329772, + 0.013062916696071625, + 0.01850860007107258, + 0.009279538877308369, + 0.01655706949532032, + 0.010877602733671665, + 0.016857551410794258, + 0.019527049735188484, + 0.011835819110274315, + 0.01711367629468441, + 0.01248865108937025, + 0.009019659832119942, + 0.006405878812074661, + 0.007300970144569874, + 0.012023995630443096, + 0.013667694292962551, + 0.00863430742174387, + 0.03822760656476021, + 0.012213817797601223, + 0.021680977195501328, + 0.01462994422763586, + 0.011014560237526894, + 0.013828746043145657, + 0.008731450885534286, + 0.005828607361763716, + 0.007448816671967506, + 0.014810544438660145, + 0.006055513862520456, + 0.005012590903788805, + 0.0054769194684922695, + 0.0064497753046453, + 0.003569883294403553, + 0.004295332357287407, + 0.004082933068275452, + 0.003976041916757822, + 0.023081019520759583, + 0.010207769460976124, + 0.01313434075564146, + 0.03024168498814106, + 0.005462854169309139, + 0.009912200272083282, + 0.0051467870362102985, + 0.00809311680495739, + 0.014209259301424026, + 0.01963351108133793, + 0.014221456833183765, + 0.04592536762356758, + 0.015431728214025497, + 0.04751664400100708, + 0.01754276640713215, + 0.02072923071682453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "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 scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.12708574533462524 + ], + "opposite_logprob_tokens": [ + -2.1270856857299805 + ], + "logprob_reward": -0.12708574533462524, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708574533462524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05049196630716324, + 0.01993725076317787, + 0.03381191939115524, + 0.23912367224693298, + 0.034112926572561264, + 0.13743151724338531, + 0.04212512820959091, + 0.018907692283391953, + 0.041440919041633606, + 0.3826170265674591 + ], + "credit_map": [ + 0.0036796003114432096, + 0.005553212948143482, + 0.007002844009548426, + 0.00624387152493, + 0.005061106290668249, + 0.008186228573322296, + 0.007562898099422455, + 0.007202201057225466, + 0.002093761693686247, + 0.003778384067118168, + 0.0033346025738865137, + 0.0023445598781108856, + 0.0016910814447328448, + 0.0024687701370567083, + 0.0017954485956579447, + 0.0024306399282068014, + 0.003220419865101576, + 0.005234277807176113, + 0.0033648924436420202, + 0.003522042417898774, + 0.004534565843641758, + 0.003582216100767255, + 0.0036209605168551207, + 0.006732540670782328, + 0.02624821849167347, + 0.026293840259313583, + 0.030394265428185463, + 0.037719570100307465, + 0.027551576495170593, + 0.017017951235175133, + 0.029715580865740776, + 0.04418264701962471, + 0.0057019987143576145, + 0.004027131479233503, + 0.003216126235201955, + 0.004203634802252054, + 0.004478820133954287, + 0.004665290005505085, + 0.003670822363346815, + 0.0041491021402180195, + 0.012623117305338383, + 0.005401439033448696, + 0.02156832069158554, + 0.010708407498896122, + 0.014900881797075272, + 0.009504873305559158, + 0.04647977277636528, + 0.016244681552052498, + 0.0049706436693668365, + 0.004312844946980476, + 0.0037019795272499323, + 0.006198406219482422, + 0.00603469368070364, + 0.006116076372563839, + 0.0049242740496993065, + 0.005866204854100943, + 0.0021234997548162937, + 0.002090852241963148, + 0.002248193370178342, + 0.0024405582807958126, + 0.0033585610799491405, + 0.001977709122002125, + 0.002894229954108596, + 0.0017740859184414148, + 0.006363658234477043, + 0.004119289107620716, + 0.004660122096538544, + 0.009068189188838005, + 0.006011238321661949, + 0.003983411006629467, + 0.003006518818438053, + 0.004228488076478243, + 0.019218645989894867, + 0.0310710109770298, + 0.02393193356692791, + 0.11425424367189407, + 0.016912635415792465, + 0.13156427443027496, + 0.018640536814928055, + 0.027023719623684883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.12708574533462524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.11292393505573273 + ], + "opposite_logprob_tokens": [ + -2.2379238605499268 + ], + "logprob_reward": -0.11292393505573273, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292393505573273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04951747879385948, + 0.050560902804136276, + 0.04999290406703949, + 0.22473038733005524, + 0.06711429357528687, + 0.20420043170452118, + 0.04529574513435364, + 0.044981468468904495, + 0.05297607555985451, + 0.2106303572654724 + ], + "credit_map": [ + 0.007273347582668066, + 0.004012060351669788, + 0.0072526754811406136, + 0.005003807134926319, + 0.004173360764980316, + 0.008718973025679588, + 0.004465519916266203, + 0.008617731742560863, + 0.004683734383434057, + 0.007854270748794079, + 0.008802690543234348, + 0.004202821757644415, + 0.005163620691746473, + 0.00714262155815959, + 0.00455475552007556, + 0.008156388998031616, + 0.005740134511142969, + 0.008856662549078465, + 0.006026997696608305, + 0.005541305523365736, + 0.006630007643252611, + 0.0038322838954627514, + 0.005960777867585421, + 0.007404733449220657, + 0.025172656401991844, + 0.02614903263747692, + 0.02898762933909893, + 0.03400981426239014, + 0.025893718004226685, + 0.015419079922139645, + 0.028301605954766273, + 0.04079684615135193, + 0.009188137017190456, + 0.010297579690814018, + 0.006894456222653389, + 0.009335682727396488, + 0.007121601141989231, + 0.00834349449723959, + 0.008548391051590443, + 0.007384949363768101, + 0.01274732407182455, + 0.016181811690330505, + 0.011067807674407959, + 0.023163944482803345, + 0.029616517946124077, + 0.022855516523122787, + 0.058380376547575, + 0.030187122523784637, + 0.005209149327129126, + 0.0036287358961999416, + 0.003654302330687642, + 0.01028499472886324, + 0.005166003946214914, + 0.004850484896451235, + 0.006650052964687347, + 0.005852020811289549, + 0.007545014843344688, + 0.004406921099871397, + 0.010235763154923916, + 0.006224237848073244, + 0.002725359285250306, + 0.005795195698738098, + 0.003297582268714905, + 0.004751395434141159, + 0.011678672395646572, + 0.005236255470663309, + 0.007577542681246996, + 0.00912074837833643, + 0.004533914849162102, + 0.00502789905294776, + 0.003521401435136795, + 0.006279645953327417, + 0.012896982952952385, + 0.016928723081946373, + 0.01450746413320303, + 0.05922331660985947, + 0.008633526042103767, + 0.07584276795387268, + 0.009049348533153534, + 0.013548229821026325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.11292393505573273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on rain-slicked city streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject wearing a fitted, waterproof black technical jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject slumped over the handlebars of a road bike while pedaling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a continuous barrel roll while tracking the cyclists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red taillights and neon storefront signs reflected on the black 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": [ + -0.004640049301087856 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640049301087856, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640049301087856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2658873498439789, + 0.045897796750068665, + 0.10937739163637161, + 0.04636293277144432, + 0.07022670656442642, + 0.0986737385392189, + 0.10238132625818253, + 0.05934397876262665, + 0.09418316930532455, + 0.10766564309597015 + ], + "credit_map": [ + 0.01961132138967514, + 0.015101131051778793, + 0.05169890448451042, + 0.017775753512978554, + 0.021307412534952164, + 0.055019717663526535, + 0.015600258484482765, + 0.06977283954620361, + 0.005144122522324324, + 0.005045673809945583, + 0.00666244188323617, + 0.0038515988271683455, + 0.006141521967947483, + 0.006645821034908295, + 0.0038320780731737614, + 0.008574542589485645, + 0.010665343143045902, + 0.014530054293572903, + 0.00907056126743555, + 0.009804186411201954, + 0.019632479175925255, + 0.011764195747673512, + 0.011728989891707897, + 0.02218157984316349, + 0.006918143946677446, + 0.006306328810751438, + 0.007181735243648291, + 0.00852553453296423, + 0.004653139039874077, + 0.0031418625731021166, + 0.0039825295098125935, + 0.005653657019138336, + 0.010449392721056938, + 0.009860733523964882, + 0.007795242592692375, + 0.00988913606852293, + 0.008589883334934711, + 0.009507767856121063, + 0.006957094185054302, + 0.00717745628207922, + 0.008405577391386032, + 0.007949607446789742, + 0.009417545050382614, + 0.009008679538965225, + 0.011330620385706425, + 0.007724904455244541, + 0.03194275125861168, + 0.012894055806100368, + 0.01919008046388626, + 0.013587054796516895, + 0.011147670447826385, + 0.01644463837146759, + 0.009109401144087315, + 0.005582279060035944, + 0.010709084570407867, + 0.016611110419034958, + 0.010731387883424759, + 0.007869427092373371, + 0.012271028012037277, + 0.009770131669938564, + 0.005441313609480858, + 0.004677312448620796, + 0.004824423231184483, + 0.0037589550483971834, + 0.019830014556646347, + 0.007809552364051342, + 0.015708452090620995, + 0.02232680469751358, + 0.005762950051575899, + 0.01038566417992115, + 0.004881777800619602, + 0.007477957289665937, + 0.009265867061913013, + 0.01234510075300932, + 0.00959912408143282, + 0.027349932119250298, + 0.007241467013955116, + 0.024537494406104088, + 0.007087682373821735, + 0.010238973423838615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.004640049301087856 + } + ], + "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/0000025/group_11/index_tensor([5747, 5747])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_11/index_tensor([5747, 5747])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a6873c4b41a725cb47f57adbdcde344f7c512f01 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_11/index_tensor([5747, 5747])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion close-up, the camera panning closely to the action, focuses on a warrior clad in weathered steel armor. In the dim light, the armor, covered in fine scratches and slight rust, reflects a cold metallic luster. His muscles are tense, and across his chiseled face, sweat and dust mingle, flowing down his furrowed brow and tense jaw. Wisps of dark hair fall from the edge of his helmet, plastering to his damp forehead. His eyes are wide, filled with determination and savage power, as he swings a broadsword, also scarred by battle, from right to left with all his might. In slow motion, the sword's edge slices through the air, creating a visible, distorted current, blurring the background into a blurry, dark-toned chaos, with faint traces of flying dust visible. The entire picture presents a faded retro film texture, with extremely low color saturation, showing a sepia tone washed by time, and is covered with obvious film grain, fine scratches and occasional flashes of light, creating a tense, primitive and epic atmosphere.", + "seed": 9372, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_11/index_5747_seed_9372_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": -0.10827606692910194, + "vlm_logprob": -0.10827606692910194, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.10827606692910194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022815579548478127 + ], + "opposite_logprob_tokens": [ + -4.147815704345703 + ], + "logprob_reward": -0.022815579548478127, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022815579548478127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2703039050102234, + 0.066989466547966, + 0.03484470769762993, + 0.03142453730106354, + 0.0857146829366684, + 0.08661474287509918, + 0.04414645582437515, + 0.03362993523478508, + 0.24498596787452698, + 0.10134562104940414 + ], + "credit_map": [ + 0.011267720721662045, + 0.015835706144571304, + 0.01130607258528471, + 0.011916391551494598, + 0.09719505906105042, + 0.07497569173574448, + 0.03277307003736496, + 0.015034202486276627, + 0.006132545415312052, + 0.008879439905285835, + 0.005260462872684002, + 0.005743463523685932, + 0.00935010053217411, + 0.01379189733415842, + 0.00938570499420166, + 0.008445856161415577, + 0.0033197528682649136, + 0.004248949233442545, + 0.0030397481750696898, + 0.003409093711525202, + 0.004853335674852133, + 0.005389899481087923, + 0.005055180285125971, + 0.005528748035430908, + 0.004296277184039354, + 0.006074964534491301, + 0.0025899570900946856, + 0.0031969405245035887, + 0.006183863617479801, + 0.0035281076561659575, + 0.002636965364217758, + 0.002917459001764655, + 0.006494531407952309, + 0.01694268174469471, + 0.012794890441000462, + 0.009356656111776829, + 0.013857688754796982, + 0.012232772074639797, + 0.006882146000862122, + 0.007153315935283899, + 0.004648232366889715, + 0.004714084789156914, + 0.009562925435602665, + 0.004286276176571846, + 0.007822884246706963, + 0.02627653256058693, + 0.005626918748021126, + 0.023676887154579163, + 0.0064889369532465935, + 0.0063987066969275475, + 0.004314300138503313, + 0.004680220037698746, + 0.0066745467483997345, + 0.007460716180503368, + 0.003629193874076009, + 0.004499835893511772, + 0.0049625602550804615, + 0.007978502660989761, + 0.003669579979032278, + 0.0033554164692759514, + 0.002921492327004671, + 0.003279010998085141, + 0.004238039255142212, + 0.0032253344543278217, + 0.026332475244998932, + 0.03730647265911102, + 0.017538823187351227, + 0.021884053945541382, + 0.04067673906683922, + 0.046842947602272034, + 0.034203216433525085, + 0.020201247185468674, + 0.01580657809972763, + 0.005881412886083126, + 0.005748980212956667, + 0.009134747087955475, + 0.025832198560237885, + 0.013727639801800251, + 0.01523625012487173, + 0.009977820329368114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.022815579548478127 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34864354133605957 + ], + "opposite_logprob_tokens": [ + -1.2236435413360596 + ], + "logprob_reward": -0.34864354133605957, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34864354133605957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12332656979560852, + 0.11909162998199463, + 0.06045292317867279, + 0.03751753643155098, + 0.03328561782836914, + 0.21076253056526184, + 0.11157345026731491, + 0.12848302721977234, + 0.07599513232707977, + 0.0995115116238594 + ], + "credit_map": [ + 0.00846731849014759, + 0.007980681024491787, + 0.007517328951507807, + 0.0079634515568614, + 0.03721727803349495, + 0.03049408085644245, + 0.010374976322054863, + 0.013311462476849556, + 0.011951104737818241, + 0.020744113251566887, + 0.007867156527936459, + 0.010144130326807499, + 0.017315659672021866, + 0.022618215531110764, + 0.013755140826106071, + 0.014696118421852589, + 0.006007789634168148, + 0.00683543412014842, + 0.005117308348417282, + 0.006709638517349958, + 0.009903071448206902, + 0.008099607191979885, + 0.009252329356968403, + 0.008527747355401516, + 0.0049633486196398735, + 0.007158922962844372, + 0.00459061935544014, + 0.005045188590884209, + 0.0051201460883021355, + 0.0038002615328878164, + 0.003049589693546295, + 0.00378946284763515, + 0.0025915533769875765, + 0.006242420058697462, + 0.006368242669850588, + 0.002603310626000166, + 0.007621722761541605, + 0.004172082990407944, + 0.001855816226452589, + 0.0018304692348465323, + 0.023166056722402573, + 0.017445499077439308, + 0.02766512706875801, + 0.0122710932046175, + 0.021963637322187424, + 0.04163968563079834, + 0.01625172793865204, + 0.05035972222685814, + 0.02019745297729969, + 0.023982448503375053, + 0.010269679129123688, + 0.009175349958240986, + 0.015366297215223312, + 0.020306775346398354, + 0.006354350596666336, + 0.0059211039915680885, + 0.012996600940823555, + 0.015450694598257542, + 0.011476661078631878, + 0.010862229391932487, + 0.023760680109262466, + 0.023811737075448036, + 0.017746763303875923, + 0.01237767655402422, + 0.008714036084711552, + 0.013560405001044273, + 0.0050886692479252815, + 0.005010603927075863, + 0.012399441562592983, + 0.013793994672596455, + 0.011414551176130772, + 0.006013436242938042, + 0.018389694392681122, + 0.005863249767571688, + 0.004037749487906694, + 0.007217801176011562, + 0.03305350989103317, + 0.011701309122145176, + 0.011338847689330578, + 0.007909361273050308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.34864354133605957 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129235103726387 + ], + "opposite_logprob_tokens": [ + -2.2379236221313477 + ], + "logprob_reward": -0.1129235103726387, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129235103726387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18095098435878754, + 0.09146557003259659, + 0.08096542954444885, + 0.05287131294608116, + 0.06747180223464966, + 0.20901258289813995, + 0.07850446552038193, + 0.05816669017076492, + 0.07205050438642502, + 0.10854063928127289 + ], + "credit_map": [ + 0.015570361167192459, + 0.026038674637675285, + 0.014016752131283283, + 0.01605870947241783, + 0.031420476734638214, + 0.03034657984972, + 0.0214428398758173, + 0.02605658397078514, + 0.007551763206720352, + 0.013674202375113964, + 0.006262432783842087, + 0.011397257447242737, + 0.014490931294858456, + 0.016465382650494576, + 0.010952913202345371, + 0.010670684278011322, + 0.006439628545194864, + 0.008980284444987774, + 0.006315396632999182, + 0.007795721758157015, + 0.014897732064127922, + 0.013542748987674713, + 0.0123890470713377, + 0.010604866780340672, + 0.007729612290859222, + 0.010802419856190681, + 0.004237579647451639, + 0.006394645199179649, + 0.008708388544619083, + 0.005391060374677181, + 0.004544251598417759, + 0.00506335636600852, + 0.005432284437119961, + 0.012879864312708378, + 0.00889999233186245, + 0.007504007313400507, + 0.013662705197930336, + 0.008289571851491928, + 0.006066511385142803, + 0.00473686633631587, + 0.016216041520237923, + 0.01260329969227314, + 0.01433593686670065, + 0.015996841713786125, + 0.030451452359557152, + 0.051407791674137115, + 0.02850041538476944, + 0.03950079530477524, + 0.01160570327192545, + 0.014880523085594177, + 0.0056031388230621815, + 0.009927879087626934, + 0.011292732320725918, + 0.011498832143843174, + 0.007306940853595734, + 0.0063887122087180614, + 0.007241578307002783, + 0.009294805116951466, + 0.005614745896309614, + 0.004110358189791441, + 0.007220817729830742, + 0.00989589188247919, + 0.00682402728125453, + 0.007964466698467731, + 0.008287540636956692, + 0.014346328563988209, + 0.006245652213692665, + 0.010384052991867065, + 0.009068929590284824, + 0.00905574019998312, + 0.008414529263973236, + 0.00624772859737277, + 0.014657828956842422, + 0.009569797664880753, + 0.008589453995227814, + 0.01059215422719717, + 0.017627937719225883, + 0.019376695156097412, + 0.016791898757219315, + 0.011334875598549843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.1129235103726387 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329937182366848 + ], + "opposite_logprob_tokens": [ + -3.7732994556427 + ], + "logprob_reward": -0.02329937182366848, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329937182366848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.385640025138855, + 0.06440219283103943, + 0.04323795810341835, + 0.03829190880060196, + 0.05293535441160202, + 0.09417527914047241, + 0.060782305896282196, + 0.052072104066610336, + 0.09461797028779984, + 0.11384496092796326 + ], + "credit_map": [ + 0.030244776979088783, + 0.07118205726146698, + 0.03710634633898735, + 0.03926357999444008, + 0.05798524245619774, + 0.04855813831090927, + 0.06425904482603073, + 0.037040825933218, + 0.006283158902078867, + 0.010435990057885647, + 0.006848826073110104, + 0.006784002762287855, + 0.008675510995090008, + 0.01138884387910366, + 0.006452054716646671, + 0.007533809170126915, + 0.003834580071270466, + 0.005128087010234594, + 0.004339409992098808, + 0.0040239919908344746, + 0.008548843674361706, + 0.007213394157588482, + 0.004976851865649223, + 0.005172797478735447, + 0.0058090901002287865, + 0.008125495165586472, + 0.003242135513573885, + 0.004030429758131504, + 0.005939853843301535, + 0.0039216093719005585, + 0.0032511225435882807, + 0.003972173202782869, + 0.0043161287903785706, + 0.009548241272568703, + 0.006110391579568386, + 0.004782395903021097, + 0.014207780361175537, + 0.006413496099412441, + 0.004051747266203165, + 0.0035051717422902584, + 0.007284120190888643, + 0.006079765968024731, + 0.005302450153976679, + 0.00445218663662672, + 0.01232664193958044, + 0.022867856547236443, + 0.0090373195707798, + 0.026824945583939552, + 0.008999691344797611, + 0.010539192706346512, + 0.0043611773289740086, + 0.003793169744312763, + 0.013500242494046688, + 0.008938443847000599, + 0.005343038123100996, + 0.005307354032993317, + 0.005651064217090607, + 0.0058848499320447445, + 0.004807451274245977, + 0.0050387512892484665, + 0.009380306117236614, + 0.009565112181007862, + 0.005967971868813038, + 0.005776598583906889, + 0.011914966627955437, + 0.024333039298653603, + 0.00738117890432477, + 0.007826009765267372, + 0.009886123239994049, + 0.014620831236243248, + 0.011196600273251534, + 0.007459220942109823, + 0.014550318941473961, + 0.0083461944013834, + 0.007932249456644058, + 0.01413771416991949, + 0.0197713915258646, + 0.01728740707039833, + 0.018184766173362732, + 0.01363491453230381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.02329937182366848 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369833156466484 + ], + "opposite_logprob_tokens": [ + -3.408698320388794 + ], + "logprob_reward": -0.03369833156466484, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369833156466484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09626276791095734, + 0.07705007493495941, + 0.07087131589651108, + 0.04986313357949257, + 0.09880389273166656, + 0.34042373299598694, + 0.08093016594648361, + 0.09768524020910263, + 0.052577029913663864, + 0.03553264960646629 + ], + "credit_map": [ + 0.012639818713068962, + 0.01173426117748022, + 0.009599407203495502, + 0.0088090430945158, + 0.012211417779326439, + 0.014223919250071049, + 0.012122485786676407, + 0.014922413975000381, + 0.007464100606739521, + 0.011734615080058575, + 0.005819085985422134, + 0.0085384426638484, + 0.009512564167380333, + 0.013940330594778061, + 0.013219330459833145, + 0.006821607239544392, + 0.006426628213375807, + 0.008761327713727951, + 0.008151836693286896, + 0.005882308352738619, + 0.014346235431730747, + 0.010865302756428719, + 0.008533207699656487, + 0.007904470898211002, + 0.00784310419112444, + 0.00813788641244173, + 0.004053981509059668, + 0.00531811872497201, + 0.00877341441810131, + 0.004959548823535442, + 0.005627221427857876, + 0.00514985853806138, + 0.006337882485240698, + 0.020174862816929817, + 0.011998141184449196, + 0.006521472241729498, + 0.027378955855965614, + 0.012665131129324436, + 0.007564385421574116, + 0.0061630671843886375, + 0.01592192053794861, + 0.0195165928453207, + 0.02146161161363125, + 0.015987953171133995, + 0.02945285476744175, + 0.1028137058019638, + 0.020006882026791573, + 0.11526218801736832, + 0.01687837392091751, + 0.01264486089348793, + 0.00719537353143096, + 0.006239095702767372, + 0.007658911403268576, + 0.01697724685072899, + 0.0061111776158213615, + 0.007225125562399626, + 0.012301878072321415, + 0.017922818660736084, + 0.009178275242447853, + 0.00809542927891016, + 0.011347580701112747, + 0.01665128953754902, + 0.013126825913786888, + 0.009061135351657867, + 0.00624444056302309, + 0.004358076024800539, + 0.004238548688590527, + 0.004336593672633171, + 0.00790292676538229, + 0.011530553922057152, + 0.00847822055220604, + 0.005487670190632343, + 0.005277096293866634, + 0.0020671600941568613, + 0.003024633042514324, + 0.004163587000221014, + 0.009067446924746037, + 0.00393633171916008, + 0.0050290729850530624, + 0.0029673236422240734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.03369833156466484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up, the camera panning closely to the action, focuses on a warrior clad in weathered steel armor. In the dim light, the armor, covered in fine scratches and slight rust, reflects a cold metallic luster. His muscles are tense, and across his chiseled face, sweat and dust mingle, flowing down his furrowed brow and tense jaw. Wisps of dark hair fall from the edge of his helmet, plastering to his damp forehead. His eyes are wide, filled with determination and savage power, as he swings a broadsword, also scarred by battle, from right to left with all his might. In slow motion, the sword's edge slices through the air, creating a visible, distorted current, blurring the background into a blurry, dark-toned chaos, with faint traces of flying dust visible. The entire picture presents a faded retro film texture, with extremely low color saturation, showing a sepia tone washed by time, and is covered with obvious film grain, fine scratches and occasional flashes of light, creating a tense, primitive and epic atmosphere.", + "seed": 9373, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_11/index_5747_seed_9373_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": -0.1622741725295782, + "vlm_logprob": -0.1622741725295782, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.1622741725295782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01850694604218006 + ], + "opposite_logprob_tokens": [ + -4.39350700378418 + ], + "logprob_reward": -0.01850694604218006, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01850694604218006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19886773824691772, + 0.09301529824733734, + 0.04652266949415207, + 0.04297782853245735, + 0.054415591061115265, + 0.1521756649017334, + 0.08701249957084656, + 0.16678905487060547, + 0.06258067488670349, + 0.09564296156167984 + ], + "credit_map": [ + 0.024490347132086754, + 0.03587310388684273, + 0.022844571620225906, + 0.02424752339720726, + 0.03569206967949867, + 0.022234754636883736, + 0.017663542181253433, + 0.01582181639969349, + 0.010961952619254589, + 0.01611291617155075, + 0.007833865471184254, + 0.014836102724075317, + 0.011123694479465485, + 0.01318899355828762, + 0.010359571315348148, + 0.0085982047021389, + 0.0067246356047689915, + 0.007739246357232332, + 0.005777050741016865, + 0.004512004088610411, + 0.005105564370751381, + 0.006197439972311258, + 0.004540611989796162, + 0.005926123820245266, + 0.0043249111622571945, + 0.00467579485848546, + 0.0041802129708230495, + 0.0060117789544165134, + 0.008400479331612587, + 0.0049360222183167934, + 0.0042081959545612335, + 0.006240436807274818, + 0.005318585783243179, + 0.005687411408871412, + 0.0071190460585057735, + 0.006448057945817709, + 0.009720887988805771, + 0.007012272719293833, + 0.007515340577811003, + 0.005593989044427872, + 0.03612172231078148, + 0.03876897692680359, + 0.015238305553793907, + 0.012552314437925816, + 0.01036342978477478, + 0.01631789281964302, + 0.012387175112962723, + 0.010425855405628681, + 0.007453962229192257, + 0.008974524214863777, + 0.013264580629765987, + 0.015051458030939102, + 0.008253955282270908, + 0.009699089452624321, + 0.009163117967545986, + 0.015151821076869965, + 0.014888638630509377, + 0.015148084610700607, + 0.023018695414066315, + 0.024779899045825005, + 0.020692458376288414, + 0.030759388580918312, + 0.018796712160110474, + 0.01870518922805786, + 0.007436033338308334, + 0.00634576054289937, + 0.006664748769253492, + 0.0068858303129673, + 0.010708030313253403, + 0.007749401032924652, + 0.011311381123960018, + 0.0054794903844594955, + 0.007293986156582832, + 0.010831236839294434, + 0.012314476072788239, + 0.012263125739991665, + 0.012488320469856262, + 0.01441938430070877, + 0.00930327083915472, + 0.016729166731238365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.01850694604218006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289015829563141 + ], + "opposite_logprob_tokens": [ + -1.0539015531539917 + ], + "logprob_reward": -0.4289015829563141, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289015829563141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08171986043453217, + 0.33680298924446106, + 0.07407539337873459, + 0.06895096600055695, + 0.03576658293604851, + 0.1443376988172531, + 0.031067214906215668, + 0.07704007625579834, + 0.0707644447684288, + 0.07947475463151932 + ], + "credit_map": [ + 0.010734143666923046, + 0.011077804490923882, + 0.011727829463779926, + 0.007686791010200977, + 0.012347790412604809, + 0.010085255838930607, + 0.009791634045541286, + 0.008268613368272781, + 0.03322042152285576, + 0.08107014000415802, + 0.02910269796848297, + 0.045601777732372284, + 0.04602542519569397, + 0.049790725111961365, + 0.02624306082725525, + 0.025748731568455696, + 0.009099210612475872, + 0.015790320932865143, + 0.006386057008057833, + 0.0067269569262862206, + 0.00784066691994667, + 0.009092445485293865, + 0.008614697493612766, + 0.010525036603212357, + 0.007536767516285181, + 0.006518838461488485, + 0.007122404873371124, + 0.009065752848982811, + 0.008204248733818531, + 0.00856288056820631, + 0.00737554207444191, + 0.014564534649252892, + 0.003857214003801346, + 0.002528612269088626, + 0.00843470823019743, + 0.006185575854033232, + 0.003762425621971488, + 0.004726390354335308, + 0.0038945702835917473, + 0.0023770881816744804, + 0.03787245601415634, + 0.017507128417491913, + 0.01566292904317379, + 0.009992535226047039, + 0.009583993814885616, + 0.029479026794433594, + 0.011413177475333214, + 0.012826457619667053, + 0.0031596263870596886, + 0.0025967108085751534, + 0.00577149773016572, + 0.004702894948422909, + 0.0027024527080357075, + 0.0035518223885446787, + 0.002552617806941271, + 0.006029592826962471, + 0.006226806901395321, + 0.00687258318066597, + 0.011653480120003223, + 0.013875446282327175, + 0.007306327112019062, + 0.01218027900904417, + 0.009275426156818867, + 0.009649725630879402, + 0.006884423084557056, + 0.009105343371629715, + 0.011206231079995632, + 0.008651470765471458, + 0.007559237536042929, + 0.007973367348313332, + 0.01223884429782629, + 0.0071455263532698154, + 0.00559203652665019, + 0.01018008403480053, + 0.01095996331423521, + 0.008212769404053688, + 0.015836358070373535, + 0.010976792313158512, + 0.008173830807209015, + 0.009542924351990223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.4289015829563141 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813218832015991 + ], + "opposite_logprob_tokens": [ + -1.4063218832015991 + ], + "logprob_reward": -0.2813218832015991, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813218832015991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13319553434848785, + 0.06845079362392426, + 0.04836515709757805, + 0.036586590111255646, + 0.11460630595684052, + 0.0923217386007309, + 0.13930727541446686, + 0.1694263517856598, + 0.08222779631614685, + 0.11551238596439362 + ], + "credit_map": [ + 0.0135260708630085, + 0.01560153253376484, + 0.01898338831961155, + 0.011983056552708149, + 0.029787380248308182, + 0.016788870096206665, + 0.016970327123999596, + 0.009554905816912651, + 0.007049340754747391, + 0.011341004632413387, + 0.007803498767316341, + 0.010994386859238148, + 0.009159835055470467, + 0.008066382259130478, + 0.00705724349245429, + 0.006979098543524742, + 0.004706513602286577, + 0.010682603344321251, + 0.004149114713072777, + 0.004772558342665434, + 0.006808941718190908, + 0.005583019461482763, + 0.0053445687517523766, + 0.0063178399577736855, + 0.003740814048796892, + 0.006599729880690575, + 0.0035815052688121796, + 0.004341276362538338, + 0.007526052184402943, + 0.0038326021749526262, + 0.0036437888629734516, + 0.0033208192326128483, + 0.009619297459721565, + 0.012238392606377602, + 0.013502037152647972, + 0.01348815206438303, + 0.019572172313928604, + 0.013176771812140942, + 0.019044289365410805, + 0.013965190388262272, + 0.019588300958275795, + 0.010230532847344875, + 0.010615591891109943, + 0.006569492164999247, + 0.007811075542122126, + 0.02136581763625145, + 0.008648516610264778, + 0.007492417003959417, + 0.01629987545311451, + 0.018178889527916908, + 0.020622404292225838, + 0.01957152970135212, + 0.013535086996853352, + 0.01806945540010929, + 0.01712094247341156, + 0.015909096226096153, + 0.01488560251891613, + 0.0188199020922184, + 0.020337900146842003, + 0.02234845794737339, + 0.02135273441672325, + 0.028963860124349594, + 0.023114247247576714, + 0.01960364542901516, + 0.010001115500926971, + 0.010542037896811962, + 0.008548444136977196, + 0.007386788260191679, + 0.011514049023389816, + 0.014998680911958218, + 0.01216078270226717, + 0.007075896952301264, + 0.007986835204064846, + 0.015210514888167381, + 0.014493808150291443, + 0.01627214439213276, + 0.02079709805548191, + 0.016810722649097443, + 0.009842348285019398, + 0.014098916202783585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.2813218832015991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059328556060791 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.02059328556060791, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059328556060791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13614881038665771, + 0.07148328423500061, + 0.027910254895687103, + 0.04264644905924797, + 0.06820548325777054, + 0.2074076533317566, + 0.07085391134023666, + 0.13790778815746307, + 0.147437185049057, + 0.08999918401241302 + ], + "credit_map": [ + 0.015931501984596252, + 0.017223138362169266, + 0.016846243292093277, + 0.012142494320869446, + 0.029038462787866592, + 0.015054736286401749, + 0.015153221786022186, + 0.014759006910026073, + 0.00703393341973424, + 0.013850950635969639, + 0.007029821630567312, + 0.008997898548841476, + 0.008258199319243431, + 0.010162286460399628, + 0.01004759781062603, + 0.006102595943957567, + 0.0021496950648725033, + 0.0037750331684947014, + 0.0031533821020275354, + 0.0028459406457841396, + 0.003368576057255268, + 0.002623786684125662, + 0.003997580148279667, + 0.00599626312032342, + 0.004718255717307329, + 0.0067029413767158985, + 0.004535273648798466, + 0.005107061471790075, + 0.007806581445038319, + 0.004046850837767124, + 0.004486167803406715, + 0.005243320018053055, + 0.007438627537339926, + 0.005480961408466101, + 0.011429146863520145, + 0.012607326731085777, + 0.00845502968877554, + 0.008769196458160877, + 0.008264477364718914, + 0.005760715343058109, + 0.05207555368542671, + 0.03711281716823578, + 0.02100900188088417, + 0.01695014350116253, + 0.013923035003244877, + 0.030835701152682304, + 0.01805788278579712, + 0.017443524673581123, + 0.010117800906300545, + 0.005802135914564133, + 0.010623381473124027, + 0.0117591368034482, + 0.006324526853859425, + 0.007421096786856651, + 0.008551431819796562, + 0.010254395194351673, + 0.0114901727065444, + 0.011058970354497433, + 0.02283240295946598, + 0.02813846804201603, + 0.013340906240046024, + 0.01433505117893219, + 0.021093329414725304, + 0.01561848446726799, + 0.01836002990603447, + 0.012106423266232014, + 0.01818891242146492, + 0.01928933896124363, + 0.02378256991505623, + 0.019077828153967857, + 0.02671978436410427, + 0.00991229061037302, + 0.007778811268508434, + 0.010448961518704891, + 0.008465852588415146, + 0.015355881303548813, + 0.013193895108997822, + 0.01340966671705246, + 0.008675090968608856, + 0.012671014294028282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.02059328556060791 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062047164887189865 + ], + "opposite_logprob_tokens": [ + -2.812047243118286 + ], + "logprob_reward": -0.062047164887189865, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062047164887189865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10099557042121887, + 0.08318426460027695, + 0.07339195162057877, + 0.06615198403596878, + 0.12626758217811584, + 0.22393998503684998, + 0.04810198023915291, + 0.07547233253717422, + 0.07874248176813126, + 0.12375187873840332 + ], + "credit_map": [ + 0.012427926063537598, + 0.020354026928544044, + 0.01105230301618576, + 0.011996493674814701, + 0.016186203807592392, + 0.010646489448845387, + 0.011487968266010284, + 0.0068441652692854404, + 0.007463766261935234, + 0.01610037498176098, + 0.006567277014255524, + 0.015459597110748291, + 0.00974471215158701, + 0.009920732118189335, + 0.01085757464170456, + 0.007070229388773441, + 0.006955210585147142, + 0.009498104453086853, + 0.011054114438593388, + 0.008036896586418152, + 0.00958974938839674, + 0.008456673473119736, + 0.01038717944175005, + 0.00941401906311512, + 0.00862949714064598, + 0.006663792300969362, + 0.007171493023633957, + 0.00656162342056632, + 0.013614479452371597, + 0.006981082260608673, + 0.007600588258355856, + 0.008929429575800896, + 0.017258064821362495, + 0.010357275605201721, + 0.02046581357717514, + 0.016871098428964615, + 0.015448122285306454, + 0.01761980727314949, + 0.014383338391780853, + 0.01386405248194933, + 0.01691201515495777, + 0.06905215233564377, + 0.0117882639169693, + 0.014719071798026562, + 0.011527230963110924, + 0.0770665630698204, + 0.009589536115527153, + 0.013285144232213497, + 0.005957945249974728, + 0.004971109330654144, + 0.008165177889168262, + 0.0055494424887001514, + 0.0038441598881036043, + 0.0054816133342683315, + 0.004823884926736355, + 0.009308649227023125, + 0.00652362871915102, + 0.007866894826292992, + 0.008494211360812187, + 0.009112129919230938, + 0.010546108707785606, + 0.013488024473190308, + 0.011371226981282234, + 0.008070111274719238, + 0.0101465517655015, + 0.006954798009246588, + 0.013122811913490295, + 0.010105959139764309, + 0.007935389876365662, + 0.012026575393974781, + 0.010488626547157764, + 0.007961775176227093, + 0.008386978879570961, + 0.015908258035779, + 0.015122473239898682, + 0.01707478240132332, + 0.02063312940299511, + 0.018097219988703728, + 0.010419505648314953, + 0.018109535798430443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.062047164887189865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up, the camera panning closely to the action, focuses on a warrior clad in weathered steel armor. In the dim light, the armor, covered in fine scratches and slight rust, reflects a cold metallic luster. His muscles are tense, and across his chiseled face, sweat and dust mingle, flowing down his furrowed brow and tense jaw. Wisps of dark hair fall from the edge of his helmet, plastering to his damp forehead. His eyes are wide, filled with determination and savage power, as he swings a broadsword, also scarred by battle, from right to left with all his might. In slow motion, the sword's edge slices through the air, creating a visible, distorted current, blurring the background into a blurry, dark-toned chaos, with faint traces of flying dust visible. The entire picture presents a faded retro film texture, with extremely low color saturation, showing a sepia tone washed by time, and is covered with obvious film grain, fine scratches and occasional flashes of light, creating a tense, primitive and epic atmosphere.", + "seed": 9374, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_11/index_5747_seed_9374_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": -0.11056242994964123, + "vlm_logprob": -0.11056242994964123, + "vlm_yes_no_margin": 2.699999976158142, + "vlm_token_credit": -0.11056242994964123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019283680245280266 + ], + "opposite_logprob_tokens": [ + -4.394283771514893 + ], + "logprob_reward": -0.019283680245280266, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.019283680245280266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08500988036394119, + 0.03518034517765045, + 0.17999008297920227, + 0.046861741691827774, + 0.15967364609241486, + 0.12715628743171692, + 0.07361196726560593, + 0.18639476597309113, + 0.04144105687737465, + 0.06468023359775543 + ], + "credit_map": [ + 0.009121570736169815, + 0.010855965316295624, + 0.008699231781065464, + 0.007129697594791651, + 0.02154153399169445, + 0.01068741176277399, + 0.009614602662622929, + 0.007359870243817568, + 0.003910285420715809, + 0.0044824182987213135, + 0.005400311201810837, + 0.005303898826241493, + 0.004628797993063927, + 0.003918529953807592, + 0.003190780058503151, + 0.004345323890447617, + 0.012955322861671448, + 0.021418042480945587, + 0.014623734168708324, + 0.025512374937534332, + 0.040123291313648224, + 0.026382537558674812, + 0.020233236253261566, + 0.01874154433608055, + 0.003969783429056406, + 0.004945063963532448, + 0.00508081866428256, + 0.0053092604503035545, + 0.006665416527539492, + 0.007817448116838932, + 0.00667315861210227, + 0.006400793790817261, + 0.022153066471219063, + 0.02795981802046299, + 0.012509535998106003, + 0.017201995477080345, + 0.027550630271434784, + 0.022046079859137535, + 0.015044840052723885, + 0.01520768366754055, + 0.01658688671886921, + 0.010662904009222984, + 0.0251852385699749, + 0.011432512663304806, + 0.0182422436773777, + 0.02751273848116398, + 0.008308321237564087, + 0.009225444868206978, + 0.008882628753781319, + 0.007341937627643347, + 0.00543627142906189, + 0.005445655435323715, + 0.0086797084659338, + 0.013585030101239681, + 0.010149623267352581, + 0.014091115444898605, + 0.012133783660829067, + 0.015477539971470833, + 0.05079659819602966, + 0.012553857639431953, + 0.05631372332572937, + 0.014258278533816338, + 0.014103779569268227, + 0.010757208801805973, + 0.0057778251357376575, + 0.007511962205171585, + 0.005199471954256296, + 0.0035111356992274523, + 0.004875611513853073, + 0.004183659795671701, + 0.003098046174272895, + 0.007283343002200127, + 0.008161636069417, + 0.009183320216834545, + 0.004498438443988562, + 0.007350489962846041, + 0.010085594840347767, + 0.007309026550501585, + 0.010520213283598423, + 0.007571512833237648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.019283680245280266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344786286354065 + ], + "opposite_logprob_tokens": [ + -1.3134478330612183 + ], + "logprob_reward": -0.31344786286354065, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344786286354065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08484280854463577, + 0.04543879255652428, + 0.24170681834220886, + 0.04112295061349869, + 0.09204644709825516, + 0.031674209982156754, + 0.06281358748674393, + 0.26406121253967285, + 0.0382571779191494, + 0.09803605079650879 + ], + "credit_map": [ + 0.009300129488110542, + 0.012699582614004612, + 0.00841422937810421, + 0.007819791324436665, + 0.019500983878970146, + 0.010686768218874931, + 0.011673255823552608, + 0.00474807433784008, + 0.003705583745613694, + 0.006189668085426092, + 0.005059434566646814, + 0.006112896837294102, + 0.006748138926923275, + 0.006151588633656502, + 0.005128714721649885, + 0.006342765875160694, + 0.014614216051995754, + 0.0324380062520504, + 0.017722900956869125, + 0.02434730902314186, + 0.05775299668312073, + 0.0362858809530735, + 0.035556644201278687, + 0.022988859564065933, + 0.003259334247559309, + 0.005339533556252718, + 0.005739113781601191, + 0.004589480347931385, + 0.005184631794691086, + 0.006143738515675068, + 0.0057435003109276295, + 0.005123617127537727, + 0.013046672567725182, + 0.01732753776013851, + 0.007858691737055779, + 0.006603705231100321, + 0.01818777807056904, + 0.011431428603827953, + 0.009655356407165527, + 0.007935279048979282, + 0.00403660349547863, + 0.003571414155885577, + 0.005619693081825972, + 0.0031079770997166634, + 0.005102154798805714, + 0.0045655760914087296, + 0.0027704539243131876, + 0.0029003366362303495, + 0.007997401989996433, + 0.01350262202322483, + 0.008767456747591496, + 0.0064900643192231655, + 0.0087567875161767, + 0.00703774718567729, + 0.004055861849337816, + 0.006205643527209759, + 0.017404580488801003, + 0.020084328949451447, + 0.07022961229085922, + 0.01337342243641615, + 0.08868616819381714, + 0.019139666110277176, + 0.019004609435796738, + 0.016138790175318718, + 0.0039226822555065155, + 0.004981614649295807, + 0.004681366961449385, + 0.004350109025835991, + 0.005089235957711935, + 0.0070615471340715885, + 0.004107221029698849, + 0.004063401836901903, + 0.008916657418012619, + 0.011393343098461628, + 0.007840867154300213, + 0.009727525524795055, + 0.014155898243188858, + 0.017144035547971725, + 0.015443224459886551, + 0.013414490036666393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.31344786286354065 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270337849855423 + ], + "opposite_logprob_tokens": [ + -2.1270337104797363 + ], + "logprob_reward": -0.1270337849855423, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270337849855423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10882235318422318, + 0.1194404736161232, + 0.11141259968280792, + 0.07214172929525375, + 0.07859326153993607, + 0.07248258590698242, + 0.028488794341683388, + 0.24816711246967316, + 0.07446401566267014, + 0.08598711341619492 + ], + "credit_map": [ + 0.012198740616440773, + 0.018042288720607758, + 0.013978101313114166, + 0.008460084907710552, + 0.02116716280579567, + 0.013425972312688828, + 0.01350562646985054, + 0.008044376038014889, + 0.011424372904002666, + 0.015080842189490795, + 0.01793397217988968, + 0.016807308420538902, + 0.013525390066206455, + 0.014685392379760742, + 0.013202151283621788, + 0.016781041398644447, + 0.009508662857115269, + 0.011457414366304874, + 0.008595195598900318, + 0.010195106267929077, + 0.025698700919747353, + 0.017612194642424583, + 0.011634614318609238, + 0.01671069860458374, + 0.0059064640663564205, + 0.008252143859863281, + 0.0091030802577734, + 0.007225142326205969, + 0.009528531692922115, + 0.011357536539435387, + 0.00918732862919569, + 0.011581498198211193, + 0.011804202571511269, + 0.015871861949563026, + 0.005120398011058569, + 0.005878963507711887, + 0.013801001943647861, + 0.009781971573829651, + 0.00852144043892622, + 0.007813416421413422, + 0.006818308029323816, + 0.006726179737597704, + 0.012657669372856617, + 0.005826056003570557, + 0.014339215122163296, + 0.01432393305003643, + 0.005580486264079809, + 0.006210730876773596, + 0.004341197200119495, + 0.003802061313763261, + 0.0026935029309242964, + 0.002177879447117448, + 0.0032134507782757282, + 0.003411676501855254, + 0.0036950246430933475, + 0.005154000595211983, + 0.01667974516749382, + 0.012078292667865753, + 0.07542210817337036, + 0.01298926305025816, + 0.08879945427179337, + 0.018054917454719543, + 0.015061020851135254, + 0.009082291275262833, + 0.009063987992703915, + 0.014736229553818703, + 0.00649214256554842, + 0.0074323262088000774, + 0.00818541832268238, + 0.009834494441747665, + 0.008004079572856426, + 0.010715327225625515, + 0.0068004559725522995, + 0.009282899089157581, + 0.009127652272582054, + 0.00954978633671999, + 0.012030611746013165, + 0.013392993248999119, + 0.010368339717388153, + 0.01543436385691166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.1270337849855423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810177743434906 + ], + "opposite_logprob_tokens": [ + -3.2881016731262207 + ], + "logprob_reward": -0.03810177743434906, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810177743434906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12913429737091064, + 0.12445520609617233, + 0.05389656499028206, + 0.0415426604449749, + 0.042230796068906784, + 0.17516963183879852, + 0.02426019310951233, + 0.32367193698883057, + 0.018765952438116074, + 0.06687266379594803 + ], + "credit_map": [ + 0.015378251671791077, + 0.02002202346920967, + 0.011926707811653614, + 0.012119793333113194, + 0.02797711081802845, + 0.017498303204774857, + 0.014722679741680622, + 0.009489435702562332, + 0.010846899822354317, + 0.014157461933791637, + 0.011022103950381279, + 0.02445627935230732, + 0.015084074810147285, + 0.015732314437627792, + 0.014179671183228493, + 0.018976405262947083, + 0.0038817792665213346, + 0.004530868958681822, + 0.004645136650651693, + 0.00400378555059433, + 0.01640867441892624, + 0.009371276944875717, + 0.004999879747629166, + 0.006055165082216263, + 0.003763376735150814, + 0.004282074514776468, + 0.004807707387953997, + 0.0053178840316832066, + 0.005365649238228798, + 0.008252818137407303, + 0.004692144226282835, + 0.005061008967459202, + 0.007078411988914013, + 0.005129947792738676, + 0.0045694997534155846, + 0.003846447914838791, + 0.006281172391027212, + 0.005339923780411482, + 0.0037894626148045063, + 0.006195932161062956, + 0.02463039755821228, + 0.018262149766087532, + 0.026091359555721283, + 0.01752941682934761, + 0.028270406648516655, + 0.025014357641339302, + 0.01692124269902706, + 0.018450317904353142, + 0.0037759796250611544, + 0.00265599531121552, + 0.0023990615736693144, + 0.0020924590062350035, + 0.002927940571680665, + 0.003160305554047227, + 0.0022717437241226435, + 0.004976710770279169, + 0.020614199340343475, + 0.03264976292848587, + 0.07582343369722366, + 0.01734280213713646, + 0.10386358201503754, + 0.024242538958787918, + 0.02875956892967224, + 0.02037607692182064, + 0.0026920298114418983, + 0.004367790184915066, + 0.001631745952181518, + 0.0019505753880366683, + 0.0026526404544711113, + 0.001939487410709262, + 0.001374626299366355, + 0.0021570581011474133, + 0.0069289137609303, + 0.00799018144607544, + 0.005582914687693119, + 0.00791953131556511, + 0.011622963473200798, + 0.010567002929747105, + 0.007538359146565199, + 0.008722803555428982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.03810177743434906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054945044219493866 + ], + "opposite_logprob_tokens": [ + -2.9299449920654297 + ], + "logprob_reward": -0.054945044219493866, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054945044219493866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07221473008394241, + 0.051017191261053085, + 0.11884185671806335, + 0.1572408825159073, + 0.1254260390996933, + 0.15273651480674744, + 0.056929849088191986, + 0.10538451373577118, + 0.07939809560775757, + 0.08081027120351791 + ], + "credit_map": [ + 0.007080337032675743, + 0.012204366736114025, + 0.008058257400989532, + 0.007268289104104042, + 0.014281745068728924, + 0.00822682399302721, + 0.009597012773156166, + 0.005497896112501621, + 0.005162919871509075, + 0.006204860284924507, + 0.0059247431345283985, + 0.006630136165767908, + 0.007146861869841814, + 0.0072625381872057915, + 0.005081901326775551, + 0.007603227160871029, + 0.007471299264580011, + 0.011863398365676403, + 0.009685144759714603, + 0.015019360929727554, + 0.027450306341052055, + 0.017469393089413643, + 0.013323419727385044, + 0.016559533774852753, + 0.015970531851053238, + 0.01999453827738762, + 0.019987979903817177, + 0.01645311340689659, + 0.018271327018737793, + 0.02234121784567833, + 0.019515912979841232, + 0.024706276133656502, + 0.018515009433031082, + 0.019421793520450592, + 0.009166283532977104, + 0.011451903730630875, + 0.019144907593727112, + 0.017277942970395088, + 0.015004649758338928, + 0.015443551354110241, + 0.01744815707206726, + 0.014017488807439804, + 0.02113126590847969, + 0.01299543771892786, + 0.01718062534928322, + 0.04049362242221832, + 0.013734489679336548, + 0.015735430642962456, + 0.008044176734983921, + 0.009998460300266743, + 0.0078116715885698795, + 0.004756948910653591, + 0.008061492815613747, + 0.008085662499070168, + 0.0036772231105715036, + 0.006494214292615652, + 0.007113764062523842, + 0.01205387432128191, + 0.014261621981859207, + 0.008905221708118916, + 0.029047435149550438, + 0.00979496631771326, + 0.013813069090247154, + 0.010394558310508728, + 0.00903709139674902, + 0.01691988855600357, + 0.007310936693102121, + 0.006608705967664719, + 0.010253841988742352, + 0.009951642714440823, + 0.007380757946521044, + 0.011935228481888771, + 0.0075607579201459885, + 0.008281946182250977, + 0.008093509823083878, + 0.006365460809320211, + 0.012333204969763756, + 0.013008374720811844, + 0.011249640956521034, + 0.013917375355958939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.054945044219493866 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up, the camera panning closely to the action, focuses on a warrior clad in weathered steel armor. In the dim light, the armor, covered in fine scratches and slight rust, reflects a cold metallic luster. His muscles are tense, and across his chiseled face, sweat and dust mingle, flowing down his furrowed brow and tense jaw. Wisps of dark hair fall from the edge of his helmet, plastering to his damp forehead. His eyes are wide, filled with determination and savage power, as he swings a broadsword, also scarred by battle, from right to left with all his might. In slow motion, the sword's edge slices through the air, creating a visible, distorted current, blurring the background into a blurry, dark-toned chaos, with faint traces of flying dust visible. The entire picture presents a faded retro film texture, with extremely low color saturation, showing a sepia tone washed by time, and is covered with obvious film grain, fine scratches and occasional flashes of light, creating a tense, primitive and epic atmosphere.", + "seed": 9375, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_11/index_5747_seed_9375_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": -0.13070952612906694, + "vlm_logprob": -0.13070952612906694, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.13070952612906694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014388498850166798 + ], + "opposite_logprob_tokens": [ + -5.014388561248779 + ], + "logprob_reward": -0.014388498850166798, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.014388498850166798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11396295577287674, + 0.045318324118852615, + 0.26835715770721436, + 0.02683459222316742, + 0.09095519036054611, + 0.05506514012813568, + 0.15973767638206482, + 0.07363752275705338, + 0.05883962661027908, + 0.10729178041219711 + ], + "credit_map": [ + 0.013975925743579865, + 0.017216309905052185, + 0.010767323896288872, + 0.012664807960391045, + 0.029821638017892838, + 0.011104945093393326, + 0.010531826876103878, + 0.007880172692239285, + 0.006230429280549288, + 0.006455209571868181, + 0.0035893190652132034, + 0.005505524575710297, + 0.006017422303557396, + 0.005817075725644827, + 0.006773723755031824, + 0.004929621238261461, + 0.010626829229295254, + 0.005833052564412355, + 0.010412875562906265, + 0.09913123399019241, + 0.016151705756783485, + 0.009651454165577888, + 0.10246876627206802, + 0.014081258326768875, + 0.002489475766196847, + 0.004553657490760088, + 0.00332978880032897, + 0.0030806721188127995, + 0.004004522692412138, + 0.003117986489087343, + 0.003585912985727191, + 0.0026725742500275373, + 0.005301678087562323, + 0.00979112833738327, + 0.008338858373463154, + 0.011111859232187271, + 0.018922358751296997, + 0.010976706631481647, + 0.016916656866669655, + 0.009595939889550209, + 0.006315943785011768, + 0.00525064067915082, + 0.006227500736713409, + 0.006162730511277914, + 0.0053888349793851376, + 0.008029458113014698, + 0.007749868556857109, + 0.009940166026353836, + 0.01688348688185215, + 0.025240641087293625, + 0.010566489771008492, + 0.016357874497771263, + 0.03779751434922218, + 0.02020793780684471, + 0.0206233449280262, + 0.012060385197401047, + 0.008315719664096832, + 0.013178439810872078, + 0.0065657831728458405, + 0.008381298743188381, + 0.01260350365191698, + 0.010486808605492115, + 0.009797393344342709, + 0.004308579955250025, + 0.006586415693163872, + 0.006009520031511784, + 0.0038786702789366245, + 0.005762971471995115, + 0.016737796366214752, + 0.00944339670240879, + 0.006229315418750048, + 0.004191537853330374, + 0.012446636334061623, + 0.013157406821846962, + 0.004981116857379675, + 0.007385904900729656, + 0.024743501096963882, + 0.014655690640211105, + 0.020043382421135902, + 0.009878149256110191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.014388498850166798 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134472668170929 + ], + "opposite_logprob_tokens": [ + -1.3134472370147705 + ], + "logprob_reward": -0.3134472668170929, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134472668170929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04234623536467552, + 0.08439870923757553, + 0.2491164356470108, + 0.03328121453523636, + 0.0543450266122818, + 0.06240138038992882, + 0.14603374898433685, + 0.03396828845143318, + 0.26098325848579407, + 0.033125799149274826 + ], + "credit_map": [ + 0.0061178081668913364, + 0.004574148450046778, + 0.00393815478309989, + 0.00586217874661088, + 0.009658725932240486, + 0.0035746342036873102, + 0.005368933547288179, + 0.003251653630286455, + 0.010231483727693558, + 0.011033378541469574, + 0.008691921830177307, + 0.008029078133404255, + 0.011985483579337597, + 0.013038886711001396, + 0.014106757938861847, + 0.007281715981662273, + 0.009610931389033794, + 0.005420411936938763, + 0.009238502010703087, + 0.0945528969168663, + 0.012096344493329525, + 0.007612906862050295, + 0.09535050392150879, + 0.015233932062983513, + 0.003407256444916129, + 0.00521839177235961, + 0.003920397721230984, + 0.0039260853081941605, + 0.004800072871148586, + 0.003837637137621641, + 0.004293554928153753, + 0.003877816954627633, + 0.004284228663891554, + 0.006861743982881308, + 0.005784062203019857, + 0.006560541223734617, + 0.010150179266929626, + 0.007126742508262396, + 0.007331789471209049, + 0.006245740223675966, + 0.009542200714349747, + 0.006347620859742165, + 0.006239051930606365, + 0.005808062851428986, + 0.006504652090370655, + 0.007111691869795322, + 0.008526397868990898, + 0.012321698479354382, + 0.015693368390202522, + 0.020291265100240707, + 0.008851238526403904, + 0.019533870741724968, + 0.0265957061201334, + 0.018045922741293907, + 0.024180088192224503, + 0.012842281721532345, + 0.003976744133979082, + 0.004912431351840496, + 0.002902975305914879, + 0.004916197154670954, + 0.005409014876931906, + 0.004874022211879492, + 0.005062731448560953, + 0.0019141711527481675, + 0.02053719013929367, + 0.019747627899050713, + 0.017071524634957314, + 0.030168773606419563, + 0.0778626874089241, + 0.05013098567724228, + 0.02659880928695202, + 0.01886563189327717, + 0.00347185879945755, + 0.0028614336624741554, + 0.0022848923690617085, + 0.0026530111208558083, + 0.006126537453383207, + 0.005911556538194418, + 0.006339217070490122, + 0.003477293998003006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.3134472668170929 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556085884571075 + ], + "opposite_logprob_tokens": [ + -1.6005609035491943 + ], + "logprob_reward": -0.22556085884571075, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556085884571075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1642569601535797, + 0.058803409337997437, + 0.2066469043493271, + 0.06130998954176903, + 0.15509015321731567, + 0.08818937838077545, + 0.11402764916419983, + 0.042091093957424164, + 0.0828041359782219, + 0.026780303567647934 + ], + "credit_map": [ + 0.015018516220152378, + 0.023713409900665283, + 0.01804300583899021, + 0.018801460042595863, + 0.04031124338507652, + 0.015530124306678772, + 0.020462997257709503, + 0.01237618736922741, + 0.007010302972048521, + 0.007343723438680172, + 0.006867813412100077, + 0.006878996267914772, + 0.007280334830284119, + 0.006454366259276867, + 0.009195288643240929, + 0.007772584445774555, + 0.008244816213846207, + 0.005221836268901825, + 0.008565355092287064, + 0.07720769941806793, + 0.01130032166838646, + 0.006969501730054617, + 0.07763423025608063, + 0.011503140442073345, + 0.0048364936374127865, + 0.009215048514306545, + 0.0059308847412467, + 0.0074502271600067616, + 0.00996081717312336, + 0.007216230034828186, + 0.00972835160791874, + 0.006971937138587236, + 0.011570539325475693, + 0.02230221778154373, + 0.01744408719241619, + 0.018208399415016174, + 0.02595747634768486, + 0.017532967031002045, + 0.025331029668450356, + 0.01674344390630722, + 0.010951529256999493, + 0.011446970514953136, + 0.007529923692345619, + 0.008282904513180256, + 0.012653693556785583, + 0.012026185169816017, + 0.010264941491186619, + 0.015033232048153877, + 0.012443674728274345, + 0.016600610688328743, + 0.0065319109708070755, + 0.010531884618103504, + 0.025077776983380318, + 0.012913969345390797, + 0.01790052093565464, + 0.012027304619550705, + 0.004884028807282448, + 0.0059561473317444324, + 0.0033120110165327787, + 0.005440735723823309, + 0.006236211396753788, + 0.008320356719195843, + 0.005442594178020954, + 0.0024990125093609095, + 0.007868154905736446, + 0.00879735965281725, + 0.006030444987118244, + 0.0063553499057888985, + 0.02295314334332943, + 0.01327300164848566, + 0.011019917204976082, + 0.006506766192615032, + 0.002997671253979206, + 0.0030382878612726927, + 0.002853092970326543, + 0.0030649788677692413, + 0.004290881101042032, + 0.0038918787613511086, + 0.0037987076211720705, + 0.0028448060620576143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.22556085884571075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108546286821365 + ], + "opposite_logprob_tokens": [ + -3.2881085872650146 + ], + "logprob_reward": -0.038108546286821365, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108546286821365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.133433997631073, + 0.21528850495815277, + 0.14957177639007568, + 0.0378316305577755, + 0.11907829344272614, + 0.03436620533466339, + 0.11338645219802856, + 0.07734496891498566, + 0.04397674277424812, + 0.07572140544652939 + ], + "credit_map": [ + 0.016132397577166557, + 0.01885819435119629, + 0.011941914446651936, + 0.02052357606589794, + 0.02923470176756382, + 0.01364027801901102, + 0.013453077524900436, + 0.009649863466620445, + 0.026968596503138542, + 0.03171674907207489, + 0.018084397539496422, + 0.019967732951045036, + 0.029723793268203735, + 0.019696932286024094, + 0.04140883684158325, + 0.027721481397747993, + 0.006795014254748821, + 0.0032878685742616653, + 0.00557221844792366, + 0.04688754677772522, + 0.00847200769931078, + 0.00524621969088912, + 0.06421295553445816, + 0.009097951464354992, + 0.0033919101115316153, + 0.004230871330946684, + 0.003725901246070862, + 0.0050160689279437065, + 0.0070365662686526775, + 0.004561270121484995, + 0.005622528493404388, + 0.004246515221893787, + 0.01060287281870842, + 0.0149046890437603, + 0.015866316854953766, + 0.012284313328564167, + 0.023778660222887993, + 0.014024470001459122, + 0.015153724700212479, + 0.012463254854083061, + 0.004531083162873983, + 0.004636950325220823, + 0.0025850089732557535, + 0.002683335682377219, + 0.0048595331609249115, + 0.003938502166420221, + 0.004044115077704191, + 0.007087677251547575, + 0.013800677843391895, + 0.016207817941904068, + 0.006671728566288948, + 0.012011456303298473, + 0.02243967354297638, + 0.01383352093398571, + 0.01792866736650467, + 0.010492906905710697, + 0.008210545405745506, + 0.010686123743653297, + 0.00971563346683979, + 0.009975205175578594, + 0.011569821275770664, + 0.008544783107936382, + 0.0124964052811265, + 0.006146453320980072, + 0.0033444827422499657, + 0.003160403575748205, + 0.0032551574986428022, + 0.006700250320136547, + 0.00974071305245161, + 0.006661368999630213, + 0.007682621944695711, + 0.0034317420795559883, + 0.010784302838146687, + 0.005852132569998503, + 0.0036816697102040052, + 0.0043273684568703175, + 0.011275039985775948, + 0.010439109057188034, + 0.008994889445602894, + 0.020366894081234932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.038108546286821365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a warrior wearing weathered steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior swinging a broadsword from right to left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a faded retro film texture with low color saturation and a sepia tone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning closely to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204245984554291 + ], + "opposite_logprob_tokens": [ + -2.812042474746704 + ], + "logprob_reward": -0.06204245984554291, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204245984554291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15685337781906128, + 0.06998825073242188, + 0.20371492207050323, + 0.028783190995454788, + 0.13567352294921875, + 0.07313648611307144, + 0.08024495840072632, + 0.12809932231903076, + 0.04988919943571091, + 0.07361681014299393 + ], + "credit_map": [ + 0.018107466399669647, + 0.02442556619644165, + 0.016595494002103806, + 0.01826036348938942, + 0.036966823041439056, + 0.015148205682635307, + 0.014936738647520542, + 0.012412706390023232, + 0.00920442957431078, + 0.009643231518566608, + 0.00787736289203167, + 0.009363553486764431, + 0.006205724086612463, + 0.005924475844949484, + 0.011111599393188953, + 0.010657870210707188, + 0.008956271223723888, + 0.004849095828831196, + 0.007923333905637264, + 0.06787208467721939, + 0.012342448346316814, + 0.00909244641661644, + 0.08167921751737595, + 0.011000017635524273, + 0.0023118089884519577, + 0.0036800834350287914, + 0.003054539207369089, + 0.0036734279710799456, + 0.003953763283789158, + 0.003130754455924034, + 0.005366363096982241, + 0.0036124498583376408, + 0.010470603592693806, + 0.020053569227457047, + 0.016969921067357063, + 0.012306900694966316, + 0.028555026277899742, + 0.018854591995477676, + 0.01706608012318611, + 0.011396828107535839, + 0.008888177573680878, + 0.01061287522315979, + 0.011501402594149113, + 0.0074734813533723354, + 0.008225028403103352, + 0.007846418768167496, + 0.00840495154261589, + 0.01018415205180645, + 0.008633125573396683, + 0.011823365464806557, + 0.00483235577121377, + 0.010962463915348053, + 0.015360568650066853, + 0.009773924015462399, + 0.01097685843706131, + 0.007882291451096535, + 0.014517316594719887, + 0.018055766820907593, + 0.012460152618587017, + 0.019558852538466454, + 0.019747942686080933, + 0.01942477934062481, + 0.016532009467482567, + 0.0078024971298873425, + 0.006171181797981262, + 0.0061181215569376945, + 0.006140690762549639, + 0.006542314775288105, + 0.007490450516343117, + 0.006124763749539852, + 0.006508428603410721, + 0.004793250933289528, + 0.008069287985563278, + 0.007830542512238026, + 0.0042229825630784035, + 0.0049217925406992435, + 0.013651889748871326, + 0.010386372916400433, + 0.008478794246912003, + 0.0160551518201828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.06204245984554291 + } + ], + "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/0000025/group_12/index_tensor([4417, 4417])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_12/index_tensor([4417, 4417])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..914716355bc495a5dbaedaf3fe6c8176dd9c4bd1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_12/index_tensor([4417, 4417])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot captures a massive American flag fluttering violently against a vast, flawless azure sky. Rendered at a high frame rate, the footage captures every movement with surreal, crystal clarity. Under harsh direct sunlight, the flag's heavy fabric snaps and ripples in the powerful, unseen wind. Crimson and bright white stripes undulate in powerful, chaotic waves, while the navy blue canton, densely embroidered with 50 stars, twists and surges with immense force. The intense light meticulously defines the texture of the woven material, creating shimmering highlights at the tops of the creases and casting deep, sharp shadows within the valleys. The flag clings to the halyard atop the tall flagpole, its edge whipping violently, every ripple and twist captured without a hint of motion blur, creating a powerful and awe-inspiring atmosphere of raw, untamed energy and enduring power.", + "seed": 15322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_12/index_4417_seed_15322_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": -0.44440235793590543, + "vlm_logprob": -0.44440235793590543, + "vlm_yes_no_margin": 1.475000023841858, + "vlm_token_credit": -0.44440235793590543, + "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 shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8634039163589478 + ], + "opposite_logprob_tokens": [ + -0.6134039163589478 + ], + "logprob_reward": -0.8634039163589478, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8634039163589478, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1021205335855484, + 0.04776868224143982, + 0.273693323135376, + 0.06611819565296173, + 0.04359550029039383, + 0.05717833340167999, + 0.18864388763904572, + 0.06420641392469406, + 0.1273321509361267, + 0.029342951253056526 + ], + "credit_map": [ + 0.021529177203774452, + 0.013622959144413471, + 0.0073042940348386765, + 0.0054511865600943565, + 0.009745658375322819, + 0.017664702609181404, + 0.007043780758976936, + 0.01975877955555916, + 0.00938326958566904, + 0.006907855160534382, + 0.003209085203707218, + 0.0027815867215394974, + 0.006923560053110123, + 0.008638597093522549, + 0.0036212638951838017, + 0.006303470116108656, + 0.05416984111070633, + 0.05214373767375946, + 0.014175248332321644, + 0.011781328357756138, + 0.04045329988002777, + 0.04761506989598274, + 0.02432332932949066, + 0.029031502082943916, + 0.009378880262374878, + 0.00788868498057127, + 0.003272740636020899, + 0.004549214616417885, + 0.013250377029180527, + 0.010277265682816505, + 0.00480266846716404, + 0.012698368169367313, + 0.007313525304198265, + 0.005135145504027605, + 0.0024663996882736683, + 0.0029403753578662872, + 0.005122760310769081, + 0.009673752821981907, + 0.0036944139283150434, + 0.007249132730066776, + 0.009765587747097015, + 0.005533761810511351, + 0.0038581332191824913, + 0.004085642751306295, + 0.005450225435197353, + 0.011217869818210602, + 0.008778191171586514, + 0.008488927967846394, + 0.011041132733225822, + 0.009158080443739891, + 0.00448025856167078, + 0.011867362074553967, + 0.007354397792369127, + 0.007576946634799242, + 0.07877150923013687, + 0.05839420482516289, + 0.006154024973511696, + 0.004005715250968933, + 0.002693148795515299, + 0.004530610051006079, + 0.0035084655974060297, + 0.017329011112451553, + 0.015098268166184425, + 0.010887173004448414, + 0.02028699405491352, + 0.011447121389210224, + 0.006203114986419678, + 0.013705401681363583, + 0.010178483091294765, + 0.018969709053635597, + 0.023069605231285095, + 0.023471735417842865, + 0.003363749012351036, + 0.002582266926765442, + 0.0017688844818621874, + 0.0023277958389371634, + 0.002436133334413171, + 0.0034723884891718626, + 0.005561638157814741, + 0.007830098271369934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.8634039163589478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04872018098831177 + ], + "opposite_logprob_tokens": [ + -3.048720121383667 + ], + "logprob_reward": -0.04872018098831177, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04872018098831177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08564572781324387, + 0.15047577023506165, + 0.1473381370306015, + 0.10200277715921402, + 0.1274290680885315, + 0.04849725589156151, + 0.04869598522782326, + 0.10777994990348816, + 0.10806383192539215, + 0.07407144457101822 + ], + "credit_map": [ + 0.016597753390669823, + 0.009200787171721458, + 0.00523640401661396, + 0.005047147162258625, + 0.008947809226810932, + 0.012010408565402031, + 0.008061391301453114, + 0.020544027909636497, + 0.019931292161345482, + 0.01908811926841736, + 0.006006136070936918, + 0.00972310733050108, + 0.026882367208600044, + 0.028784653171896935, + 0.012394312769174576, + 0.02766578644514084, + 0.0266577135771513, + 0.02779892459511757, + 0.008596247993409634, + 0.006226275581866503, + 0.024791568517684937, + 0.025423986837267876, + 0.011156987398862839, + 0.016686435788869858, + 0.02375400811433792, + 0.011185158975422382, + 0.005762065295130014, + 0.0055074007250368595, + 0.016068361699581146, + 0.016735291108489037, + 0.005445963703095913, + 0.017544517293572426, + 0.01709115318953991, + 0.015687551349401474, + 0.0074325245805084705, + 0.010666720569133759, + 0.016130827367305756, + 0.027697313576936722, + 0.014701391570270061, + 0.018021581694483757, + 0.007470337674021721, + 0.006332776043564081, + 0.0034067239612340927, + 0.003998998086899519, + 0.004969350062310696, + 0.009653283283114433, + 0.006760922260582447, + 0.005904861725866795, + 0.0035358613822609186, + 0.0026974924840033054, + 0.0015007625333964825, + 0.0026309730019420385, + 0.0020812188740819693, + 0.003288512583822012, + 0.01777275837957859, + 0.015188409015536308, + 0.008691081777215004, + 0.0068391673266887665, + 0.00618822081014514, + 0.009251457639038563, + 0.006312768906354904, + 0.02635318785905838, + 0.0262665543705225, + 0.01787751540541649, + 0.022615421563386917, + 0.013358457013964653, + 0.008335454389452934, + 0.006594576872885227, + 0.010506385937333107, + 0.014141902327537537, + 0.011305785737931728, + 0.0212058387696743, + 0.007667218334972858, + 0.00472941854968667, + 0.003133359830826521, + 0.0039059759583324194, + 0.004482804797589779, + 0.011194247752428055, + 0.011656828224658966, + 0.027301592752337456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.04872018098831177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029830604791641235 + ], + "opposite_logprob_tokens": [ + -3.5298306941986084 + ], + "logprob_reward": -0.029830604791641235, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029830604791641235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17747366428375244, + 0.13058900833129883, + 0.06561417877674103, + 0.0741884633898735, + 0.20873360335826874, + 0.07489350438117981, + 0.058103132992982864, + 0.0355846993625164, + 0.09997142851352692, + 0.07484833896160126 + ], + "credit_map": [ + 0.02208782732486725, + 0.01911279559135437, + 0.008663590997457504, + 0.014789153821766376, + 0.019332394003868103, + 0.024926725775003433, + 0.017121383920311928, + 0.051439784467220306, + 0.016370300203561783, + 0.01898309960961342, + 0.006211595609784126, + 0.00950214359909296, + 0.02212168090045452, + 0.025755910202860832, + 0.011557245627045631, + 0.02008702978491783, + 0.009433425031602383, + 0.007756506092846394, + 0.0031524403020739555, + 0.004276568070054054, + 0.010337292216718197, + 0.010552107356488705, + 0.007712157908827066, + 0.012393679469823837, + 0.008649276569485664, + 0.008615318685770035, + 0.0038651772774755955, + 0.005594773683696985, + 0.01155886147171259, + 0.010416877456009388, + 0.008922021836042404, + 0.01656615547835827, + 0.02453145757317543, + 0.025638632476329803, + 0.01112622581422329, + 0.020085958763957024, + 0.02370697259902954, + 0.04464348778128624, + 0.032134395092725754, + 0.026866475120186806, + 0.011739085428416729, + 0.011405836790800095, + 0.0051236264407634735, + 0.005125514697283506, + 0.009463313035666943, + 0.013309531845152378, + 0.008908244781196117, + 0.00981835089623928, + 0.006503542419523001, + 0.0068369521759450436, + 0.003629225306212902, + 0.005133906379342079, + 0.00672967778518796, + 0.008392333053052425, + 0.012828404083848, + 0.008049092255532742, + 0.00466932961717248, + 0.0029704610351473093, + 0.002068143803626299, + 0.003208136884495616, + 0.0024394928477704525, + 0.007657858543097973, + 0.006175183691084385, + 0.006396092474460602, + 0.014671387150883675, + 0.009017598815262318, + 0.007086120545864105, + 0.007913975045084953, + 0.011102336458861828, + 0.013793205842375755, + 0.01495322398841381, + 0.02143358439207077, + 0.00928072351962328, + 0.00639998447149992, + 0.0035311938263475895, + 0.007196349557489157, + 0.006763881538063288, + 0.009428226388990879, + 0.018391121178865433, + 0.013856858946383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.029830604791641235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.1371681690216064 + ], + "opposite_logprob_tokens": [ + -0.38716819882392883 + ], + "logprob_reward": -1.1371681690216064, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371681690216064, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06596365571022034, + 0.06336425244808197, + 0.12062432616949081, + 0.1448151171207428, + 0.058091774582862854, + 0.040396250784397125, + 0.20155750215053558, + 0.053742121905088425, + 0.1220988854765892, + 0.12934617698192596 + ], + "credit_map": [ + 0.01042504608631134, + 0.009314948692917824, + 0.004439921118319035, + 0.0035363193601369858, + 0.008117329329252243, + 0.010961916297674179, + 0.006095117889344692, + 0.013073052279651165, + 0.006244202610105276, + 0.007944047451019287, + 0.0026733705308288336, + 0.004694066010415554, + 0.011159265413880348, + 0.012915697880089283, + 0.006485865917056799, + 0.011247726157307625, + 0.02287496253848076, + 0.02301831729710102, + 0.006041520740836859, + 0.005230123642832041, + 0.0184532031416893, + 0.021434485912322998, + 0.010126236826181412, + 0.013445465825498104, + 0.0191999152302742, + 0.014840968884527683, + 0.008269214071333408, + 0.010519886389374733, + 0.026236142963171005, + 0.019822537899017334, + 0.011598660610616207, + 0.034327782690525055, + 0.007607294712215662, + 0.007095012813806534, + 0.002565529430285096, + 0.0064580440521240234, + 0.00563187338411808, + 0.00941692665219307, + 0.010723073966801167, + 0.00859401561319828, + 0.006604759022593498, + 0.0053010545670986176, + 0.0024099547881633043, + 0.0031952050048857927, + 0.005101701710373163, + 0.006811888422816992, + 0.005257728509604931, + 0.005713953636586666, + 0.011645345948636532, + 0.010829074308276176, + 0.004735963884741068, + 0.013055182993412018, + 0.00838354043662548, + 0.007972115650773048, + 0.08310208469629288, + 0.06183416023850441, + 0.006863373331725597, + 0.003971562255173922, + 0.0034574775490909815, + 0.0056640212424099445, + 0.004362262319773436, + 0.009229648858308792, + 0.012500216253101826, + 0.007693554740399122, + 0.02103506028652191, + 0.013126315549015999, + 0.006275473162531853, + 0.015901047736406326, + 0.011773591861128807, + 0.010300868190824986, + 0.026990555226802826, + 0.01669597066938877, + 0.00974327139556408, + 0.007579665631055832, + 0.005683159921318293, + 0.008769943378865719, + 0.012026329524815083, + 0.021789491176605225, + 0.015378102660179138, + 0.04837620630860329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -1.1371681690216064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14288891851902008 + ], + "opposite_logprob_tokens": [ + -2.0178890228271484 + ], + "logprob_reward": -0.14288891851902008, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14288891851902008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07636847347021103, + 0.09300592541694641, + 0.1568281352519989, + 0.09273780882358551, + 0.04495803639292717, + 0.07003351300954819, + 0.07044469565153122, + 0.04219517484307289, + 0.14852672815322876, + 0.2049015611410141 + ], + "credit_map": [ + 0.011055514216423035, + 0.008496114052832127, + 0.0038531164173036814, + 0.0041980319656431675, + 0.007471705786883831, + 0.014112401753664017, + 0.006702887825667858, + 0.02047870121896267, + 0.018985845148563385, + 0.012206070125102997, + 0.005582825280725956, + 0.00476992828771472, + 0.012851210311055183, + 0.01869506761431694, + 0.007265584077686071, + 0.012649388052523136, + 0.028697917237877846, + 0.022752651944756508, + 0.008268112316727638, + 0.006917697377502918, + 0.028119821101427078, + 0.023675749078392982, + 0.013212413527071476, + 0.025183767080307007, + 0.010743930004537106, + 0.009770285338163376, + 0.005660568363964558, + 0.00641874223947525, + 0.020160574465990067, + 0.014234387315809727, + 0.007697596214711666, + 0.01805173233151436, + 0.00718131335452199, + 0.0057535432279109955, + 0.003204930108040571, + 0.0025292071513831615, + 0.0063340142369270325, + 0.009602750651538372, + 0.003620253410190344, + 0.006732023786753416, + 0.011566611006855965, + 0.007745089940726757, + 0.0051652416586875916, + 0.006870297249406576, + 0.006050539668649435, + 0.011999569833278656, + 0.01293159369379282, + 0.007704570423811674, + 0.005206869449466467, + 0.004594593774527311, + 0.0019140254007652402, + 0.004356320947408676, + 0.0036865067668259144, + 0.002943456405773759, + 0.027750656008720398, + 0.01999226212501526, + 0.005978258792310953, + 0.003998578060418367, + 0.0029155004303902388, + 0.0037425137124955654, + 0.0028593989554792643, + 0.0076940967701375484, + 0.009727630764245987, + 0.005279200151562691, + 0.024817675352096558, + 0.017321938648819923, + 0.008984179235994816, + 0.012707095593214035, + 0.019375983625650406, + 0.027069322764873505, + 0.018551843240857124, + 0.01969868689775467, + 0.014514590613543987, + 0.011739331297576427, + 0.008238046430051327, + 0.01783006824553013, + 0.01834798976778984, + 0.03344094753265381, + 0.036369360983371735, + 0.06442122906446457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.14288891851902008 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a massive American flag fluttering violently against a vast, flawless azure sky. Rendered at a high frame rate, the footage captures every movement with surreal, crystal clarity. Under harsh direct sunlight, the flag's heavy fabric snaps and ripples in the powerful, unseen wind. Crimson and bright white stripes undulate in powerful, chaotic waves, while the navy blue canton, densely embroidered with 50 stars, twists and surges with immense force. The intense light meticulously defines the texture of the woven material, creating shimmering highlights at the tops of the creases and casting deep, sharp shadows within the valleys. The flag clings to the halyard atop the tall flagpole, its edge whipping violently, every ripple and twist captured without a hint of motion blur, creating a powerful and awe-inspiring atmosphere of raw, untamed energy and enduring power.", + "seed": 15323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_12/index_4417_seed_15323_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": -0.30397042017430065, + "vlm_logprob": -0.30397042017430065, + "vlm_yes_no_margin": 2.2750000238418577, + "vlm_token_credit": -0.30397042017430065, + "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 taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10693324357271194 + ], + "opposite_logprob_tokens": [ + -2.731933355331421 + ], + "logprob_reward": -0.10693324357271194, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.10693324357271194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23793666064739227, + 0.08682220429182053, + 0.14916378259658813, + 0.08101923018693924, + 0.027481859549880028, + 0.04085727035999298, + 0.2725875675678253, + 0.03972586244344711, + 0.043782833963632584, + 0.020622719079256058 + ], + "credit_map": [ + 0.04192456603050232, + 0.0319255031645298, + 0.0165924783796072, + 0.014770379289984703, + 0.01997820846736431, + 0.041460055857896805, + 0.013177436776459217, + 0.058108046650886536, + 0.016431324183940887, + 0.016586508601903915, + 0.0073626781813800335, + 0.006343282293528318, + 0.005481070373207331, + 0.014566763304173946, + 0.0060199229046702385, + 0.014030655845999718, + 0.04763823002576828, + 0.014726041816174984, + 0.014223289676010609, + 0.006568079814314842, + 0.009179665707051754, + 0.025715725496411324, + 0.006345119792968035, + 0.024767620489001274, + 0.024981478229165077, + 0.00924538355320692, + 0.008936884813010693, + 0.003262508427724242, + 0.005165679380297661, + 0.014567509293556213, + 0.0030510115902870893, + 0.01180877722799778, + 0.008477071300148964, + 0.0032395683228969574, + 0.0029345371294766665, + 0.001563321566209197, + 0.0019932978320866823, + 0.004918364342302084, + 0.0016976363258436322, + 0.0026580635458230972, + 0.011396828107535839, + 0.003794218646362424, + 0.0030549014918506145, + 0.0031964366789907217, + 0.0035696374252438545, + 0.006791764870285988, + 0.004052025731652975, + 0.005001457873731852, + 0.051072243601083755, + 0.027759557589888573, + 0.02023281157016754, + 0.0228677149862051, + 0.02317132242023945, + 0.044950954616069794, + 0.02582061104476452, + 0.05671233683824539, + 0.006571416277438402, + 0.006617972161620855, + 0.002649336587637663, + 0.0032946299761533737, + 0.004196562338620424, + 0.006484661716967821, + 0.0033566716592758894, + 0.0065546100959181786, + 0.008521834388375282, + 0.004117364063858986, + 0.0029379723127931356, + 0.0036489726044237614, + 0.0033206443767994642, + 0.007551565766334534, + 0.0041242437437176704, + 0.00956023670732975, + 0.00492988433688879, + 0.002020074985921383, + 0.0016934230225160718, + 0.0012631189310923219, + 0.0018261716468259692, + 0.0031331528443843126, + 0.0027274740859866142, + 0.003029419342055917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.10693324357271194 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012580911628901958 + ], + "opposite_logprob_tokens": [ + -4.387580871582031 + ], + "logprob_reward": -0.012580911628901958, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012580911628901958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09108948707580566, + 0.25701919198036194, + 0.06485297530889511, + 0.0749073177576065, + 0.04635122790932655, + 0.03758912533521652, + 0.3057883679866791, + 0.06372275203466415, + 0.026859983801841736, + 0.03181959316134453 + ], + "credit_map": [ + 0.014170252718031406, + 0.012092492543160915, + 0.005520880222320557, + 0.004514235537499189, + 0.0042062243446707726, + 0.01700729876756668, + 0.005353015847504139, + 0.028225086629390717, + 0.0572243332862854, + 0.023143522441387177, + 0.015768075361847878, + 0.009946812875568867, + 0.01374797709286213, + 0.04814562201499939, + 0.015637362375855446, + 0.07340548187494278, + 0.01659744791686535, + 0.008763007819652557, + 0.005398864857852459, + 0.004395464435219765, + 0.0048644267953932285, + 0.01039139088243246, + 0.0061434791423380375, + 0.008298889733850956, + 0.02118389494717121, + 0.007522474974393845, + 0.007206047885119915, + 0.00248859659768641, + 0.0047943368554115295, + 0.014004066586494446, + 0.0028330583591014147, + 0.014874843880534172, + 0.016123462468385696, + 0.004719733726233244, + 0.004331925418227911, + 0.0017658910946920514, + 0.0024177946615964174, + 0.009030405431985855, + 0.0021389215253293514, + 0.005823094397783279, + 0.00690930150449276, + 0.0033954638056457043, + 0.002809214172884822, + 0.0031064560171216726, + 0.0036923233419656754, + 0.006341049913316965, + 0.003888749284669757, + 0.007446570787578821, + 0.08379988372325897, + 0.02336282655596733, + 0.019589757546782494, + 0.009242748841643333, + 0.02426963672041893, + 0.06211662292480469, + 0.01950155757367611, + 0.06390534341335297, + 0.017377784475684166, + 0.007061701267957687, + 0.0045174118131399155, + 0.003911422099918127, + 0.0040052211843431, + 0.011100780218839645, + 0.005164974834769964, + 0.010583453811705112, + 0.005809292197227478, + 0.0022469009272754192, + 0.0017136618262156844, + 0.001737086335197091, + 0.0017255342099815607, + 0.005249474663287401, + 0.002060693223029375, + 0.00631733937188983, + 0.006580808199942112, + 0.001979514490813017, + 0.0019330885261297226, + 0.0014157629339024425, + 0.0012803277932107449, + 0.005850658752024174, + 0.0035742938052862883, + 0.009205138310790062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.012580911628901958 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033720459789037704 + ], + "opposite_logprob_tokens": [ + -3.4087204933166504 + ], + "logprob_reward": -0.033720459789037704, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033720459789037704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18876111507415771, + 0.19827446341514587, + 0.13645678758621216, + 0.14813154935836792, + 0.038096826523542404, + 0.03379227966070175, + 0.05120482295751572, + 0.06483425199985504, + 0.11246946454048157, + 0.027978431433439255 + ], + "credit_map": [ + 0.045702606439590454, + 0.019953355193138123, + 0.011324908584356308, + 0.013223959133028984, + 0.01603420078754425, + 0.031978804618120193, + 0.015332553535699844, + 0.03521072492003441, + 0.04430415481328964, + 0.02907366119325161, + 0.011730373837053776, + 0.011669655330479145, + 0.012997983954846859, + 0.03468471020460129, + 0.01145845465362072, + 0.042355459183454514, + 0.03123052418231964, + 0.021110359579324722, + 0.008674771524965763, + 0.006680777296423912, + 0.009730656631290913, + 0.02379179559648037, + 0.00683456938713789, + 0.028403323143720627, + 0.03271230682730675, + 0.020816154778003693, + 0.011865412816405296, + 0.007689393125474453, + 0.011998002417385578, + 0.02529279515147209, + 0.006990497466176748, + 0.030766991898417473, + 0.013398214243352413, + 0.0037092582788318396, + 0.003719848580658436, + 0.0016738520935177803, + 0.002503146417438984, + 0.007323689758777618, + 0.0018577836453914642, + 0.003911034669727087, + 0.0054521579295396805, + 0.002705906517803669, + 0.0018408073810860515, + 0.0033934861421585083, + 0.0030601031612604856, + 0.0058928136713802814, + 0.004155492875725031, + 0.007291513495147228, + 0.010970317758619785, + 0.005228197667747736, + 0.00404641218483448, + 0.0037994347512722015, + 0.0035299547016620636, + 0.00854135025292635, + 0.005619834177196026, + 0.009469321928918362, + 0.015904860571026802, + 0.00443632947281003, + 0.004623765591531992, + 0.002413366921246052, + 0.002916601486504078, + 0.01297104824334383, + 0.0036049385089427233, + 0.017963340505957603, + 0.025794856250286102, + 0.009000304155051708, + 0.0077798510901629925, + 0.007419445086270571, + 0.007373038679361343, + 0.021532200276851654, + 0.008832268416881561, + 0.024737505242228508, + 0.004819320980459452, + 0.0018384280847385526, + 0.001823827507905662, + 0.0018335104687139392, + 0.00243980810046196, + 0.004380000755190849, + 0.004187532234936953, + 0.006656006444245577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.033720459789037704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.22374427318573 + ], + "opposite_logprob_tokens": [ + -0.34874430298805237 + ], + "logprob_reward": -1.22374427318573, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.22374427318573, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09723580628633499, + 0.1409136950969696, + 0.06307587027549744, + 0.07185184955596924, + 0.08575456589460373, + 0.22732892632484436, + 0.16709649562835693, + 0.028326332569122314, + 0.08160999417304993, + 0.03680647909641266 + ], + "credit_map": [ + 0.021228404715657234, + 0.014115686528384686, + 0.008437348529696465, + 0.006622637156397104, + 0.006535561755299568, + 0.016464421525597572, + 0.005732817575335503, + 0.01809893548488617, + 0.036989565938711166, + 0.020591167733073235, + 0.009051318280398846, + 0.007546469569206238, + 0.009430880658328533, + 0.0243220217525959, + 0.008537321351468563, + 0.02444494515657425, + 0.01547915581613779, + 0.0092189721763134, + 0.004897169768810272, + 0.0036792526952922344, + 0.005698004737496376, + 0.009950333274900913, + 0.004124389495700598, + 0.010028592310845852, + 0.022439608350396156, + 0.008100283332169056, + 0.006919109728187323, + 0.003256511176005006, + 0.0046222941018640995, + 0.013276910409331322, + 0.0029026910196989775, + 0.010334443300962448, + 0.023656116798520088, + 0.011033878661692142, + 0.00593153340741992, + 0.004131438676267862, + 0.005100288428366184, + 0.015265213325619698, + 0.006768191698938608, + 0.013867909088730812, + 0.05027996376156807, + 0.02295190468430519, + 0.02028968185186386, + 0.014444831758737564, + 0.015839101746678352, + 0.039281316101551056, + 0.015521654859185219, + 0.04872047156095505, + 0.03706072270870209, + 0.01492922380566597, + 0.00980412121862173, + 0.007652385625988245, + 0.0186625923961401, + 0.030287472531199455, + 0.010246663354337215, + 0.03845329210162163, + 0.006003941409289837, + 0.0031214228365570307, + 0.0021137327421456575, + 0.001505457446910441, + 0.001934060244821012, + 0.0051805404946208, + 0.0021758079528808594, + 0.006291370838880539, + 0.022918706759810448, + 0.00504037132486701, + 0.005554649047553539, + 0.00436884630471468, + 0.0045614042319357395, + 0.016389893367886543, + 0.005796208046376705, + 0.016979917883872986, + 0.008864008821547031, + 0.0036294106394052505, + 0.002217658329755068, + 0.003160493914037943, + 0.002177000278607011, + 0.00534412544220686, + 0.005633959546685219, + 0.005779825150966644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -1.22374427318573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14287321269512177 + ], + "opposite_logprob_tokens": [ + -2.0178732872009277 + ], + "logprob_reward": -0.14287321269512177, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14287321269512177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12659384310245514, + 0.15691252052783966, + 0.06654014438390732, + 0.09998560696840286, + 0.046669892966747284, + 0.16176922619342804, + 0.1766047477722168, + 0.060267508029937744, + 0.05605389177799225, + 0.0486026294529438 + ], + "credit_map": [ + 0.019171619787812233, + 0.014686536975204945, + 0.007736416067928076, + 0.006715863477438688, + 0.011791720055043697, + 0.022024504840373993, + 0.006917267572134733, + 0.03754990175366402, + 0.03445957973599434, + 0.021055174991488457, + 0.009045247919857502, + 0.011157621629536152, + 0.013113084249198437, + 0.0242142491042614, + 0.011358249932527542, + 0.03250931203365326, + 0.014883496798574924, + 0.007009243592619896, + 0.005498299840837717, + 0.0050747510977089405, + 0.004601843189448118, + 0.010819677263498306, + 0.004403991159051657, + 0.014248834922909737, + 0.019084343686699867, + 0.013143115676939487, + 0.006139023695141077, + 0.005714735947549343, + 0.009067695587873459, + 0.017624540254473686, + 0.004735645838081837, + 0.02447650581598282, + 0.011075201444327831, + 0.008366418071091175, + 0.004371594171971083, + 0.0021630285773426294, + 0.003323386423289776, + 0.008231591433286667, + 0.0032274334225803614, + 0.005911238957196474, + 0.029370347037911415, + 0.016024237498641014, + 0.01341192051768303, + 0.011833010241389275, + 0.010462991893291473, + 0.027912048622965813, + 0.01173741277307272, + 0.04101727157831192, + 0.04668907821178436, + 0.01722625643014908, + 0.010248641483485699, + 0.013079429045319557, + 0.012386485002934933, + 0.03222234919667244, + 0.013770457357168198, + 0.030982038006186485, + 0.01807829551398754, + 0.006229841150343418, + 0.003997958730906248, + 0.0027176314033567905, + 0.0035155850928276777, + 0.01054232008755207, + 0.004349435213953257, + 0.01083644013851881, + 0.009108168072998524, + 0.004543478600680828, + 0.003951793536543846, + 0.004691163543611765, + 0.0037364044692367315, + 0.010116491466760635, + 0.005533350631594658, + 0.014373041689395905, + 0.010071971453726292, + 0.004777750000357628, + 0.003458042861893773, + 0.004282071720808744, + 0.005160672124475241, + 0.006848194636404514, + 0.006804383359849453, + 0.007199543062597513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.14287321269512177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a massive American flag fluttering violently against a vast, flawless azure sky. Rendered at a high frame rate, the footage captures every movement with surreal, crystal clarity. Under harsh direct sunlight, the flag's heavy fabric snaps and ripples in the powerful, unseen wind. Crimson and bright white stripes undulate in powerful, chaotic waves, while the navy blue canton, densely embroidered with 50 stars, twists and surges with immense force. The intense light meticulously defines the texture of the woven material, creating shimmering highlights at the tops of the creases and casting deep, sharp shadows within the valleys. The flag clings to the halyard atop the tall flagpole, its edge whipping violently, every ripple and twist captured without a hint of motion blur, creating a powerful and awe-inspiring atmosphere of raw, untamed energy and enduring power.", + "seed": 15324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_12/index_4417_seed_15324_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": -0.31875520646572114, + "vlm_logprob": -0.31875520646572114, + "vlm_yes_no_margin": 2.2749999523162843, + "vlm_token_credit": -0.31875520646572114, + "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 taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03230779990553856 + ], + "opposite_logprob_tokens": [ + -4.0323076248168945 + ], + "logprob_reward": -0.03230779990553856, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03230779990553856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20443201065063477, + 0.09076634794473648, + 0.08360454440116882, + 0.14552295207977295, + 0.06629907339811325, + 0.0733434334397316, + 0.08384377509355545, + 0.055403418838977814, + 0.1540132611989975, + 0.04277116805315018 + ], + "credit_map": [ + 0.04583863168954849, + 0.02388618513941765, + 0.019909042865037918, + 0.01495551411062479, + 0.017404993996024132, + 0.027248304337263107, + 0.03174719959497452, + 0.023442130535840988, + 0.01091261301189661, + 0.006274914834648371, + 0.006061190739274025, + 0.004339344799518585, + 0.007836275734007359, + 0.01674169860780239, + 0.009492450393736362, + 0.02910786122083664, + 0.01813519187271595, + 0.013062257319688797, + 0.007760179694741964, + 0.004706880077719688, + 0.00840581301599741, + 0.012606076896190643, + 0.006377530284225941, + 0.01255061849951744, + 0.02804073691368103, + 0.020654572173953056, + 0.011493860743939877, + 0.009890155866742134, + 0.018964743241667747, + 0.018621740862727165, + 0.017126675695180893, + 0.02073047310113907, + 0.0173816978931427, + 0.009062053635716438, + 0.007908552885055542, + 0.004437681287527084, + 0.00618124520406127, + 0.008572463877499104, + 0.0066907769069075584, + 0.006064601708203554, + 0.012635752558708191, + 0.008655975572764874, + 0.010222848504781723, + 0.006971672177314758, + 0.006748703774064779, + 0.00952167809009552, + 0.007506698369979858, + 0.011080103926360607, + 0.02199563756585121, + 0.011903427541255951, + 0.010610844939947128, + 0.00428372249007225, + 0.007827657274901867, + 0.012682009488344193, + 0.005573874339461327, + 0.008966607041656971, + 0.013189642690122128, + 0.007492808625102043, + 0.005486249458044767, + 0.00335527746938169, + 0.0051707010716199875, + 0.00856087077409029, + 0.004778699018061161, + 0.007369170431047678, + 0.047793298959732056, + 0.015452953986823559, + 0.01995187997817993, + 0.006620990578085184, + 0.012147936038672924, + 0.02345210313796997, + 0.008839716203510761, + 0.019754381850361824, + 0.011392227374017239, + 0.004306115210056305, + 0.005248943343758583, + 0.002634417498484254, + 0.0035874899476766586, + 0.00733232730999589, + 0.0026590072084218264, + 0.005610640626400709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.03230779990553856 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983442321419716 + ], + "opposite_logprob_tokens": [ + -3.529834508895874 + ], + "logprob_reward": -0.02983442321419716, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983442321419716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10012741386890411, + 0.07468116283416748, + 0.0603095106780529, + 0.11737588047981262, + 0.03246554732322693, + 0.1273714154958725, + 0.12001405656337738, + 0.19189675152301788, + 0.054197296500205994, + 0.12156093865633011 + ], + "credit_map": [ + 0.019755681976675987, + 0.018840931355953217, + 0.010791189968585968, + 0.0048148310743272305, + 0.013089228421449661, + 0.015696752816438675, + 0.008805699646472931, + 0.008333093486726284, + 0.00896554533392191, + 0.015020006336271763, + 0.004307144321501255, + 0.004369417205452919, + 0.005088035482913256, + 0.010498706251382828, + 0.010025744326412678, + 0.016406560316681862, + 0.011730906553566456, + 0.009893112815916538, + 0.006539749447256327, + 0.0034354703966528177, + 0.006444558966904879, + 0.00874052383005619, + 0.004288352560251951, + 0.009236833080649376, + 0.033172838389873505, + 0.013408803381025791, + 0.012198921293020248, + 0.007560516707599163, + 0.01185590773820877, + 0.01660616137087345, + 0.011826290749013424, + 0.010746439918875694, + 0.007937734946608543, + 0.004457194823771715, + 0.0033159449230879545, + 0.0020975987426936626, + 0.0030819163657724857, + 0.004834023304283619, + 0.0027623323258012533, + 0.003978803288191557, + 0.03190390020608902, + 0.026349253952503204, + 0.015187161043286324, + 0.00629560649394989, + 0.010413207113742828, + 0.01984216831624508, + 0.005316460505127907, + 0.01206367090344429, + 0.019730011001229286, + 0.01715262606739998, + 0.01199681032449007, + 0.009239057078957558, + 0.011785637587308884, + 0.017087114974856377, + 0.01528521440923214, + 0.01773759350180626, + 0.03344181552529335, + 0.022404378280043602, + 0.022044457495212555, + 0.015052113682031631, + 0.022054096683859825, + 0.025406697764992714, + 0.0248115174472332, + 0.026681676506996155, + 0.016254527494311333, + 0.004786255769431591, + 0.006917600519955158, + 0.0023374243173748255, + 0.005564265418797731, + 0.008867050521075726, + 0.0032365128863602877, + 0.006233659107238054, + 0.03951577469706535, + 0.010150549933314323, + 0.01971428282558918, + 0.006052700337022543, + 0.008849251084029675, + 0.01940336637198925, + 0.006726352497935295, + 0.011148665100336075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.02983442321419716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812128305435181 + ], + "opposite_logprob_tokens": [ + -3.288121223449707 + ], + "logprob_reward": -0.03812128305435181, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812128305435181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25089460611343384, + 0.13809168338775635, + 0.08828505128622055, + 0.03911028057336807, + 0.0828617736697197, + 0.09121765196323395, + 0.039397191256284714, + 0.08699440211057663, + 0.04589628055691719, + 0.1372510939836502 + ], + "credit_map": [ + 0.038243718445301056, + 0.04866601154208183, + 0.03170047700405121, + 0.02268095873296261, + 0.03020879253745079, + 0.03513883799314499, + 0.02074838988482952, + 0.023507408797740936, + 0.018995089456439018, + 0.024291669949889183, + 0.008632340468466282, + 0.0076826936565339565, + 0.012955467216670513, + 0.02087552286684513, + 0.010685853660106659, + 0.03397304192185402, + 0.021160827949643135, + 0.012200026772916317, + 0.00675631407648325, + 0.007031959015876055, + 0.008384392596781254, + 0.012267882004380226, + 0.0077922227792441845, + 0.012691417708992958, + 0.008150333538651466, + 0.005598639603704214, + 0.0033339106012135744, + 0.002598677296191454, + 0.005064767319709063, + 0.005465395282953978, + 0.005048801191151142, + 0.003849752712994814, + 0.011704823933541775, + 0.013966870494186878, + 0.006136428564786911, + 0.009071615524590015, + 0.008169170469045639, + 0.011441434733569622, + 0.01036921888589859, + 0.012002203613519669, + 0.012309188023209572, + 0.016441717743873596, + 0.008672291412949562, + 0.00933592114597559, + 0.010073969140648842, + 0.011982308700680733, + 0.010235374793410301, + 0.01216687262058258, + 0.007798584643751383, + 0.005060320254415274, + 0.00534659530967474, + 0.0024956255219876766, + 0.003591529792174697, + 0.005787796340882778, + 0.003888447070494294, + 0.005428288131952286, + 0.016731910407543182, + 0.012201768346130848, + 0.012030580081045628, + 0.004653105977922678, + 0.011142591945827007, + 0.013498680666089058, + 0.0074028377421200275, + 0.009332915768027306, + 0.010471406392753124, + 0.005748445633798838, + 0.00517145125195384, + 0.0022925836965441704, + 0.004518980160355568, + 0.007281915750354528, + 0.0038178579416126013, + 0.0065936376340687275, + 0.029551859945058823, + 0.017119605094194412, + 0.014905106276273727, + 0.01044419128447771, + 0.012541603296995163, + 0.023369107395410538, + 0.011130880564451218, + 0.018188731744885445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03812128305435181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.313529372215271 + ], + "opposite_logprob_tokens": [ + -0.3135293424129486 + ], + "logprob_reward": -1.313529372215271, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.313529372215271, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20332951843738556, + 0.1560535877943039, + 0.13423703610897064, + 0.05589351803064346, + 0.09812171012163162, + 0.03516174107789993, + 0.08558116853237152, + 0.09246826171875, + 0.0879456028342247, + 0.05120781064033508 + ], + "credit_map": [ + 0.034674614667892456, + 0.027614399790763855, + 0.01913974992930889, + 0.015548991039395332, + 0.016895322129130363, + 0.02538572996854782, + 0.04280181601643562, + 0.021268900483846664, + 0.014112381264567375, + 0.031668271869421005, + 0.009275591932237148, + 0.008711338974535465, + 0.0193877462297678, + 0.022807596251368523, + 0.02069513127207756, + 0.029395537450909615, + 0.024411706253886223, + 0.016404984518885612, + 0.011789048090577126, + 0.008411348797380924, + 0.012880665250122547, + 0.01964205875992775, + 0.012179242447018623, + 0.028517989441752434, + 0.007833282463252544, + 0.01392098143696785, + 0.003453858196735382, + 0.0037872223183512688, + 0.005582932848483324, + 0.008387381210923195, + 0.0066660866141319275, + 0.006261770147830248, + 0.029809173196554184, + 0.014594176784157753, + 0.008744894526898861, + 0.0057226927019655704, + 0.008113150484859943, + 0.016385551542043686, + 0.005605135578662157, + 0.009146937169134617, + 0.007449484895914793, + 0.004636388272047043, + 0.0035071526654064655, + 0.002398641547188163, + 0.003414237406104803, + 0.005762776359915733, + 0.0032852820586413145, + 0.004707775544375181, + 0.009443067945539951, + 0.011548582464456558, + 0.008908230811357498, + 0.006955472286790609, + 0.010346528142690659, + 0.013595974072813988, + 0.007607327774167061, + 0.017175983637571335, + 0.011894595809280872, + 0.013246802613139153, + 0.008974817581474781, + 0.006289519835263491, + 0.01228783093392849, + 0.01233150064945221, + 0.015218116343021393, + 0.012225082144141197, + 0.017160026356577873, + 0.012625095434486866, + 0.008862625807523727, + 0.004363140556961298, + 0.008467729203402996, + 0.013382414355874062, + 0.0067733279429376125, + 0.016311246901750565, + 0.015386066399514675, + 0.007604776415973902, + 0.006081585306674242, + 0.002118474105373025, + 0.0042392779141664505, + 0.008381161838769913, + 0.0029457525815814734, + 0.00445071654394269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -1.313529372215271 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17998315393924713 + ], + "opposite_logprob_tokens": [ + -1.804983139038086 + ], + "logprob_reward": -0.17998315393924713, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17998315393924713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17873001098632812, + 0.1286637932062149, + 0.20545747876167297, + 0.11333424597978592, + 0.10483187437057495, + 0.0589229017496109, + 0.039007894694805145, + 0.09702068567276001, + 0.042955342680215836, + 0.03107578307390213 + ], + "credit_map": [ + 0.02780274860560894, + 0.03888503089547157, + 0.013685530982911587, + 0.012725651264190674, + 0.013824693858623505, + 0.02495586685836315, + 0.03246454894542694, + 0.014385927468538284, + 0.021503781899809837, + 0.02993261069059372, + 0.008878969587385654, + 0.006640574894845486, + 0.01042136736214161, + 0.019171789288520813, + 0.006850693840533495, + 0.025263987481594086, + 0.033106956630945206, + 0.034562595188617706, + 0.017829058691859245, + 0.013449590653181076, + 0.022268492728471756, + 0.03062385693192482, + 0.016109775751829147, + 0.03750711679458618, + 0.03148237615823746, + 0.015758998692035675, + 0.012529713101685047, + 0.00626731850206852, + 0.010612514801323414, + 0.017255296930670738, + 0.010302213951945305, + 0.00912579894065857, + 0.03240234777331352, + 0.012458410114049911, + 0.008794985711574554, + 0.00743058230727911, + 0.007425409276038408, + 0.016079891473054886, + 0.007334442809224129, + 0.0129057876765728, + 0.012955712154507637, + 0.0069451904855668545, + 0.007110167760401964, + 0.005751027725636959, + 0.004235554952174425, + 0.0077874730341136456, + 0.007623687852174044, + 0.006514081731438637, + 0.007509928196668625, + 0.006227681878954172, + 0.003530279267579317, + 0.0031486263033002615, + 0.004053645301610231, + 0.005761665292084217, + 0.003362819319590926, + 0.0054132468067109585, + 0.011132853105664253, + 0.010904766619205475, + 0.00827306229621172, + 0.006585671100765467, + 0.011498291045427322, + 0.014142514206469059, + 0.01718299090862274, + 0.017300529405474663, + 0.00897392351180315, + 0.006025935057550669, + 0.00435556098818779, + 0.001774562057107687, + 0.005025225225836039, + 0.006659210193902254, + 0.0036511330399662256, + 0.006489786319434643, + 0.008245378732681274, + 0.0032527847215533257, + 0.004426541738212109, + 0.0016974039608612657, + 0.002285704715177417, + 0.0056252554059028625, + 0.0018387766322121024, + 0.0037039320450276136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.17998315393924713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a massive American flag fluttering violently against a vast, flawless azure sky. Rendered at a high frame rate, the footage captures every movement with surreal, crystal clarity. Under harsh direct sunlight, the flag's heavy fabric snaps and ripples in the powerful, unseen wind. Crimson and bright white stripes undulate in powerful, chaotic waves, while the navy blue canton, densely embroidered with 50 stars, twists and surges with immense force. The intense light meticulously defines the texture of the woven material, creating shimmering highlights at the tops of the creases and casting deep, sharp shadows within the valleys. The flag clings to the halyard atop the tall flagpole, its edge whipping violently, every ripple and twist captured without a hint of motion blur, creating a powerful and awe-inspiring atmosphere of raw, untamed energy and enduring power.", + "seed": 15325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_12/index_4417_seed_15325_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": -0.32120292335748674, + "vlm_logprob": -0.32120292335748674, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.32120292335748674, + "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 taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01573069579899311 + ], + "opposite_logprob_tokens": [ + -5.140730857849121 + ], + "logprob_reward": -0.01573069579899311, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01573069579899311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009266308508813381, + 0.004252767190337181, + 0.008391308598220348, + 0.948056161403656, + 0.0020160807762295008, + 0.012825759127736092, + 0.009110230021178722, + 0.0020286832004785538, + 0.0016190564492717385, + 0.0024336539208889008 + ], + "credit_map": [ + 0.0022398210130631924, + 0.0017612590454518795, + 0.0008013731567189097, + 0.00038398339529521763, + 0.0006322170956991613, + 0.0025318560656160116, + 0.00033701551728881896, + 0.0005787816480733454, + 0.0008246077923104167, + 0.0012835902161896229, + 0.00046156084863469005, + 0.0002970585192088038, + 0.00032789030228741467, + 0.0005531295319087803, + 0.00020331534324213862, + 0.00030161411268636584, + 0.0019727738108485937, + 0.0026494308840483427, + 0.0006845142925158143, + 0.0004234350344631821, + 0.0008940123952925205, + 0.0005624915356747806, + 0.0006063690525479615, + 0.0005982806324027479, + 0.2475813627243042, + 0.4245610535144806, + 0.034966059029102325, + 0.013931294903159142, + 0.11058275401592255, + 0.08814053237438202, + 0.013947559520602226, + 0.0143455034121871, + 0.00033080248977057636, + 0.0004583674017339945, + 0.00022522240760736167, + 9.845324530033395e-05, + 0.00015213301230687648, + 0.0002969469060190022, + 8.811146108200774e-05, + 0.00036604382330551744, + 0.002079940401017666, + 0.003322426462545991, + 0.0017808524426072836, + 0.0007631938206031919, + 0.0008508570026606321, + 0.001095819054171443, + 0.0006014059181325138, + 0.0023312633857131004, + 0.0020028201397508383, + 0.0017190384678542614, + 0.0020116877276450396, + 0.0003898020659107715, + 0.0005261512706056237, + 0.0008523787255398929, + 0.000529889774043113, + 0.0010784611804410815, + 0.0003284546546638012, + 0.00047333119437098503, + 0.00027569846133701503, + 0.00011894434283021837, + 0.0001880056515801698, + 0.00018117066065315157, + 0.00011799443018389866, + 0.000345083448337391, + 0.0004036462341900915, + 0.0005164168542250991, + 0.00019172778411302716, + 8.355256431968883e-05, + 0.000139945448609069, + 9.521201718598604e-05, + 7.249190093716606e-05, + 0.00011606349289650097, + 0.00035322626354172826, + 0.0006259760702960193, + 0.0002271098201163113, + 0.0001509134453954175, + 0.0003703501133713871, + 0.0003792770439758897, + 0.00011363883095327765, + 0.00021316208585631102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.01573069579899311 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026381956413388252 + ], + "opposite_logprob_tokens": [ + -3.6513819694519043 + ], + "logprob_reward": -0.026381956413388252, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026381956413388252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062077220529317856, + 0.0636870339512825, + 0.1211017519235611, + 0.5306124091148376, + 0.03656264394521713, + 0.06673364341259003, + 0.03750387579202652, + 0.037151481956243515, + 0.026747265830636024, + 0.017822779715061188 + ], + "credit_map": [ + 0.013923963531851768, + 0.013444412499666214, + 0.00778698455542326, + 0.003996196668595076, + 0.005502958316355944, + 0.008473003283143044, + 0.003321317723020911, + 0.005628379061818123, + 0.015012306161224842, + 0.016477545723319054, + 0.006282454822212458, + 0.003945716191083193, + 0.004971718415617943, + 0.00469816243276, + 0.005307906772941351, + 0.006991216447204351, + 0.022487519308924675, + 0.023784007877111435, + 0.009892617352306843, + 0.0077388230711221695, + 0.02073659747838974, + 0.015327549539506435, + 0.009792742319405079, + 0.011341884732246399, + 0.13735151290893555, + 0.23640812933444977, + 0.02033260464668274, + 0.008900417946279049, + 0.06299301236867905, + 0.0486561618745327, + 0.007960883900523186, + 0.008009654469788074, + 0.010501392185688019, + 0.007844578474760056, + 0.0036648004315793514, + 0.0013818095903843641, + 0.0020582114811986685, + 0.00628753611817956, + 0.0013746994081884623, + 0.0034496134612709284, + 0.01126865204423666, + 0.010452826507389545, + 0.012163622304797173, + 0.0038188304752111435, + 0.004676495213061571, + 0.005835510324686766, + 0.002858698135241866, + 0.01565900258719921, + 0.007708876859396696, + 0.00873593706637621, + 0.007450449746102095, + 0.002144008642062545, + 0.0027103573083877563, + 0.0033935855608433485, + 0.0016469537513330579, + 0.003713707672432065, + 0.007462724577635527, + 0.009060087613761425, + 0.005754854064434767, + 0.0019496857421472669, + 0.002987087471410632, + 0.0027934126555919647, + 0.0020454751793295145, + 0.005098151508718729, + 0.004925544373691082, + 0.008414787240326405, + 0.003014126792550087, + 0.0013265797169879079, + 0.003052199725061655, + 0.0017732518026605248, + 0.001089835655875504, + 0.0031509401742368937, + 0.0028095184825360775, + 0.004294977057725191, + 0.001961521804332733, + 0.0009731437312439084, + 0.0014490497997030616, + 0.002849883632734418, + 0.0008027087897062302, + 0.0026819745544344187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.026381956413388252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867884889245033 + ], + "opposite_logprob_tokens": [ + -3.0486788749694824 + ], + "logprob_reward": -0.04867884889245033, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867884889245033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009635957889258862, + 0.007347447797656059, + 0.008833184838294983, + 0.9198063015937805, + 0.0021587845403701067, + 0.012618064880371094, + 0.008844075724482536, + 0.005954358726739883, + 0.007189114112406969, + 0.017612682655453682 + ], + "credit_map": [ + 0.0026007643900811672, + 0.0012169895926490426, + 0.0008212280226871371, + 0.00037224424886517227, + 0.0006981338956393301, + 0.0029333794955164194, + 0.000406141160055995, + 0.0005870771128684282, + 0.0012106425128877163, + 0.0016770726069808006, + 0.0007983361720107496, + 0.0005100917187519372, + 0.000669126573484391, + 0.0009440111462026834, + 0.0008031784673221409, + 0.0007349893567152321, + 0.00176862848456949, + 0.0018002876313403249, + 0.0006246693665161729, + 0.0005348065751604736, + 0.0008868129807524383, + 0.0009008863708004355, + 0.0012736301869153976, + 0.0010434632422402501, + 0.23973776400089264, + 0.41209569573402405, + 0.03395851328969002, + 0.013521756045520306, + 0.10753456503152847, + 0.08557727187871933, + 0.013502005487680435, + 0.01387878693640232, + 0.0003109776589553803, + 0.0005014189518988132, + 0.00020473074982874095, + 0.00012702294043265283, + 0.00028928721440024674, + 0.0002791489823721349, + 0.00014494781498797238, + 0.00030125046032480896, + 0.002309259958565235, + 0.0017420257208868861, + 0.002462149830535054, + 0.0005537413526326418, + 0.0006456326227635145, + 0.0011890166206285357, + 0.0005990212666802108, + 0.0031172188464552164, + 0.0019247664604336023, + 0.0024401594419032335, + 0.0008421828388236463, + 0.00041125048301182687, + 0.0006555883446708322, + 0.0008038830128498375, + 0.0004082363157067448, + 0.0013580097584053874, + 0.0014227567007765174, + 0.0016541715012863278, + 0.0007604992133565247, + 0.00022330941283144057, + 0.0004323525936342776, + 0.00039564131293445826, + 0.00034706684527918696, + 0.0007185615249909461, + 0.00106232485268265, + 0.001654324820265174, + 0.0007925236131995916, + 0.0004199556424282491, + 0.0011519461404532194, + 0.0005927695310674608, + 0.0004557160718832165, + 0.0010595542844384909, + 0.0026416678447276354, + 0.002568101743236184, + 0.002038130536675453, + 0.0009112686384469271, + 0.0020741750486195087, + 0.002999448450282216, + 0.001133417128585279, + 0.0032464752439409494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.04867884889245033 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.3135493993759155 + ], + "opposite_logprob_tokens": [ + -0.31354936957359314 + ], + "logprob_reward": -1.3135493993759155, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3135493993759155, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.007260708138346672, + 0.014161071740090847, + 0.00896814838051796, + 0.9262403845787048, + 0.006296580191701651, + 0.017050880938768387, + 0.007997876964509487, + 0.0046738870441913605, + 0.0035826184321194887, + 0.0037678503431379795 + ], + "credit_map": [ + 0.0019088247790932655, + 0.001225702348165214, + 0.00046454567927867174, + 0.00029406382236629725, + 0.0004989682929590344, + 0.0020424562972038984, + 0.0002871556207537651, + 0.0005389915895648301, + 0.0035583525896072388, + 0.0029609219636768103, + 0.0017119435360655189, + 0.0007984201656654477, + 0.0010563783580437303, + 0.0011994187952950597, + 0.0010212418856099248, + 0.0018543944461271167, + 0.0014495417708531022, + 0.0018840370466932654, + 0.0008062849519774318, + 0.000513591745402664, + 0.0009858862031251192, + 0.0009542091283947229, + 0.0008967746398411691, + 0.0014778224285691977, + 0.24157047271728516, + 0.41440725326538086, + 0.0341661274433136, + 0.013570962473750114, + 0.10809463262557983, + 0.0862913653254509, + 0.014010060578584671, + 0.014129501767456532, + 0.0012167816748842597, + 0.001529228058643639, + 0.00044564480776898563, + 0.00035958734224550426, + 0.0004464654193725437, + 0.000997996423393488, + 0.0002999968419317156, + 0.0010008793324232101, + 0.0030910740606486797, + 0.002993826987221837, + 0.003052076091989875, + 0.0007920284988358617, + 0.0015552469994872808, + 0.0017150716157630086, + 0.0007825957727618515, + 0.0030689616687595844, + 0.0013598192017525434, + 0.0017662630416452885, + 0.0005646999925374985, + 0.000384991813916713, + 0.0007929918938316405, + 0.0011259624734520912, + 0.0005135430837981403, + 0.0014896055217832327, + 0.0011081892298534513, + 0.0012873011874035, + 0.0006965327775105834, + 0.0002236026048194617, + 0.00036004578578285873, + 0.0002903908025473356, + 0.0002292072313139215, + 0.00047861761413514614, + 0.0008379325154237449, + 0.0013579906662926078, + 0.00032494132756255567, + 0.00018461533181834966, + 0.00030360667733475566, + 0.0002637751749716699, + 0.0001228180481120944, + 0.00018693838501349092, + 0.0005726309027522802, + 0.0008181036100722849, + 0.0004261434078216553, + 0.000216139160329476, + 0.000442728225607425, + 0.0006119097233749926, + 0.00016655241779517382, + 0.0005136429099366069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -1.3135493993759155 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking up at the flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a vast, flawless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the American flag shown with crimson and bright white stripes and a navy blue canton with 50 stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flag fluttering violently with visible motion blur absent (crystal clarity)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flagpole appear tall with the flag clinging to the halyard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016737163066864 + ], + "opposite_logprob_tokens": [ + -1.7016737461090088 + ], + "logprob_reward": -0.2016737163066864, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016737163066864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0035733296535909176, + 0.003374747931957245, + 0.004972368944436312, + 0.955291211605072, + 0.004173142835497856, + 0.018198426812887192, + 0.0030461985152214766, + 0.002765882760286331, + 0.0018234887393191457, + 0.0027811776380985975 + ], + "credit_map": [ + 0.0008131759241223335, + 0.0006319262902252376, + 0.00027341305394656956, + 0.00019801451708190143, + 0.0003199805214535445, + 0.0009150447440333664, + 0.0001419507898390293, + 0.0002798241621349007, + 0.0008081186679191887, + 0.0007837058510631323, + 0.00039153650868684053, + 0.00015504872135352343, + 0.0002921214036177844, + 0.00037680697278119624, + 0.00032568976166658103, + 0.00024172032135538757, + 0.0009540211176499724, + 0.0017017755890265107, + 0.0004965246189385653, + 0.00025627881404943764, + 0.0004790586535818875, + 0.0004043036897201091, + 0.0003202623629476875, + 0.0003601447388064116, + 0.24941863119602203, + 0.42738789319992065, + 0.0352676585316658, + 0.014054752886295319, + 0.11145175248384476, + 0.08880570530891418, + 0.014362305402755737, + 0.01454260852187872, + 0.0004901979700662196, + 0.0010261139832437038, + 0.0004110366862732917, + 0.0002404098049737513, + 0.0003943317860830575, + 0.0006996187148615718, + 0.00021469641069415957, + 0.0006967377848923206, + 0.0029726766515523195, + 0.004514372441917658, + 0.0035834691952914, + 0.001056653680279851, + 0.0016599280061200261, + 0.0017533834325149655, + 0.0008026894647628069, + 0.001855255221016705, + 0.0005472840857692063, + 0.0005741170607507229, + 0.0005526656750589609, + 0.0002159578143619001, + 0.000257998937740922, + 0.0003975986910518259, + 0.00018974700651597232, + 0.000310829549562186, + 0.0007077729678712785, + 0.0007211833144538105, + 0.0005857022479176521, + 0.00013722071889787912, + 0.00014595828542951494, + 0.00016604388656560332, + 0.0001141997126978822, + 0.00018780192476697266, + 0.00041335250716656446, + 0.0005676529253832996, + 0.00017471041064709425, + 9.248179412679747e-05, + 0.000198421097593382, + 0.00012214590969961137, + 6.940512685105205e-05, + 0.00018531885871198028, + 0.0004494462627917528, + 0.0007037735194899142, + 0.0003798829857259989, + 0.00014586671022698283, + 0.0004689132620114833, + 0.0003185180830769241, + 0.00012137871090089902, + 0.00019339834398124367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.2016737163066864 + } + ], + "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/0000025/group_13/index_tensor([9319, 9319])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_13/index_tensor([9319, 9319])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7c19c2910b65fe749e4ffe61c45662e0567d4f04 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_13/index_tensor([9319, 9319])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene begins with a medium-wide shot, rising steadily upward to reveal a terrifying, expressionistic landscape painted with chaotic, emotive brushstrokes. A raging river of molten lava and lustrous obsidian churns violently. The liquid gold radiates an intense internal heat, casting a flickering, infernal light, while sharp, cooling obsidian fragments swirl like black daggers in the powerful current. Navigating this vortex is a lone, male and female figure atop a simple raft composed of dark, rough logs lashed with thick ropes. The figure is clad in a heavy, shapeless black robe and wears a mask resembling a stylized frog, its features carved with stark, unsettling lines and painted a pale, ghostly white, starkly capturing the incandescent light from below. With intense concentration, the figure grasps a long, rough wooden stick and deliberately and forcefully plunges it into the swirling current, pushing aside jagged rocks. Their movements resemble a desperate ballet against the river's wrath. The entire scene is rendered in thick, fresco-like textures and distorted shapes; light from the molten gold creates dramatic, high-contrast black and whites, casting long, dancing shadows across the bruised purple and indigo cliffs that tilt oppressively inward. As the camera completes its ascent from a high-angle, bird's-eye view, the figure and raft shrink to a tiny, solitary speck within a vast, swirling vortex of light and dark, emphasizing an atmosphere of profound isolation and an epic, desperate struggle against overwhelming, primal power.", + "seed": 17719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_13/index_9319_seed_17719_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": -0.18182055354118348, + "vlm_logprob": -0.18182055354118348, + "vlm_yes_no_margin": 1.9, + "vlm_token_credit": -0.18182055354118348, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.40424492955207825 + ], + "opposite_logprob_tokens": [ + -1.1542448997497559 + ], + "logprob_reward": -0.40424492955207825, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.40424492955207825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18368662893772125, + 0.19123321771621704, + 0.026197293773293495, + 0.020529204979538918, + 0.133891299366951, + 0.05716067925095558, + 0.03450324386358261, + 0.24294140934944153, + 0.07197168469429016, + 0.037885308265686035 + ], + "credit_map": [ + 0.019530029967427254, + 0.0158134326338768, + 0.019048497080802917, + 0.025883665308356285, + 0.0236965399235487, + 0.018606087192893028, + 0.03679366782307625, + 0.024314705282449722, + 0.020372387021780014, + 0.03490844741463661, + 0.023285679519176483, + 0.021008534356951714, + 0.025694435462355614, + 0.021863579750061035, + 0.02144693210721016, + 0.02265322022140026, + 0.0033365271519869566, + 0.0038046964909881353, + 0.002675356576219201, + 0.0033725635148584843, + 0.0039433944039046764, + 0.003045033197849989, + 0.002866612281650305, + 0.0031531096901744604, + 0.0034425663761794567, + 0.00277468329295516, + 0.002474401146173477, + 0.0037396110128611326, + 0.0018822506535798311, + 0.0026640600990504026, + 0.001489141141064465, + 0.002062492771074176, + 0.013150849379599094, + 0.009449329227209091, + 0.017409447580575943, + 0.041805703192949295, + 0.010508221574127674, + 0.013577982783317566, + 0.012636210769414902, + 0.015353563241660595, + 0.00831314455717802, + 0.008597146719694138, + 0.004053914453834295, + 0.004893498960882425, + 0.005925959441810846, + 0.014019375666975975, + 0.004572475329041481, + 0.006785163655877113, + 0.005839048884809017, + 0.004076703451573849, + 0.0030931197106838226, + 0.004102219361811876, + 0.0033562383614480495, + 0.007983176968991756, + 0.002266193740069866, + 0.0037865405902266502, + 0.032826248556375504, + 0.035849809646606445, + 0.03426191955804825, + 0.026501208543777466, + 0.024388372898101807, + 0.027799872681498528, + 0.02101135440170765, + 0.04030263051390648, + 0.011016051284968853, + 0.009584357962012291, + 0.010175333358347416, + 0.0068934704177081585, + 0.0035806309897452593, + 0.010603315196931362, + 0.004726008512079716, + 0.015392515808343887, + 0.005279047414660454, + 0.004710909444838762, + 0.005757337901741266, + 0.005811756011098623, + 0.0033735218457877636, + 0.0027642296627163887, + 0.006120907608419657, + 0.004067597910761833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.40424492955207825 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871699959039688 + ], + "opposite_logprob_tokens": [ + -3.0487170219421387 + ], + "logprob_reward": -0.04871699959039688, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871699959039688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2614569664001465, + 0.12301593273878098, + 0.11040176451206207, + 0.08676354587078094, + 0.06784872710704803, + 0.16233417391777039, + 0.03909046947956085, + 0.04664032533764839, + 0.056466247886419296, + 0.04598178714513779 + ], + "credit_map": [ + 0.024153800681233406, + 0.045599352568387985, + 0.027233222499489784, + 0.030714279040694237, + 0.03422926738858223, + 0.03142399713397026, + 0.03678110986948013, + 0.0313219353556633, + 0.012901795096695423, + 0.025468945503234863, + 0.01577962189912796, + 0.014037027955055237, + 0.01726214960217476, + 0.011777753941714764, + 0.012947596609592438, + 0.012841039337217808, + 0.01260290015488863, + 0.02520669624209404, + 0.0112159363925457, + 0.014157175086438656, + 0.012967132031917572, + 0.012061424553394318, + 0.011733735911548138, + 0.010456757619976997, + 0.010096902027726173, + 0.013147791847586632, + 0.01154483761638403, + 0.013220599852502346, + 0.012763068079948425, + 0.0075676352716982365, + 0.01061172317713499, + 0.007810990326106548, + 0.008077943697571754, + 0.006074638105928898, + 0.006496184039860964, + 0.015559641644358635, + 0.009824014268815517, + 0.00930548831820488, + 0.005876403301954269, + 0.006634415127336979, + 0.023379076272249222, + 0.023890187963843346, + 0.01596607267856598, + 0.020987341180443764, + 0.01559901051223278, + 0.03165597468614578, + 0.01481242012232542, + 0.01604410447180271, + 0.005941367708146572, + 0.004559868946671486, + 0.003479171311482787, + 0.005197728518396616, + 0.0028062101919203997, + 0.009641244076192379, + 0.0033147933427244425, + 0.004150086548179388, + 0.006507787387818098, + 0.004212293308228254, + 0.005444187670946121, + 0.007564817555248737, + 0.003506359877064824, + 0.007643768563866615, + 0.005157619249075651, + 0.006603495683521032, + 0.008951336145401001, + 0.0061199478805065155, + 0.008182511664927006, + 0.004833812825381756, + 0.003246472217142582, + 0.008505688980221748, + 0.004513184539973736, + 0.012113295495510101, + 0.009552584029734135, + 0.004948820453137159, + 0.007567753084003925, + 0.0053352550603449345, + 0.003084818134084344, + 0.005301882047206163, + 0.0044170250184834, + 0.005773650016635656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04871699959039688 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07011917978525162 + ], + "opposite_logprob_tokens": [ + -2.6951191425323486 + ], + "logprob_reward": -0.07011917978525162, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07011917978525162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27592864632606506, + 0.13170857727527618, + 0.07484720647335052, + 0.1445624977350235, + 0.05204392597079277, + 0.08794644474983215, + 0.040178082883358, + 0.06499313563108444, + 0.07472029328346252, + 0.053071167320013046 + ], + "credit_map": [ + 0.027784014120697975, + 0.03468064218759537, + 0.02581288293004036, + 0.0297250933945179, + 0.02852031961083412, + 0.05198780447244644, + 0.03297974541783333, + 0.04443814605474472, + 0.013431806117296219, + 0.028034836053848267, + 0.015945758670568466, + 0.015614527277648449, + 0.015664393082261086, + 0.013754227198660374, + 0.014300165697932243, + 0.014962847344577312, + 0.008644839748740196, + 0.014935708604753017, + 0.0075529529713094234, + 0.01042852457612753, + 0.009528066031634808, + 0.008773994632065296, + 0.0079086534678936, + 0.007074467837810516, + 0.019662655889987946, + 0.028078721836209297, + 0.015589639544487, + 0.027444161474704742, + 0.014281834475696087, + 0.012502887286245823, + 0.014670651406049728, + 0.012331937439739704, + 0.00651449803262949, + 0.006219156086444855, + 0.004508531652390957, + 0.012928339652717113, + 0.0038899879436939955, + 0.006421795580536127, + 0.004648720379918814, + 0.006912897806614637, + 0.012059016153216362, + 0.01732933148741722, + 0.006989487912505865, + 0.00882479827851057, + 0.009239737875759602, + 0.018495624884963036, + 0.0069109066389501095, + 0.00809753779321909, + 0.006487393751740456, + 0.005217601079493761, + 0.0026531131006777287, + 0.004130433779209852, + 0.0029586413875222206, + 0.009772860445082188, + 0.002720624441280961, + 0.006237414665520191, + 0.00989177543669939, + 0.010080945678055286, + 0.006812451872974634, + 0.010895275510847569, + 0.004783628974109888, + 0.007481696084141731, + 0.004828635137528181, + 0.010218723677098751, + 0.01304940227419138, + 0.007828902453184128, + 0.013063442893326283, + 0.006458228919655085, + 0.003404841758310795, + 0.009362604469060898, + 0.00473166536539793, + 0.016821201890707016, + 0.012590152211487293, + 0.004884378518909216, + 0.01340665016323328, + 0.003806040622293949, + 0.003517565317451954, + 0.005437903106212616, + 0.0035978006199002266, + 0.005830672569572926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.07011917978525162 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256530523300171 + ], + "opposite_logprob_tokens": [ + -1.600653052330017 + ], + "logprob_reward": -0.2256530523300171, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256530523300171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.206404909491539, + 0.18223878741264343, + 0.07649604976177216, + 0.04446312040090561, + 0.05222538486123085, + 0.13674499094486237, + 0.03480011597275734, + 0.1404801309108734, + 0.07814331352710724, + 0.04800323024392128 + ], + "credit_map": [ + 0.02412840910255909, + 0.022263778373599052, + 0.02232346124947071, + 0.028508957475423813, + 0.02798897586762905, + 0.02161233499646187, + 0.03328307718038559, + 0.02629593387246132, + 0.018676448613405228, + 0.032977718859910965, + 0.02307368814945221, + 0.021494997665286064, + 0.02704276144504547, + 0.019189098849892616, + 0.01955011673271656, + 0.020233968272805214, + 0.008305957540869713, + 0.017587382346391678, + 0.008270720951259136, + 0.011846582405269146, + 0.009810362011194229, + 0.0067439270205795765, + 0.007797035854309797, + 0.006134084891527891, + 0.006072872318327427, + 0.006748890038579702, + 0.004720193799585104, + 0.006952083669602871, + 0.00540783628821373, + 0.005778050050139427, + 0.003540834179148078, + 0.005242359824478626, + 0.00663326820358634, + 0.0043861400336027145, + 0.005049422383308411, + 0.010110904462635517, + 0.006474870722740889, + 0.006724752485752106, + 0.005300575401633978, + 0.007545449770987034, + 0.018557382747530937, + 0.02740943431854248, + 0.011687668971717358, + 0.016871144995093346, + 0.01296079158782959, + 0.0253494530916214, + 0.010041586123406887, + 0.013867528177797794, + 0.005586653482168913, + 0.0047958311624825, + 0.0034588270355015993, + 0.005068219732493162, + 0.0033890206832438707, + 0.0047284625470638275, + 0.002987943822517991, + 0.004785158671438694, + 0.018991459161043167, + 0.02025519125163555, + 0.017891930416226387, + 0.014288169331848621, + 0.013285910710692406, + 0.017712654545903206, + 0.009646600112318993, + 0.02840821072459221, + 0.0129042137414217, + 0.00745819415897131, + 0.012639903463423252, + 0.005532814655452967, + 0.005585788283497095, + 0.009473775513470173, + 0.0058535123243927956, + 0.018695101141929626, + 0.010149833746254444, + 0.00387350982055068, + 0.010163995437324047, + 0.00619462551549077, + 0.0023148253094404936, + 0.0037935846485197544, + 0.007197780068963766, + 0.004315076395869255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.2256530523300171 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036860644817352 + ], + "opposite_logprob_tokens": [ + -1.91036856174469 + ], + "logprob_reward": -0.16036860644817352, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036860644817352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1260584592819214, + 0.1871534138917923, + 0.09013132005929947, + 0.06957146525382996, + 0.043569888919591904, + 0.11240806430578232, + 0.03663690760731697, + 0.04921751096844673, + 0.16518306732177734, + 0.12006992846727371 + ], + "credit_map": [ + 0.01516862865537405, + 0.015733463689684868, + 0.013864035718142986, + 0.017040573060512543, + 0.014972301200032234, + 0.01569926179945469, + 0.017091307789087296, + 0.01648888736963272, + 0.018949933350086212, + 0.038571491837501526, + 0.01931271329522133, + 0.02247915416955948, + 0.02824792079627514, + 0.01840176060795784, + 0.020424924790859222, + 0.02076551504433155, + 0.011031133122742176, + 0.0188773013651371, + 0.009682240895926952, + 0.013786000199615955, + 0.010907833464443684, + 0.0075949388556182384, + 0.010073336772620678, + 0.008178532123565674, + 0.007821573875844479, + 0.014790358953177929, + 0.007984205149114132, + 0.014286428689956665, + 0.007261163555085659, + 0.0052287704311311245, + 0.006286263465881348, + 0.00591269601136446, + 0.004856890067458153, + 0.003552838461473584, + 0.0040191542357206345, + 0.010000322014093399, + 0.0058055175468325615, + 0.006619221530854702, + 0.003545583225786686, + 0.005170361138880253, + 0.014528491534292698, + 0.019049175083637238, + 0.006407622713595629, + 0.013017416931688786, + 0.01953013800084591, + 0.016414698213338852, + 0.009623867459595203, + 0.013836649246513844, + 0.0053059738129377365, + 0.003530170302838087, + 0.0035399834159761667, + 0.00535746430978179, + 0.0024610518012195826, + 0.0066944872960448265, + 0.003199238097295165, + 0.006548536941409111, + 0.006803299766033888, + 0.009889917448163033, + 0.005369292106479406, + 0.006724025588482618, + 0.0047426181845366955, + 0.004900535102933645, + 0.002981441328302026, + 0.007806385401636362, + 0.025183679535984993, + 0.015356934629380703, + 0.02583383210003376, + 0.012123738415539265, + 0.011030876077711582, + 0.020078247413039207, + 0.012057713232934475, + 0.04351804777979851, + 0.019676392897963524, + 0.01245816983282566, + 0.018657715991139412, + 0.01304925512522459, + 0.008650127798318863, + 0.010952681303024292, + 0.023192832246422768, + 0.013432757928967476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.16036860644817352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a medium-wide shot, rising steadily upward to reveal a terrifying, expressionistic landscape painted with chaotic, emotive brushstrokes. A raging river of molten lava and lustrous obsidian churns violently. The liquid gold radiates an intense internal heat, casting a flickering, infernal light, while sharp, cooling obsidian fragments swirl like black daggers in the powerful current. Navigating this vortex is a lone, male and female figure atop a simple raft composed of dark, rough logs lashed with thick ropes. The figure is clad in a heavy, shapeless black robe and wears a mask resembling a stylized frog, its features carved with stark, unsettling lines and painted a pale, ghostly white, starkly capturing the incandescent light from below. With intense concentration, the figure grasps a long, rough wooden stick and deliberately and forcefully plunges it into the swirling current, pushing aside jagged rocks. Their movements resemble a desperate ballet against the river's wrath. The entire scene is rendered in thick, fresco-like textures and distorted shapes; light from the molten gold creates dramatic, high-contrast black and whites, casting long, dancing shadows across the bruised purple and indigo cliffs that tilt oppressively inward. As the camera completes its ascent from a high-angle, bird's-eye view, the figure and raft shrink to a tiny, solitary speck within a vast, swirling vortex of light and dark, emphasizing an atmosphere of profound isolation and an epic, desperate struggle against overwhelming, primal power.", + "seed": 17720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_13/index_9319_seed_17720_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": -0.22368809022009373, + "vlm_logprob": -0.22368809022009373, + "vlm_yes_no_margin": 1.9999999940395354, + "vlm_token_credit": -0.22368809022009373, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08847866207361221 + ], + "opposite_logprob_tokens": [ + -2.5884785652160645 + ], + "logprob_reward": -0.08847866207361221, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08847866207361221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36161085963249207, + 0.07203428447246552, + 0.03841625899076462, + 0.033244773745536804, + 0.089534230530262, + 0.07575362920761108, + 0.06111139431595802, + 0.04765361547470093, + 0.03066817857325077, + 0.18997277319431305 + ], + "credit_map": [ + 0.023527873679995537, + 0.025153491646051407, + 0.03132109344005585, + 0.022489061579108238, + 0.0479930154979229, + 0.018850203603506088, + 0.07910491526126862, + 0.11317119747400284, + 0.0067778900265693665, + 0.008129692636430264, + 0.008974221535027027, + 0.010535335168242455, + 0.006540884263813496, + 0.0058823684230446815, + 0.0072771161794662476, + 0.01791677251458168, + 0.0016960749635472894, + 0.002698094118386507, + 0.0036717390175908804, + 0.012871887534856796, + 0.001996999839320779, + 0.0025681930128484964, + 0.0026243010070174932, + 0.010288969613611698, + 0.00403601722791791, + 0.005020165350288153, + 0.004753199405968189, + 0.005169853568077087, + 0.0036663010250777006, + 0.003889035200700164, + 0.002691046567633748, + 0.004019157029688358, + 0.00546813802793622, + 0.006699538789689541, + 0.02610628493130207, + 0.00669383630156517, + 0.008813957683742046, + 0.005887200124561787, + 0.005843590013682842, + 0.024021688848733902, + 0.007016563322395086, + 0.00980914942920208, + 0.008899176493287086, + 0.010767607018351555, + 0.0077225626446306705, + 0.008575414307415485, + 0.010877784341573715, + 0.01208537444472313, + 0.006778358947485685, + 0.008663383312523365, + 0.006698863580822945, + 0.007814192213118076, + 0.008942008018493652, + 0.007445329800248146, + 0.006840430665761232, + 0.007928825914859772, + 0.0062177651561796665, + 0.0045739044435322285, + 0.007249599322676659, + 0.006671138573437929, + 0.004361312370747328, + 0.006454748567193747, + 0.005650702863931656, + 0.006474442780017853, + 0.003738360945135355, + 0.0027739694342017174, + 0.003177895676344633, + 0.003922897391021252, + 0.003214362310245633, + 0.0037803957238793373, + 0.004528270568698645, + 0.005532026756554842, + 0.03948889300227165, + 0.010022826492786407, + 0.020355697721242905, + 0.020702440291643143, + 0.011775629594922066, + 0.013574688695371151, + 0.011711155995726585, + 0.06234145537018776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.08847866207361221 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637231908738613 + ], + "opposite_logprob_tokens": [ + -3.6513724327087402 + ], + "logprob_reward": -0.02637231908738613, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637231908738613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3422883152961731, + 0.09399544447660446, + 0.06635326892137527, + 0.0471058264374733, + 0.10733207315206528, + 0.06297148764133453, + 0.021004701033234596, + 0.04366501420736313, + 0.0620364211499691, + 0.15324744582176208 + ], + "credit_map": [ + 0.021411260589957237, + 0.02084130421280861, + 0.03174308314919472, + 0.02132849581539631, + 0.04503563046455383, + 0.018000811338424683, + 0.07353956252336502, + 0.11038815975189209, + 0.006480375304818153, + 0.010045891627669334, + 0.018449692055583, + 0.02064199186861515, + 0.010829349048435688, + 0.006545237731188536, + 0.012503116391599178, + 0.008499789983034134, + 0.0027440753765404224, + 0.004449944477528334, + 0.005158631131052971, + 0.022088145837187767, + 0.00226246053352952, + 0.003984034061431885, + 0.003421723609790206, + 0.022244254127144814, + 0.003573845373466611, + 0.006357706151902676, + 0.004781365394592285, + 0.006961936596781015, + 0.010469144210219383, + 0.004321963991969824, + 0.005140617024153471, + 0.005499250255525112, + 0.006138619966804981, + 0.007177972700446844, + 0.03493523970246315, + 0.005170711316168308, + 0.009776027873158455, + 0.006438748445361853, + 0.00710440706461668, + 0.030590349808335304, + 0.006538184359669685, + 0.005805145483464003, + 0.007879572920501232, + 0.009291424416005611, + 0.010831399820744991, + 0.005407288670539856, + 0.009311161004006863, + 0.007907305844128132, + 0.002102981321513653, + 0.0023744453210383654, + 0.0029426938854157925, + 0.002938029356300831, + 0.002807346871122718, + 0.001804184284992516, + 0.0033426114823669195, + 0.0026924079284071922, + 0.0063959755934774876, + 0.00402154540643096, + 0.005363987293094397, + 0.0049525112845003605, + 0.003961136564612389, + 0.008108913898468018, + 0.004605990368872881, + 0.0062549542635679245, + 0.008051334880292416, + 0.004309833515435457, + 0.0072159189730882645, + 0.007336973212659359, + 0.004204527474939823, + 0.008606990799307823, + 0.011235979385674, + 0.011074867099523544, + 0.032066840678453445, + 0.008097163401544094, + 0.016725536435842514, + 0.01643698662519455, + 0.009830825962126255, + 0.010880440473556519, + 0.009910255670547485, + 0.04929939657449722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.02637231908738613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549609549343586 + ], + "opposite_logprob_tokens": [ + -2.9299609661102295 + ], + "logprob_reward": -0.0549609549343586, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549609549343586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25950512290000916, + 0.11493229866027832, + 0.07014967501163483, + 0.030709875747561455, + 0.06898470222949982, + 0.08825092762708664, + 0.036394763737916946, + 0.07399660348892212, + 0.15333981812000275, + 0.10373620688915253 + ], + "credit_map": [ + 0.01759076677262783, + 0.017621677368879318, + 0.036865122616291046, + 0.01554623432457447, + 0.02209862694144249, + 0.014057465828955173, + 0.05353769287467003, + 0.08218756318092346, + 0.009317133575677872, + 0.011449506506323814, + 0.02505904622375965, + 0.028854532167315483, + 0.01341985259205103, + 0.006175598129630089, + 0.01051875576376915, + 0.010137873701751232, + 0.00539054349064827, + 0.005254717543721199, + 0.0056974394246935844, + 0.019706442952156067, + 0.006195726804435253, + 0.004034505691379309, + 0.004451775457710028, + 0.019418522715568542, + 0.0029323482885956764, + 0.0032685298938304186, + 0.003504438791424036, + 0.003907071892172098, + 0.006515880115330219, + 0.0028908683452755213, + 0.003428923198953271, + 0.00426181685179472, + 0.003531969618052244, + 0.004547135438770056, + 0.019938673824071884, + 0.003998450003564358, + 0.009330861270427704, + 0.004061412997543812, + 0.004740220960229635, + 0.01883598603308201, + 0.00917703378945589, + 0.01072424091398716, + 0.011243950575590134, + 0.010212070308625698, + 0.013376402668654919, + 0.008092544041574001, + 0.013663155026733875, + 0.011761540547013283, + 0.0038972976617515087, + 0.005824972875416279, + 0.004322200547903776, + 0.004028514493256807, + 0.005509665235877037, + 0.004361028783023357, + 0.003768129739910364, + 0.004682957660406828, + 0.010708341374993324, + 0.007052656263113022, + 0.00877924170345068, + 0.008826026692986488, + 0.0070455665700137615, + 0.012349733151495457, + 0.009682593867182732, + 0.00955244805663824, + 0.019582359120249748, + 0.010301695205271244, + 0.017031380906701088, + 0.017378268763422966, + 0.012253057211637497, + 0.022106532007455826, + 0.02856704406440258, + 0.02611948549747467, + 0.020246675238013268, + 0.006136366166174412, + 0.011448158882558346, + 0.010977203026413918, + 0.007837170735001564, + 0.007532522547990084, + 0.007749296259135008, + 0.031808823347091675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0549609549343586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743402302265167 + ], + "opposite_logprob_tokens": [ + -0.9743402004241943 + ], + "logprob_reward": -0.4743402302265167, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.4743402302265167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29919376969337463, + 0.07381404936313629, + 0.0949501320719719, + 0.021342583000659943, + 0.09435559809207916, + 0.07412806898355484, + 0.020628413185477257, + 0.04983920603990555, + 0.023573148995637894, + 0.2481750249862671 + ], + "credit_map": [ + 0.020015446469187737, + 0.019017565995454788, + 0.02523050643503666, + 0.01665416546165943, + 0.035169076174497604, + 0.014552286826074123, + 0.07102806866168976, + 0.0975266620516777, + 0.005020609125494957, + 0.007059399038553238, + 0.016975924372673035, + 0.01874757930636406, + 0.00838884711265564, + 0.004105724394321442, + 0.007255387492477894, + 0.00626057293266058, + 0.004229160491377115, + 0.004988634958863258, + 0.008119217120110989, + 0.03319162502884865, + 0.0045774648897349834, + 0.005715955980122089, + 0.004772374872118235, + 0.029355691745877266, + 0.0019565976690500975, + 0.0027319008950144053, + 0.0024162023328244686, + 0.002981242723762989, + 0.004443532787263393, + 0.0019990145228803158, + 0.0024872643407434225, + 0.0023268256336450577, + 0.007172444835305214, + 0.007918906398117542, + 0.02536143735051155, + 0.006922734901309013, + 0.007811162155121565, + 0.006430350244045258, + 0.006704075261950493, + 0.02603449486196041, + 0.008182534947991371, + 0.006766649428755045, + 0.009847253561019897, + 0.01013004407286644, + 0.009727317839860916, + 0.006893770769238472, + 0.012739372439682484, + 0.009841129183769226, + 0.00210791127756238, + 0.0026356251910328865, + 0.0023387435358017683, + 0.002216402208432555, + 0.004113451112061739, + 0.001766348839737475, + 0.0027258123736828566, + 0.002724118297919631, + 0.006232965737581253, + 0.005167997907847166, + 0.004732493311166763, + 0.004991601221263409, + 0.006832835264503956, + 0.00829924363642931, + 0.006839699111878872, + 0.006742370314896107, + 0.0031100616324692965, + 0.0017733746208250523, + 0.0025380917359143496, + 0.0038196900859475136, + 0.0016072280704975128, + 0.002037741942331195, + 0.002350123366340995, + 0.006336838938295841, + 0.05311046540737152, + 0.012703540734946728, + 0.026606550440192223, + 0.02708217315375805, + 0.014553571119904518, + 0.017177769914269447, + 0.015527390874922276, + 0.08141358196735382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.4743402302265167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47428828477859497 + ], + "opposite_logprob_tokens": [ + -0.974288284778595 + ], + "logprob_reward": -0.47428828477859497, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47428828477859497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25188180804252625, + 0.038610100746154785, + 0.06422844529151917, + 0.026136377826333046, + 0.0269019715487957, + 0.0511360839009285, + 0.026748813688755035, + 0.05305465683341026, + 0.047282930463552475, + 0.41401880979537964 + ], + "credit_map": [ + 0.017468897625803947, + 0.016500215977430344, + 0.022893231362104416, + 0.013758417218923569, + 0.03200017288327217, + 0.014295972883701324, + 0.0563204362988472, + 0.07864446938037872, + 0.0036943983286619186, + 0.00422917865216732, + 0.0071847569197416306, + 0.004854283761233091, + 0.004883999470621347, + 0.003598845098167658, + 0.004231350030750036, + 0.005933287087827921, + 0.003953002858906984, + 0.004391241353005171, + 0.00495687872171402, + 0.02096428908407688, + 0.003191322321072221, + 0.003781555686146021, + 0.0032615659292787313, + 0.01972859539091587, + 0.0023671097587794065, + 0.00334433326497674, + 0.004611840005964041, + 0.0033936849795281887, + 0.0031695771031081676, + 0.0024741070810705423, + 0.0023182679433375597, + 0.004457460716366768, + 0.0021143832709640265, + 0.002595245838165283, + 0.005707521457225084, + 0.0017812354490160942, + 0.0031580983195453882, + 0.0022024984937161207, + 0.0032036087941378355, + 0.00613937946036458, + 0.005560226272791624, + 0.0074073984287679195, + 0.007188757881522179, + 0.0050289868377149105, + 0.005422110203653574, + 0.007567356340587139, + 0.004604386631399393, + 0.008356869220733643, + 0.0029056419152766466, + 0.003095446852967143, + 0.0022095481399446726, + 0.0033104124013334513, + 0.005844664294272661, + 0.0023724278435111046, + 0.003311945591121912, + 0.0036987268831580877, + 0.00462464801967144, + 0.0053334468975663185, + 0.007918138988316059, + 0.007457962725311518, + 0.0055311438627541065, + 0.0039173271507024765, + 0.006912243086844683, + 0.01135974656790495, + 0.005831222049891949, + 0.00379333458840847, + 0.004547111690044403, + 0.005057520233094692, + 0.006471463479101658, + 0.004987149965018034, + 0.009385498240590096, + 0.0072096362709999084, + 0.09134829044342041, + 0.021253105252981186, + 0.04544241353869438, + 0.04659722000360489, + 0.020145641639828682, + 0.02767898701131344, + 0.02812017686665058, + 0.13343298435211182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.47428828477859497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a medium-wide shot, rising steadily upward to reveal a terrifying, expressionistic landscape painted with chaotic, emotive brushstrokes. A raging river of molten lava and lustrous obsidian churns violently. The liquid gold radiates an intense internal heat, casting a flickering, infernal light, while sharp, cooling obsidian fragments swirl like black daggers in the powerful current. Navigating this vortex is a lone, male and female figure atop a simple raft composed of dark, rough logs lashed with thick ropes. The figure is clad in a heavy, shapeless black robe and wears a mask resembling a stylized frog, its features carved with stark, unsettling lines and painted a pale, ghostly white, starkly capturing the incandescent light from below. With intense concentration, the figure grasps a long, rough wooden stick and deliberately and forcefully plunges it into the swirling current, pushing aside jagged rocks. Their movements resemble a desperate ballet against the river's wrath. The entire scene is rendered in thick, fresco-like textures and distorted shapes; light from the molten gold creates dramatic, high-contrast black and whites, casting long, dancing shadows across the bruised purple and indigo cliffs that tilt oppressively inward. As the camera completes its ascent from a high-angle, bird's-eye view, the figure and raft shrink to a tiny, solitary speck within a vast, swirling vortex of light and dark, emphasizing an atmosphere of profound isolation and an epic, desperate struggle against overwhelming, primal power.", + "seed": 17721, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_13/index_9319_seed_17721_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": -0.2075076036155224, + "vlm_logprob": -0.2075076036155224, + "vlm_yes_no_margin": 2.2249999940395355, + "vlm_token_credit": -0.2075076036155224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04946869984269142 + ], + "opposite_logprob_tokens": [ + -3.674468755722046 + ], + "logprob_reward": -0.04946869984269142, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04946869984269142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07609572261571884, + 0.028733771294355392, + 0.1081879660487175, + 0.09798229485750198, + 0.04163578897714615, + 0.11003883183002472, + 0.15207916498184204, + 0.16821499168872833, + 0.14381226897239685, + 0.07321915775537491 + ], + "credit_map": [ + 0.00790613703429699, + 0.012312495149672031, + 0.01355819497257471, + 0.006230093073099852, + 0.005255311727523804, + 0.008259380236268044, + 0.00782048050314188, + 0.014753634110093117, + 0.00275859865359962, + 0.0029865067917853594, + 0.007223439868539572, + 0.0022475947625935078, + 0.0016365142073482275, + 0.00408016424626112, + 0.005198312923312187, + 0.0026026414707303047, + 0.025223568081855774, + 0.0065463753417134285, + 0.03511916473507881, + 0.004825894255191088, + 0.004860922694206238, + 0.014873324893414974, + 0.010012013837695122, + 0.006726700812578201, + 0.01086204033344984, + 0.008004431612789631, + 0.005900298710912466, + 0.0039709825068712234, + 0.007125608157366514, + 0.035192836076021194, + 0.007438075728714466, + 0.01948801800608635, + 0.003968824632465839, + 0.004010021686553955, + 0.004721022676676512, + 0.003971944097429514, + 0.0035806677769869566, + 0.009007877670228481, + 0.004764938727021217, + 0.0076104928739368916, + 0.026523182168602943, + 0.006230725906789303, + 0.007877673953771591, + 0.0038401647470891476, + 0.005803191103041172, + 0.011462386697530746, + 0.008622799068689346, + 0.03967871516942978, + 0.034469928592443466, + 0.0077579449862241745, + 0.007299263495951891, + 0.004401304293423891, + 0.06369215250015259, + 0.02413603849709034, + 0.006058279424905777, + 0.004264265298843384, + 0.04604635015130043, + 0.008950524963438511, + 0.0074793314561247826, + 0.004840855486690998, + 0.017660899087786674, + 0.06778518110513687, + 0.007172964513301849, + 0.00827888771891594, + 0.01984463818371296, + 0.01281282864511013, + 0.006077079102396965, + 0.006645379588007927, + 0.04884517192840576, + 0.030797766521573067, + 0.007762623019516468, + 0.011026783846318722, + 0.014701003208756447, + 0.0083336615934968, + 0.008158822543919086, + 0.009146729484200478, + 0.009322360157966614, + 0.009850207716226578, + 0.005283497739583254, + 0.00842287763953209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.04946869984269142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983858808875084 + ], + "opposite_logprob_tokens": [ + -3.5298385620117188 + ], + "logprob_reward": -0.02983858808875084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983858808875084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18280330300331116, + 0.02404906414449215, + 0.06558267772197723, + 0.07997801899909973, + 0.06272053718566895, + 0.16911420226097107, + 0.1863565295934677, + 0.1106378436088562, + 0.0737311914563179, + 0.045026686042547226 + ], + "credit_map": [ + 0.01711920276284218, + 0.02891727164387703, + 0.0489339604973793, + 0.011173185892403126, + 0.010057379491627216, + 0.022865258157253265, + 0.015133759938180447, + 0.028603259474039078, + 0.004092006478458643, + 0.003368040081113577, + 0.004668379668146372, + 0.0016343186143785715, + 0.001760211423970759, + 0.0031658632215112448, + 0.002142235403880477, + 0.00321800890378654, + 0.008669717237353325, + 0.004620450083166361, + 0.0266727302223444, + 0.002853791695088148, + 0.002428961219266057, + 0.01511833630502224, + 0.0026929418090730906, + 0.0025257463566958904, + 0.01673741638660431, + 0.004896568134427071, + 0.005737487226724625, + 0.0024799129460006952, + 0.007477636449038982, + 0.03205332159996033, + 0.0043833255767822266, + 0.006212352775037289, + 0.005554734729230404, + 0.014031047001481056, + 0.004256023559719324, + 0.00844593532383442, + 0.006278183311223984, + 0.008914945647120476, + 0.004313431680202484, + 0.010926230810582638, + 0.04377542436122894, + 0.008331915363669395, + 0.009823231026530266, + 0.004403012339025736, + 0.008272053673863411, + 0.017578281462192535, + 0.012559790164232254, + 0.06437047570943832, + 0.052794504910707474, + 0.009869237430393696, + 0.009894299320876598, + 0.010019959881901741, + 0.04374975711107254, + 0.020571134984493256, + 0.023800823837518692, + 0.015656787902116776, + 0.02800191566348076, + 0.005549929104745388, + 0.004139910917729139, + 0.0038020398933440447, + 0.01380287017673254, + 0.04482031241059303, + 0.005016193725168705, + 0.005504664499312639, + 0.010214452631771564, + 0.006338056642562151, + 0.0035402807407081127, + 0.004369538277387619, + 0.025696031749248505, + 0.014669964089989662, + 0.003748072311282158, + 0.005154786631464958, + 0.008018351159989834, + 0.005692173261195421, + 0.00430200295522809, + 0.006776805501431227, + 0.006637562066316605, + 0.0066898721270263195, + 0.002927826251834631, + 0.0039820889942348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.02983858808875084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497032031416893 + ], + "opposite_logprob_tokens": [ + -2.9299702644348145 + ], + "logprob_reward": -0.05497032031416893, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497032031416893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07517269253730774, + 0.03114558756351471, + 0.27177101373672485, + 0.03211884945631027, + 0.02736387960612774, + 0.09402517974376678, + 0.20433875918388367, + 0.17581041157245636, + 0.055410291999578476, + 0.03284328430891037 + ], + "credit_map": [ + 0.008359876461327076, + 0.006853959988802671, + 0.019850244745612144, + 0.0059045953676104546, + 0.0044161793775856495, + 0.009913366287946701, + 0.0074323066510260105, + 0.012442170642316341, + 0.003007667837664485, + 0.003924303688108921, + 0.006894276477396488, + 0.0024073778185993433, + 0.001797906355932355, + 0.006668358109891415, + 0.0029999585822224617, + 0.0034457403235137463, + 0.06211690977215767, + 0.014511823654174805, + 0.08492188155651093, + 0.012360998429358006, + 0.012586632743477821, + 0.0387122817337513, + 0.02936585620045662, + 0.017194656655192375, + 0.007025680970400572, + 0.0034053416457027197, + 0.002890752861276269, + 0.0011742471251636744, + 0.0034128606785088778, + 0.00845472700893879, + 0.0017981986748054624, + 0.00395704060792923, + 0.003069294150918722, + 0.0030972259119153023, + 0.0026314370334148407, + 0.0024791774339973927, + 0.002624324057251215, + 0.005860303062945604, + 0.0030836365185678005, + 0.004518481902778149, + 0.025948183611035347, + 0.006308292038738728, + 0.005332400090992451, + 0.00315133691765368, + 0.004794774577021599, + 0.009292391128838062, + 0.006503613665699959, + 0.03269419074058533, + 0.05591550096869469, + 0.011586072854697704, + 0.011560749262571335, + 0.014395215548574924, + 0.020994627848267555, + 0.03605229780077934, + 0.033100686967372894, + 0.02073361538350582, + 0.03706853836774826, + 0.009139127098023891, + 0.006775814574211836, + 0.005751732271164656, + 0.025802701711654663, + 0.07208463549613953, + 0.00814425852149725, + 0.011043602600693703, + 0.007711554411798716, + 0.004751608707010746, + 0.003070771461352706, + 0.004523568786680698, + 0.018058182671666145, + 0.010169564746320248, + 0.003102201735600829, + 0.004022838547825813, + 0.0047796242870390415, + 0.005335325840860605, + 0.003467479720711708, + 0.0031167748384177685, + 0.005347684491425753, + 0.004956821911036968, + 0.0022061483468860388, + 0.0036334239412099123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.05497032031416893 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47436144948005676 + ], + "opposite_logprob_tokens": [ + -0.9743614196777344 + ], + "logprob_reward": -0.47436144948005676, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47436144948005676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07485997676849365, + 0.02513998933136463, + 0.3000372350215912, + 0.045561980456113815, + 0.025348316878080368, + 0.049638472497463226, + 0.11644463241100311, + 0.28092220425605774, + 0.06344256550073624, + 0.018604645505547523 + ], + "credit_map": [ + 0.006854963023215532, + 0.007847835309803486, + 0.019070187583565712, + 0.005215649027377367, + 0.003915601409971714, + 0.011271130293607712, + 0.006980699487030506, + 0.013703903183341026, + 0.0022777824196964502, + 0.002736307680606842, + 0.004940596409142017, + 0.002445399994030595, + 0.0016319085843861103, + 0.002813991392031312, + 0.005199742503464222, + 0.00309425825253129, + 0.059799112379550934, + 0.01611456833779812, + 0.11129502952098846, + 0.013057774864137173, + 0.012147803790867329, + 0.055008381605148315, + 0.019527973607182503, + 0.013086549006402493, + 0.006853817496448755, + 0.005020367447286844, + 0.003943040035665035, + 0.0022016018629074097, + 0.011504118330776691, + 0.004721995908766985, + 0.0034022394102066755, + 0.007914794608950615, + 0.0022187770809978247, + 0.00373100652359426, + 0.002093622926622629, + 0.0022573077585548162, + 0.0023433377500623465, + 0.004855400882661343, + 0.0027752812020480633, + 0.005073579028248787, + 0.009742987342178822, + 0.0036443746648728848, + 0.004269410856068134, + 0.00238282629288733, + 0.002985169645398855, + 0.005917684640735388, + 0.003474291879683733, + 0.0172217208892107, + 0.029300207272171974, + 0.007193949539214373, + 0.007935059256851673, + 0.009702586568892002, + 0.010495243594050407, + 0.019886638969182968, + 0.019329708069562912, + 0.012601231224834919, + 0.06686713546514511, + 0.013687198981642723, + 0.010301769711077213, + 0.00758393295109272, + 0.038959432393312454, + 0.11794955283403397, + 0.01191926933825016, + 0.013653895817697048, + 0.00893974769860506, + 0.005333005916327238, + 0.003486444242298603, + 0.005130598787218332, + 0.019925368949770927, + 0.011580796912312508, + 0.003671863814815879, + 0.005374729633331299, + 0.00304879667237401, + 0.00252779689617455, + 0.001664591720327735, + 0.002740965923294425, + 0.0028803797904402018, + 0.0028436582069844007, + 0.0012099791783839464, + 0.0016884756041690707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.47436144948005676 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42889896035194397 + ], + "opposite_logprob_tokens": [ + -1.0538989305496216 + ], + "logprob_reward": -0.42889896035194397, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42889896035194397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0622013695538044, + 0.0365852415561676, + 0.17131341993808746, + 0.03223644942045212, + 0.026987411081790924, + 0.05373521149158478, + 0.3027578592300415, + 0.2637171149253845, + 0.03838912025094032, + 0.012076851911842823 + ], + "credit_map": [ + 0.0059882402420043945, + 0.007484234403818846, + 0.015167766250669956, + 0.004421583376824856, + 0.0031345831230282784, + 0.01054609939455986, + 0.006006448529660702, + 0.009452415630221367, + 0.004639194812625647, + 0.006056384649127722, + 0.008215551264584064, + 0.0026688026264309883, + 0.002109191380441189, + 0.005372083280235529, + 0.003106413409113884, + 0.00441761827096343, + 0.0360235869884491, + 0.010796544142067432, + 0.05858348682522774, + 0.007608941290527582, + 0.0072852433659136295, + 0.028194332495331764, + 0.014525613747537136, + 0.00829567201435566, + 0.0047401487827301025, + 0.0035062755923718214, + 0.0024135252460837364, + 0.0016039202455431223, + 0.007171668112277985, + 0.006475346628576517, + 0.0023686958011239767, + 0.0039568678475916386, + 0.002636580029502511, + 0.00373158510774374, + 0.003343201009556651, + 0.0023846642579883337, + 0.002307229908183217, + 0.004612787626683712, + 0.0024902552831918, + 0.005481108091771603, + 0.010518636554479599, + 0.0035899807699024677, + 0.0032733974512666464, + 0.0018380520632490516, + 0.003163546323776245, + 0.0072024548426270485, + 0.0040893759578466415, + 0.02005976438522339, + 0.033424295485019684, + 0.012931263074278831, + 0.011435527354478836, + 0.009297991171479225, + 0.11211910098791122, + 0.06800754368305206, + 0.037502896040678024, + 0.01803923025727272, + 0.05927106738090515, + 0.013648363761603832, + 0.010131402872502804, + 0.00974611472338438, + 0.035772208124399185, + 0.10655903816223145, + 0.012228667736053467, + 0.016360245645046234, + 0.004954908508807421, + 0.006735695991665125, + 0.003254548879340291, + 0.004549826495349407, + 0.006745683960616589, + 0.00512715382501483, + 0.002807913115248084, + 0.004213393200188875, + 0.0014630501391366124, + 0.0015448845224455, + 0.0010704483138397336, + 0.0012414847733452916, + 0.002013830468058586, + 0.0028759618289768696, + 0.000856287544593215, + 0.0010109046706929803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.42889896035194397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a medium-wide shot, rising steadily upward to reveal a terrifying, expressionistic landscape painted with chaotic, emotive brushstrokes. A raging river of molten lava and lustrous obsidian churns violently. The liquid gold radiates an intense internal heat, casting a flickering, infernal light, while sharp, cooling obsidian fragments swirl like black daggers in the powerful current. Navigating this vortex is a lone, male and female figure atop a simple raft composed of dark, rough logs lashed with thick ropes. The figure is clad in a heavy, shapeless black robe and wears a mask resembling a stylized frog, its features carved with stark, unsettling lines and painted a pale, ghostly white, starkly capturing the incandescent light from below. With intense concentration, the figure grasps a long, rough wooden stick and deliberately and forcefully plunges it into the swirling current, pushing aside jagged rocks. Their movements resemble a desperate ballet against the river's wrath. The entire scene is rendered in thick, fresco-like textures and distorted shapes; light from the molten gold creates dramatic, high-contrast black and whites, casting long, dancing shadows across the bruised purple and indigo cliffs that tilt oppressively inward. As the camera completes its ascent from a high-angle, bird's-eye view, the figure and raft shrink to a tiny, solitary speck within a vast, swirling vortex of light and dark, emphasizing an atmosphere of profound isolation and an epic, desperate struggle against overwhelming, primal power.", + "seed": 17722, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_13/index_9319_seed_17722_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": -0.07978042736649513, + "vlm_logprob": -0.07978042736649513, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.07978042736649513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038657598197460175 + ], + "opposite_logprob_tokens": [ + -4.0386576652526855 + ], + "logprob_reward": -0.038657598197460175, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.038657598197460175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05040755495429039, + 0.14308977127075195, + 0.06113193556666374, + 0.05491660162806511, + 0.31533312797546387, + 0.06784277409315109, + 0.07248082756996155, + 0.10878322273492813, + 0.050374679267406464, + 0.07563949376344681 + ], + "credit_map": [ + 0.004319994710385799, + 0.00880956556648016, + 0.00772503949701786, + 0.006953406613320112, + 0.005728841293603182, + 0.0075093344785273075, + 0.005464616231620312, + 0.003896756563335657, + 0.015753181651234627, + 0.023062966763973236, + 0.015208106487989426, + 0.009362712502479553, + 0.011827796697616577, + 0.008067777380347252, + 0.05188930407166481, + 0.007917921058833599, + 0.0068326061591506, + 0.00794700626283884, + 0.006503459997475147, + 0.011907003819942474, + 0.005540093407034874, + 0.00972459465265274, + 0.008928295224905014, + 0.0037488786038011312, + 0.006504733115434647, + 0.007555786054581404, + 0.007132630329579115, + 0.009232069365680218, + 0.00778315169736743, + 0.005113121587783098, + 0.006833882536739111, + 0.004761223215609789, + 0.04075450450181961, + 0.030135823413729668, + 0.04161074757575989, + 0.02263340726494789, + 0.03605948016047478, + 0.05007154121994972, + 0.032273877412080765, + 0.06179376691579819, + 0.008911577053368092, + 0.007789770606905222, + 0.009619404561817646, + 0.010454738512635231, + 0.00884191133081913, + 0.006773421540856361, + 0.008571694605052471, + 0.006880251225084066, + 0.01206789817661047, + 0.010388934053480625, + 0.012031148187816143, + 0.008799579925835133, + 0.0071739740669727325, + 0.007325769867748022, + 0.007195531856268644, + 0.00749798770993948, + 0.01207795925438404, + 0.013149382546544075, + 0.00862063467502594, + 0.009396923705935478, + 0.0286477729678154, + 0.01963690109550953, + 0.007554356474429369, + 0.009699294343590736, + 0.010778826661407948, + 0.005771266762167215, + 0.007297886069864035, + 0.007960384711623192, + 0.006665883585810661, + 0.00509687140583992, + 0.003301441902294755, + 0.0035021172370761633, + 0.00897238589823246, + 0.007695885840803385, + 0.01166799757629633, + 0.009628248400986195, + 0.009237214922904968, + 0.009264969266951084, + 0.00984951015561819, + 0.009323284961283207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.038657598197460175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033724378794431686 + ], + "opposite_logprob_tokens": [ + -3.408724308013916 + ], + "logprob_reward": -0.033724378794431686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033724378794431686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18880674242973328, + 0.28015008568763733, + 0.0566118024289608, + 0.10834009200334549, + 0.10844458639621735, + 0.048508115112781525, + 0.04824325069785118, + 0.09091440588235855, + 0.030078241601586342, + 0.03990267589688301 + ], + "credit_map": [ + 0.015847405418753624, + 0.02047637850046158, + 0.029402967542409897, + 0.025521455332636833, + 0.02286253497004509, + 0.022931531071662903, + 0.033430151641368866, + 0.018334314227104187, + 0.03052976354956627, + 0.04528262838721275, + 0.03236306086182594, + 0.021341754123568535, + 0.02150459960103035, + 0.01549080852419138, + 0.09940017014741898, + 0.014237327501177788, + 0.0061920140869915485, + 0.006662013009190559, + 0.006441083736717701, + 0.010148321278393269, + 0.007023938931524754, + 0.00944355595856905, + 0.005842192098498344, + 0.004858681932091713, + 0.012701408006250858, + 0.017960412427783012, + 0.010975060053169727, + 0.022241292521357536, + 0.007986029610037804, + 0.007573251146823168, + 0.01864437758922577, + 0.01025826670229435, + 0.015736375004053116, + 0.011812925338745117, + 0.014514671638607979, + 0.009953421540558338, + 0.010150598362088203, + 0.012421202845871449, + 0.01439578551799059, + 0.01945960335433483, + 0.0061843516305089, + 0.006696037016808987, + 0.005837378092110157, + 0.006352471653372049, + 0.005475978832691908, + 0.004893506411463022, + 0.006092472933232784, + 0.006975921336561441, + 0.008816261775791645, + 0.005968616344034672, + 0.005653128027915955, + 0.0057283369824290276, + 0.004837433807551861, + 0.005661935545504093, + 0.005859329830855131, + 0.005718212109059095, + 0.00848248228430748, + 0.011089506559073925, + 0.006594195496290922, + 0.007337918970733881, + 0.025933805853128433, + 0.016558831557631493, + 0.006973022595047951, + 0.007944643497467041, + 0.0034242598339915276, + 0.0030327518470585346, + 0.0036136191338300705, + 0.00468446733430028, + 0.006611364893615246, + 0.0037555790040642023, + 0.002581901615485549, + 0.0023742944467812777, + 0.0041714562103152275, + 0.005652308464050293, + 0.00521519361063838, + 0.004592594690620899, + 0.005005203653126955, + 0.0048501156270504, + 0.004422453697770834, + 0.005993349477648735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.033724378794431686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370167315006256 + ], + "opposite_logprob_tokens": [ + -3.4087016582489014 + ], + "logprob_reward": -0.03370167315006256, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370167315006256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20391438901424408, + 0.10216794162988663, + 0.07721465080976486, + 0.0967048779129982, + 0.11227002739906311, + 0.057147879153490067, + 0.042467471212148666, + 0.12911413609981537, + 0.10351943969726562, + 0.07547923177480698 + ], + "credit_map": [ + 0.016133099794387817, + 0.024677513167262077, + 0.03121289052069187, + 0.029095973819494247, + 0.023407815024256706, + 0.02244936302304268, + 0.04087088629603386, + 0.016066832467913628, + 0.013462157920002937, + 0.01575998216867447, + 0.00818910263478756, + 0.00905752182006836, + 0.008376037701964378, + 0.006927949376404285, + 0.033426471054553986, + 0.0069687142968177795, + 0.008559406735002995, + 0.010690193623304367, + 0.009532765485346317, + 0.019349610432982445, + 0.0070134978741407394, + 0.007486270274966955, + 0.00851688627153635, + 0.006066016387194395, + 0.010154316201806068, + 0.012183287180960178, + 0.012289855629205704, + 0.02046506479382515, + 0.007574102375656366, + 0.0071427649818360806, + 0.019163386896252632, + 0.007732104510068893, + 0.017952512949705124, + 0.012701451778411865, + 0.015542039647698402, + 0.00865682028234005, + 0.009581314399838448, + 0.014383175410330296, + 0.014538574032485485, + 0.018914127722382545, + 0.009241189807653427, + 0.004864147864282131, + 0.008064255118370056, + 0.011211007833480835, + 0.004114954732358456, + 0.006509535945951939, + 0.007350532338023186, + 0.005792249925434589, + 0.006966203451156616, + 0.00546345254406333, + 0.0047246855683624744, + 0.007640128489583731, + 0.0037737528327852488, + 0.0045170108787715435, + 0.0047264560125768185, + 0.004655780736356974, + 0.013273978605866432, + 0.013218563050031662, + 0.01296579372137785, + 0.011620541103184223, + 0.03337966650724411, + 0.023020053282380104, + 0.009429039433598518, + 0.012206503190100193, + 0.017087042331695557, + 0.011090617626905441, + 0.013565491884946823, + 0.01801932603120804, + 0.01603146828711033, + 0.011393287219107151, + 0.008540605194866657, + 0.007791605312377214, + 0.007116209249943495, + 0.010037519037723541, + 0.01219706330448389, + 0.008891713805496693, + 0.008578750304877758, + 0.00989654753357172, + 0.0077354805544018745, + 0.011025951243937016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.03370167315006256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798989176750183 + ], + "opposite_logprob_tokens": [ + -1.804898977279663 + ], + "logprob_reward": -0.1798989176750183, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798989176750183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09911038726568222, + 0.11016027629375458, + 0.08712796121835709, + 0.10839668661355972, + 0.08437502384185791, + 0.1743238866329193, + 0.09500119090080261, + 0.06883693486452103, + 0.13476257026195526, + 0.037905074656009674 + ], + "credit_map": [ + 0.00972018577158451, + 0.014366159215569496, + 0.012953512370586395, + 0.013747267425060272, + 0.011926756240427494, + 0.014341401867568493, + 0.012370227836072445, + 0.009684877470135689, + 0.009543491527438164, + 0.01851838454604149, + 0.014592812396585941, + 0.012132507748901844, + 0.009030867367982864, + 0.007935876026749611, + 0.03271499276161194, + 0.005691340658813715, + 0.009361373260617256, + 0.010699382051825523, + 0.010886183008551598, + 0.021833280101418495, + 0.00934652704745531, + 0.010299085639417171, + 0.00927338469773531, + 0.005428746808320284, + 0.011942829936742783, + 0.017462288960814476, + 0.013755048625171185, + 0.020114202052354813, + 0.008261914364993572, + 0.00860083568841219, + 0.016471700742840767, + 0.011787869036197662, + 0.011683986522257328, + 0.009995948523283005, + 0.0068574389442801476, + 0.013107608072459698, + 0.00872661080211401, + 0.011074699461460114, + 0.010490487329661846, + 0.012438240461051464, + 0.023369336500763893, + 0.022799525409936905, + 0.022913141176104546, + 0.024163417518138885, + 0.023171376436948776, + 0.018255110830068588, + 0.023171627894043922, + 0.01648036390542984, + 0.008862976916134357, + 0.01645924523472786, + 0.015051152557134628, + 0.014307733625173569, + 0.01148430909961462, + 0.010212449356913567, + 0.007016908377408981, + 0.01160641759634018, + 0.008566509000957012, + 0.008907436393201351, + 0.006040181964635849, + 0.008272122591733932, + 0.015317168086767197, + 0.010575633496046066, + 0.0055952491238713264, + 0.005562635138630867, + 0.018306627869606018, + 0.015280578285455704, + 0.017854321748018265, + 0.02636464685201645, + 0.015364875085651875, + 0.015408791601657867, + 0.012129341252148151, + 0.014053392224013805, + 0.003890692489221692, + 0.005251047667115927, + 0.005006459075957537, + 0.004760673735290766, + 0.004329179413616657, + 0.004446048755198717, + 0.004625727888196707, + 0.005595244932919741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.1798989176750183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 landscape featuring a raging river of molten lava and lustrous obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures (one male and one female) visible navigating the river on a raft made of dark logs lashed with thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one of the figures wearing a heavy black robe and a mask resembling a stylized frog with pale, ghostly white features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure deliberately plunge a long, rough wooden stick into the swirling current to push aside jagged rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium-wide shot and rise steadily upward to a high-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291956901550293 + ], + "opposite_logprob_tokens": [ + -2.237919569015503 + ], + "logprob_reward": -0.11291956901550293, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291956901550293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06906505674123764, + 0.10718341916799545, + 0.12210302799940109, + 0.0821787491440773, + 0.2764829993247986, + 0.15670177340507507, + 0.06701485812664032, + 0.056276436895132065, + 0.03991037234663963, + 0.02308330126106739 + ], + "credit_map": [ + 0.008233736269176006, + 0.010372764430940151, + 0.009145732969045639, + 0.008433681912720203, + 0.008445549756288528, + 0.00951503124088049, + 0.006698641926050186, + 0.008219911716878414, + 0.016837911680340767, + 0.015025238506495953, + 0.009593481197953224, + 0.011148335412144661, + 0.008394903503358364, + 0.01135999709367752, + 0.028026381507515907, + 0.006797163747251034, + 0.013881295919418335, + 0.014511523768305779, + 0.013708950951695442, + 0.02480201981961727, + 0.010782030411064625, + 0.020291533321142197, + 0.015636403113603592, + 0.008489266969263554, + 0.008876672945916653, + 0.011808080598711967, + 0.007796681486070156, + 0.018665237352252007, + 0.008567136712372303, + 0.006019044667482376, + 0.01177857257425785, + 0.008667314425110817, + 0.05209995061159134, + 0.023763416334986687, + 0.026991773396730423, + 0.03651471808552742, + 0.019283680245280266, + 0.060596954077482224, + 0.028870170935988426, + 0.02836233377456665, + 0.017970195040106773, + 0.02591177448630333, + 0.01735847443342209, + 0.018966050818562508, + 0.02542020007967949, + 0.016340747475624084, + 0.018857821822166443, + 0.015876509249210358, + 0.007397927809506655, + 0.011721314862370491, + 0.006629421375691891, + 0.009416839107871056, + 0.009062577039003372, + 0.007098156958818436, + 0.00705923093482852, + 0.008629383519291878, + 0.005987842567265034, + 0.006629615090787411, + 0.006004351191222668, + 0.008076167665421963, + 0.010262350551784039, + 0.008231418207287788, + 0.006209957879036665, + 0.004874724894762039, + 0.005384081043303013, + 0.004286686424165964, + 0.005145648960024118, + 0.00737745501101017, + 0.005130501464009285, + 0.004851589910686016, + 0.004170667380094528, + 0.0035637374967336655, + 0.004094780422747135, + 0.003108535660430789, + 0.005800032988190651, + 0.00269822939299047, + 0.0018499047728255391, + 0.002053246134892106, + 0.0016579465009272099, + 0.0018206234090030193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.11291956901550293 + } + ], + "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/0000025/group_14/index_tensor([2290, 2290])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_14/index_tensor([2290, 2290])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e87b99bfc9459d32a18f5271639504d4e281753e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_14/index_tensor([2290, 2290])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, set against the backdrop of the water level, captures a stern-faced hunter, his face gaunt and weathered, his long black hair plastered to his scalp, slowly and deliberately immersing himself in the dark, murky water at the heart of a dusky swamp. They are clad in worn, dark brown leather, the material cracked and waterlogged, its tarnished burrs gleaming dimly. The scene is rendered in a somber, painterly style, characterized by dramatic black-and-white contrasts and rich, painterly textures that give the gnarled, moss-covered roots in the background an almost tangible quality. As the hunters sink unflinchingly, water pours over their clenched jaws and thin lips, their posture deliberately stiff. A faint ray of light filters through the dense, oppressive canopy of trees above, capturing the steely determination in their pale, wide-open eyes, a last, sharp glimmer of defiance. Just before the dark, oily water completely obscures their faces, the light fades, leaving only faint ripples spreading across the black void.", + "seed": 278, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_14/index_2290_seed_278_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": -0.020581630524247885, + "vlm_logprob": -0.020581630524247885, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.020581630524247885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07052230089902878 + ], + "opposite_logprob_tokens": [ + -2.8205223083496094 + ], + "logprob_reward": -0.07052230089902878, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07052230089902878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057964425534009933, + 0.04771803691983223, + 0.16290108859539032, + 0.22620797157287598, + 0.24228493869304657, + 0.02696194127202034, + 0.06146563962101936, + 0.07489671558141708, + 0.04524845629930496, + 0.05435076728463173 + ], + "credit_map": [ + 0.008801649324595928, + 0.0027783147525042295, + 0.008397545665502548, + 0.007734831888228655, + 0.004183347802609205, + 0.006262802053242922, + 0.011076075956225395, + 0.008729859255254269, + 0.004073134623467922, + 0.005665051285177469, + 0.0051521179266273975, + 0.00608482863754034, + 0.007863531820476055, + 0.005448620766401291, + 0.006711598485708237, + 0.006719152443110943, + 0.009068969637155533, + 0.005696678534150124, + 0.04309150576591492, + 0.012815013527870178, + 0.009012199006974697, + 0.016009900718927383, + 0.015059000812470913, + 0.05214782431721687, + 0.02151128277182579, + 0.04438111558556557, + 0.022468846291303635, + 0.022679967805743217, + 0.03734850510954857, + 0.019228991121053696, + 0.029788775369524956, + 0.02880048379302025, + 0.024310356006026268, + 0.05171023681759834, + 0.023813560605049133, + 0.032842691987752914, + 0.03861761465668678, + 0.018357587978243828, + 0.029745159670710564, + 0.02288772165775299, + 0.0021808233577758074, + 0.0017900458769872785, + 0.004108456429094076, + 0.004184256307780743, + 0.002626943401992321, + 0.003031115746125579, + 0.0035737233702093363, + 0.005466577596962452, + 0.00681475130841136, + 0.0042188530787825584, + 0.008253602311015129, + 0.008125122636556625, + 0.005379388574510813, + 0.010935917496681213, + 0.006758230738341808, + 0.010979774408042431, + 0.007288938853889704, + 0.009397733025252819, + 0.009172601625323296, + 0.014629236422479153, + 0.008997495286166668, + 0.00783973466604948, + 0.010276633314788342, + 0.007294346112757921, + 0.004133508540689945, + 0.003921322990208864, + 0.005720593500882387, + 0.007418443448841572, + 0.004195923916995525, + 0.007216356229037046, + 0.006259459536522627, + 0.0063828458078205585, + 0.006179909221827984, + 0.004378279205411673, + 0.006406876724213362, + 0.00922208372503519, + 0.008445097133517265, + 0.008303425274789333, + 0.006240172311663628, + 0.005174923222512007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.07052230089902878 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208228327333927 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208228327333927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208228327333927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0953081026673317, + 0.05074796825647354, + 0.41969919204711914, + 0.11469411849975586, + 0.09905020892620087, + 0.03620066121220589, + 0.02967670001089573, + 0.02508745715022087, + 0.07700130343437195, + 0.052534282207489014 + ], + "credit_map": [ + 0.011230500414967537, + 0.005714216269552708, + 0.012032019905745983, + 0.0081381406635046, + 0.008388501591980457, + 0.014186338521540165, + 0.01561970915645361, + 0.01999867893755436, + 0.0071695479564368725, + 0.0043685464188456535, + 0.0068367719650268555, + 0.005959934554994106, + 0.005038073752075434, + 0.008776298724114895, + 0.006878625135868788, + 0.0057201748713850975, + 0.015151124447584152, + 0.008275774307549, + 0.15548020601272583, + 0.018212981522083282, + 0.007674946915358305, + 0.019290395081043243, + 0.014587808400392532, + 0.18102595210075378, + 0.010005544871091843, + 0.020000290125608444, + 0.01122843287885189, + 0.01637325808405876, + 0.018035385757684708, + 0.00812947005033493, + 0.019765228033065796, + 0.011156512424349785, + 0.011272131465375423, + 0.016246475279331207, + 0.014021952636539936, + 0.012049452401697636, + 0.012571107596158981, + 0.008673888631165028, + 0.010364928282797337, + 0.013850271701812744, + 0.0032956728246062994, + 0.002188289538025856, + 0.0052781254053115845, + 0.005477249622344971, + 0.003272009314969182, + 0.004746128339320421, + 0.004355079028755426, + 0.007588104344904423, + 0.002619368489831686, + 0.0026516378857195377, + 0.003908631857484579, + 0.004332962445914745, + 0.00412286352366209, + 0.0038795373402535915, + 0.004562513902783394, + 0.003599184099584818, + 0.002527696080505848, + 0.002561830449849367, + 0.0032233537640422583, + 0.004490863531827927, + 0.002438247436657548, + 0.0038782458286732435, + 0.0033221456687897444, + 0.002645072527229786, + 0.0076535833068192005, + 0.00621426198631525, + 0.00928865373134613, + 0.012653213925659657, + 0.005751878023147583, + 0.016695905476808548, + 0.009996390901505947, + 0.008747408166527748, + 0.00622668256983161, + 0.004564070608466864, + 0.007706631440669298, + 0.00684363255277276, + 0.0040152003057301044, + 0.0117581682279706, + 0.005655993241816759, + 0.005763904191553593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.014208228327333927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11173215508460999, + 0.03651827573776245, + 0.4417974352836609, + 0.028091611340641975, + 0.05235802382230759, + 0.03352329879999161, + 0.029051441699266434, + 0.03998579457402229, + 0.13786013424396515, + 0.08908182382583618 + ], + "credit_map": [ + 0.013506910763680935, + 0.0058470130898058414, + 0.015265305526554585, + 0.009337024763226509, + 0.010886083357036114, + 0.014745849184691906, + 0.016678951680660248, + 0.025465020909905434, + 0.003619159571826458, + 0.0016129547730088234, + 0.004437915049493313, + 0.004289313685148954, + 0.004205174278467894, + 0.00502985343337059, + 0.006332627031952143, + 0.006991282105445862, + 0.016951724886894226, + 0.008494225330650806, + 0.1589982807636261, + 0.019970711320638657, + 0.009866993874311447, + 0.021956823766231537, + 0.019287975504994392, + 0.18627071380615234, + 0.002618340076878667, + 0.002116731135174632, + 0.004151257686316967, + 0.004459644202142954, + 0.0024867854081094265, + 0.0036780124064534903, + 0.004162208177149296, + 0.004418636672198772, + 0.004592909011989832, + 0.003276377683505416, + 0.007612717337906361, + 0.007825116626918316, + 0.006312310229986906, + 0.005823005456477404, + 0.010584741830825806, + 0.006330849137157202, + 0.0031261437106877565, + 0.0021305440459400415, + 0.004208747297525406, + 0.0033852807246148586, + 0.004137840587645769, + 0.00504778977483511, + 0.00676093902438879, + 0.0047260164283216, + 0.003106278833001852, + 0.0020151089411228895, + 0.0038685137405991554, + 0.00416587246581912, + 0.0029316102154552937, + 0.004329187795519829, + 0.004522643052041531, + 0.004112228285521269, + 0.003805851563811302, + 0.003909410443156958, + 0.005245558451861143, + 0.008181511424481869, + 0.0038727072533220053, + 0.004893099423497915, + 0.0057318066246807575, + 0.004345854744315147, + 0.01365100871771574, + 0.009497120976448059, + 0.017139960080385208, + 0.01950780861079693, + 0.010465096682310104, + 0.031432200223207474, + 0.017767271026968956, + 0.018399685621261597, + 0.010559510439634323, + 0.008930911310017109, + 0.012391641736030579, + 0.011202733963727951, + 0.011825047433376312, + 0.01578684151172638, + 0.009734926745295525, + 0.008650217205286026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.0067326370626688 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00862894020974636 + ], + "opposite_logprob_tokens": [ + -4.758628845214844 + ], + "logprob_reward": -0.00862894020974636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00862894020974636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12040538340806961, + 0.01518929097801447, + 0.40591877698898315, + 0.020227260887622833, + 0.050249893218278885, + 0.013080863282084465, + 0.04191488027572632, + 0.05925919860601425, + 0.121962770819664, + 0.15179170668125153 + ], + "credit_map": [ + 0.02277747541666031, + 0.006678157951682806, + 0.02613581158220768, + 0.010394351556897163, + 0.006757476832717657, + 0.016800526529550552, + 0.011460940353572369, + 0.01940063200891018, + 0.002168327569961548, + 0.001387020805850625, + 0.0019786201883107424, + 0.0020788894034922123, + 0.0013151114108040929, + 0.002413171110674739, + 0.0016009350074455142, + 0.0022472150158137083, + 0.01598489284515381, + 0.00752783240750432, + 0.14710402488708496, + 0.01713998056948185, + 0.008019519969820976, + 0.024011604487895966, + 0.01451790053397417, + 0.17161302268505096, + 0.0019355203257873654, + 0.0020248694345355034, + 0.0024276268668472767, + 0.0027410495094954967, + 0.0021753411274403334, + 0.0027752018067985773, + 0.0037552700378000736, + 0.002392382360994816, + 0.0031594387255609035, + 0.003461337648332119, + 0.014385251328349113, + 0.003735510865226388, + 0.0030521689914166927, + 0.0038374990690499544, + 0.00296687800437212, + 0.015651807188987732, + 0.001246791915036738, + 0.0011058866512030363, + 0.0017877089558169246, + 0.0012101988540962338, + 0.001244746265001595, + 0.0021364621352404356, + 0.001814352348446846, + 0.002534716622903943, + 0.00442076800391078, + 0.0028984250966459513, + 0.007100739050656557, + 0.004781614523380995, + 0.00364768598228693, + 0.0072462065145373344, + 0.004937310237437487, + 0.006882131099700928, + 0.005718544125556946, + 0.0058471448719501495, + 0.008331635035574436, + 0.012381907552480698, + 0.007813393138349056, + 0.005611906759440899, + 0.007459437940269709, + 0.006095230113714933, + 0.012942367233335972, + 0.007866413332521915, + 0.016109002754092216, + 0.015008434653282166, + 0.0073330821469426155, + 0.03337119147181511, + 0.01297062635421753, + 0.016361650079488754, + 0.018210947513580322, + 0.014174616895616055, + 0.021218765527009964, + 0.02051459066569805, + 0.015627378597855568, + 0.030696365982294083, + 0.01665106602013111, + 0.014697974547743797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.00862894020974636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028160461224615574 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.0028160461224615574, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028160461224615574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19730418920516968, + 0.06353115290403366, + 0.17001239955425262, + 0.10845144838094711, + 0.07489143311977386, + 0.026846200227737427, + 0.06812937557697296, + 0.0846877321600914, + 0.1144874095916748, + 0.09165865182876587 + ], + "credit_map": [ + 0.025418398901820183, + 0.011003021150827408, + 0.031102262437343597, + 0.019929591566324234, + 0.01808117888867855, + 0.02645043097436428, + 0.04060499742627144, + 0.024714311584830284, + 0.005543839186429977, + 0.003309698309749365, + 0.007702247239649296, + 0.006811594124883413, + 0.009929872117936611, + 0.007776596583425999, + 0.011470078490674496, + 0.010987229645252228, + 0.006150578614324331, + 0.004497783724218607, + 0.059138812124729156, + 0.009011475369334221, + 0.0035581178963184357, + 0.009284159168601036, + 0.0097013795748353, + 0.06867008656263351, + 0.010497434064745903, + 0.013193896040320396, + 0.013392559252679348, + 0.014760714955627918, + 0.011003422550857067, + 0.011217102408409119, + 0.017993859946727753, + 0.016392454504966736, + 0.008806460537016392, + 0.012342873960733414, + 0.007760193664580584, + 0.009727170690894127, + 0.008389572612941265, + 0.009254904463887215, + 0.008750139735639095, + 0.009860111400485039, + 0.0022143491078168154, + 0.0029364400543272495, + 0.004167766775935888, + 0.0032914059702306986, + 0.0035092306789010763, + 0.003456939710304141, + 0.0036002236884087324, + 0.0036698428448289633, + 0.008116128854453564, + 0.005276597570627928, + 0.009012705646455288, + 0.009734226390719414, + 0.007409184705466032, + 0.010972512885928154, + 0.008075878024101257, + 0.0095321424305439, + 0.0076612248085439205, + 0.009023692458868027, + 0.011521538719534874, + 0.017153771594166756, + 0.010632362216711044, + 0.011895060539245605, + 0.008713597431778908, + 0.008086483925580978, + 0.01074109598994255, + 0.007820291444659233, + 0.016906702890992165, + 0.017820928245782852, + 0.008004934526979923, + 0.02721291035413742, + 0.014753881841897964, + 0.011226660571992397, + 0.010268471203744411, + 0.010489702224731445, + 0.012825735844671726, + 0.011138980276882648, + 0.014362214133143425, + 0.014201176352798939, + 0.009306893683969975, + 0.009065477177500725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0028160461224615574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, set against the backdrop of the water level, captures a stern-faced hunter, his face gaunt and weathered, his long black hair plastered to his scalp, slowly and deliberately immersing himself in the dark, murky water at the heart of a dusky swamp. They are clad in worn, dark brown leather, the material cracked and waterlogged, its tarnished burrs gleaming dimly. The scene is rendered in a somber, painterly style, characterized by dramatic black-and-white contrasts and rich, painterly textures that give the gnarled, moss-covered roots in the background an almost tangible quality. As the hunters sink unflinchingly, water pours over their clenched jaws and thin lips, their posture deliberately stiff. A faint ray of light filters through the dense, oppressive canopy of trees above, capturing the steely determination in their pale, wide-open eyes, a last, sharp glimmer of defiance. Just before the dark, oily water completely obscures their faces, the light fades, leaving only faint ripples spreading across the black void.", + "seed": 279, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_14/index_2290_seed_279_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": -0.012559495447203518, + "vlm_logprob": -0.012559495447203518, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.012559495447203518, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04480377212166786 + ], + "opposite_logprob_tokens": [ + -3.2948038578033447 + ], + "logprob_reward": -0.04480377212166786, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04480377212166786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1195492073893547, + 0.14224651455879211, + 0.09931184351444244, + 0.14971695840358734, + 0.041826386004686356, + 0.10791049897670746, + 0.07599844038486481, + 0.07162650674581528, + 0.09473638236522675, + 0.09707727283239365 + ], + "credit_map": [ + 0.012002832256257534, + 0.01359699759632349, + 0.01817484386265278, + 0.01218888908624649, + 0.016243716701865196, + 0.012191061861813068, + 0.01048607099801302, + 0.024664798751473427, + 0.014988274313509464, + 0.018996192142367363, + 0.018038056790828705, + 0.01044924184679985, + 0.017296042293310165, + 0.023009223863482475, + 0.02122475579380989, + 0.018244726583361626, + 0.01132832933217287, + 0.01518910750746727, + 0.014185581356287003, + 0.009318336844444275, + 0.008881875313818455, + 0.014544184319674969, + 0.011174989864230156, + 0.01468943152576685, + 0.013858851976692677, + 0.007456610444933176, + 0.02031133882701397, + 0.01106333453208208, + 0.013456415385007858, + 0.0075454567559063435, + 0.04475215822458267, + 0.03127278387546539, + 0.004033635836094618, + 0.0030768977012485266, + 0.00492045097053051, + 0.008954472839832306, + 0.005067349411547184, + 0.004646600224077702, + 0.005524380598217249, + 0.005602600052952766, + 0.008036302402615547, + 0.008362928405404091, + 0.018648331984877586, + 0.015005324967205524, + 0.01028531789779663, + 0.010568779893219471, + 0.018175559118390083, + 0.01882796175777912, + 0.007541261613368988, + 0.007788019720464945, + 0.014051900245249271, + 0.011502356268465519, + 0.007652859669178724, + 0.009906576946377754, + 0.009984138421714306, + 0.007571326568722725, + 0.0065057286992669106, + 0.005563672166317701, + 0.009363214485347271, + 0.011000332422554493, + 0.008495273999869823, + 0.005451067350804806, + 0.01517945434898138, + 0.010067768394947052, + 0.006954160984605551, + 0.0041558886878192425, + 0.020595457404851913, + 0.013487596996128559, + 0.008198651485145092, + 0.0051201325841248035, + 0.01345842331647873, + 0.02276606857776642, + 0.010962048545479774, + 0.011076221242547035, + 0.01249970868229866, + 0.014862347394227982, + 0.014374924823641777, + 0.009555288590490818, + 0.012575596570968628, + 0.011171135120093822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.04480377212166786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744833197444677 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.006744833197444677, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744833197444677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15444885194301605, + 0.08208785951137543, + 0.0366775281727314, + 0.07596824318170547, + 0.07996199280023575, + 0.13421505689620972, + 0.0592300221323967, + 0.18269285559654236, + 0.11894769221544266, + 0.07576989382505417 + ], + "credit_map": [ + 0.022107988595962524, + 0.0207948237657547, + 0.02099153958261013, + 0.011342249810695648, + 0.020502787083387375, + 0.016960501670837402, + 0.02244436927139759, + 0.019304588437080383, + 0.007890628650784492, + 0.012316468171775341, + 0.01040157862007618, + 0.010736923664808273, + 0.010297862812876701, + 0.008012179285287857, + 0.011352581903338432, + 0.011079641990363598, + 0.003342179348692298, + 0.0026806064415723085, + 0.004722345620393753, + 0.004398125223815441, + 0.008291950449347496, + 0.0035439187195152044, + 0.004236920271068811, + 0.005461482331156731, + 0.007599020376801491, + 0.006455220282077789, + 0.014959988184273243, + 0.011872346512973309, + 0.008376259356737137, + 0.005276579409837723, + 0.0069771939888596535, + 0.014451632276177406, + 0.005462080240249634, + 0.006064732093364, + 0.01098552718758583, + 0.014770958572626114, + 0.01578792929649353, + 0.005604465492069721, + 0.011287414468824863, + 0.009998886846005917, + 0.015388378873467445, + 0.008408223278820515, + 0.026337632909417152, + 0.018330484628677368, + 0.012925303541123867, + 0.01215735450387001, + 0.019279710948467255, + 0.021387966349720955, + 0.007331791799515486, + 0.006807038560509682, + 0.007526632398366928, + 0.008595578372478485, + 0.00611539464443922, + 0.008459153585135937, + 0.0061041442677378654, + 0.008290290832519531, + 0.015496543608605862, + 0.011132397688925266, + 0.019699351862072945, + 0.03230150043964386, + 0.02040756121277809, + 0.013722412288188934, + 0.04512358829379082, + 0.024809498339891434, + 0.008032296784222126, + 0.007825546897947788, + 0.03194013983011246, + 0.013771800324320793, + 0.008571089245378971, + 0.005910596344619989, + 0.01039673388004303, + 0.032499492168426514, + 0.00796088483184576, + 0.009449457749724388, + 0.008940472267568111, + 0.012144831009209156, + 0.010514369234442711, + 0.010701719671487808, + 0.006109302397817373, + 0.009948855265974998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.006744833197444677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248932633548975 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005248932633548975, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248932633548975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09096900373697281, + 0.07563766092061996, + 0.07203853130340576, + 0.07119207829236984, + 0.07510931044816971, + 0.3347535729408264, + 0.04234731197357178, + 0.11461445689201355, + 0.06862416118383408, + 0.054713934659957886 + ], + "credit_map": [ + 0.015444528311491013, + 0.008864840492606163, + 0.013297031633555889, + 0.009416447952389717, + 0.007341340649873018, + 0.014702025800943375, + 0.009656899608671665, + 0.01224588043987751, + 0.007004129700362682, + 0.0066093807108700275, + 0.010130377486348152, + 0.010380944237112999, + 0.011780159547924995, + 0.005996359512209892, + 0.013979459181427956, + 0.009756844490766525, + 0.004674824886023998, + 0.008214421570301056, + 0.010392761789262295, + 0.01062831375747919, + 0.00841559749096632, + 0.006031624507158995, + 0.012868468649685383, + 0.010812514461576939, + 0.0065915342420339584, + 0.006694466806948185, + 0.01145720761269331, + 0.009671622887253761, + 0.006683604326099157, + 0.004050250165164471, + 0.009848693385720253, + 0.016194693744182587, + 0.009594314731657505, + 0.007168760523200035, + 0.01069947425276041, + 0.007961283437907696, + 0.006593106314539909, + 0.0067602465860545635, + 0.011759272776544094, + 0.014572841115295887, + 0.03847561776638031, + 0.020046090707182884, + 0.04517941176891327, + 0.05256933718919754, + 0.03977718576788902, + 0.03632858395576477, + 0.04101017490029335, + 0.061367154121398926, + 0.003955740947276354, + 0.004503854084759951, + 0.0048951455391943455, + 0.005925008095800877, + 0.006347836460918188, + 0.0059590148739516735, + 0.0054510654881596565, + 0.005309642758220434, + 0.011710871011018753, + 0.008071904070675373, + 0.015784895047545433, + 0.020940765738487244, + 0.012261291034519672, + 0.008218246512115002, + 0.02243967354297638, + 0.015186796896159649, + 0.006972470786422491, + 0.004254543222486973, + 0.011886049062013626, + 0.013483026996254921, + 0.005741840228438377, + 0.006405217573046684, + 0.009142766706645489, + 0.010738241486251354, + 0.004753386601805687, + 0.007016486022621393, + 0.006254006642848253, + 0.010229207575321198, + 0.007160557899624109, + 0.005790800787508488, + 0.00655513908714056, + 0.00695434445515275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.005248932633548975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003185557434335351 + ], + "opposite_logprob_tokens": [ + -5.753185749053955 + ], + "logprob_reward": -0.003185557434335351, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003185557434335351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04243067651987076, + 0.04529373347759247, + 0.050732582807540894, + 0.014899522997438908, + 0.045393310487270355, + 0.08964426815509796, + 0.1282036453485489, + 0.1877877116203308, + 0.35514312982559204, + 0.04047136753797531 + ], + "credit_map": [ + 0.008700951933860779, + 0.0037453051190823317, + 0.0057882578112185, + 0.0021549202501773834, + 0.004303407855331898, + 0.007976952940225601, + 0.003782296786084771, + 0.005978586617857218, + 0.004209015052765608, + 0.0063438694924116135, + 0.006336144637316465, + 0.006116924341768026, + 0.004517410881817341, + 0.006287053693085909, + 0.004908413160592318, + 0.006574905943125486, + 0.005449303425848484, + 0.006675411947071552, + 0.006963687017560005, + 0.004171888343989849, + 0.006114740390330553, + 0.004660099279135466, + 0.007463959977030754, + 0.009233497083187103, + 0.0010626302100718021, + 0.0010738089913502336, + 0.0027027889154851437, + 0.002843210007995367, + 0.0016933458391577005, + 0.0010449255350977182, + 0.0021978062577545643, + 0.0022810085210949183, + 0.003931610845029354, + 0.003962680697441101, + 0.007367434911429882, + 0.007745291106402874, + 0.0040648095309734344, + 0.0035265653859823942, + 0.006811709608882666, + 0.007983213290572166, + 0.008477816358208656, + 0.009438745677471161, + 0.010674039833247662, + 0.014097750186920166, + 0.011340891942381859, + 0.005998050328344107, + 0.01336086168885231, + 0.016256127506494522, + 0.015244601294398308, + 0.01327008381485939, + 0.012026885524392128, + 0.01642991229891777, + 0.015214337967336178, + 0.027149969711899757, + 0.013989455066621304, + 0.014878417365252972, + 0.024030612781643867, + 0.013663049787282944, + 0.02767966128885746, + 0.03039230965077877, + 0.015957603231072426, + 0.014694641344249249, + 0.03336441516876221, + 0.028005436062812805, + 0.038139089941978455, + 0.022814489901065826, + 0.0585629902780056, + 0.05684211105108261, + 0.027598043903708458, + 0.02993849851191044, + 0.0498988963663578, + 0.07134904712438583, + 0.004244544543325901, + 0.0030926347244530916, + 0.006596820894628763, + 0.00872666947543621, + 0.005623152945190668, + 0.0028241139370948076, + 0.004742836579680443, + 0.004620596766471863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.003185557434335351 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028143818490207195 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.0028143818490207195, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028143818490207195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1031695231795311, + 0.05909239500761032, + 0.054474204778671265, + 0.07936238497495651, + 0.05299853906035423, + 0.19574211537837982, + 0.09885252267122269, + 0.1222294270992279, + 0.1810971200466156, + 0.05298179015517235 + ], + "credit_map": [ + 0.019949236884713173, + 0.008280202746391296, + 0.01601085625588894, + 0.006641521118581295, + 0.010506082326173782, + 0.01585768721997738, + 0.008245591074228287, + 0.017678342759609222, + 0.0069512068293988705, + 0.007964170537889004, + 0.007129482459276915, + 0.006729558110237122, + 0.007265182677656412, + 0.006532356608659029, + 0.008071602322161198, + 0.008448834531009197, + 0.005365276243537664, + 0.004415892995893955, + 0.008364611305296421, + 0.006276404485106468, + 0.004970750771462917, + 0.0036626753862947226, + 0.008483116514980793, + 0.01293548010289669, + 0.007034613285213709, + 0.006751153618097305, + 0.01583380065858364, + 0.010799866169691086, + 0.009510999545454979, + 0.006512255407869816, + 0.015490535646677017, + 0.007429155055433512, + 0.0036582816392183304, + 0.00345722958445549, + 0.006985521875321865, + 0.009931830689311028, + 0.007758254185318947, + 0.003375164233148098, + 0.007857552729547024, + 0.009974706918001175, + 0.015809310600161552, + 0.01634199731051922, + 0.03125511854887009, + 0.027900749817490578, + 0.021233191713690758, + 0.013661205768585205, + 0.029151195660233498, + 0.040389351546764374, + 0.012638331390917301, + 0.010708456858992577, + 0.010681135579943657, + 0.011152691207826138, + 0.012080139480531216, + 0.016116544604301453, + 0.011568523943424225, + 0.013906696811318398, + 0.014818464405834675, + 0.011734062805771828, + 0.013240494765341282, + 0.016500016674399376, + 0.0121615594252944, + 0.01682768017053604, + 0.02084709145128727, + 0.01610005833208561, + 0.027095941826701164, + 0.016843508929014206, + 0.020740332081913948, + 0.03524085506796837, + 0.01779600977897644, + 0.015336696989834309, + 0.02600763365626335, + 0.022036148235201836, + 0.005295153241604567, + 0.004244443029165268, + 0.0070443362928926945, + 0.010440286248922348, + 0.005466931965202093, + 0.0062230839394032955, + 0.007920919917523861, + 0.006346635986119509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.0028143818490207195 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, set against the backdrop of the water level, captures a stern-faced hunter, his face gaunt and weathered, his long black hair plastered to his scalp, slowly and deliberately immersing himself in the dark, murky water at the heart of a dusky swamp. They are clad in worn, dark brown leather, the material cracked and waterlogged, its tarnished burrs gleaming dimly. The scene is rendered in a somber, painterly style, characterized by dramatic black-and-white contrasts and rich, painterly textures that give the gnarled, moss-covered roots in the background an almost tangible quality. As the hunters sink unflinchingly, water pours over their clenched jaws and thin lips, their posture deliberately stiff. A faint ray of light filters through the dense, oppressive canopy of trees above, capturing the steely determination in their pale, wide-open eyes, a last, sharp glimmer of defiance. Just before the dark, oily water completely obscures their faces, the light fades, leaving only faint ripples spreading across the black void.", + "seed": 280, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_14/index_2290_seed_280_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": -0.016514858324080707, + "vlm_logprob": -0.016514858324080707, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.016514858324080707, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04345782846212387 + ], + "opposite_logprob_tokens": [ + -3.4184577465057373 + ], + "logprob_reward": -0.04345782846212387, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04345782846212387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15083137154579163, + 0.07045793533325195, + 0.05546531826257706, + 0.06046406552195549, + 0.061944976449012756, + 0.11972156912088394, + 0.1029682382941246, + 0.22673092782497406, + 0.08011326938867569, + 0.07130236178636551 + ], + "credit_map": [ + 0.015572710894048214, + 0.01655462570488453, + 0.016645990312099457, + 0.019909827038645744, + 0.016321975737810135, + 0.01890876702964306, + 0.024864668026566505, + 0.022052794694900513, + 0.006108825094997883, + 0.005492917727679014, + 0.008140189573168755, + 0.010804476216435432, + 0.008526541292667389, + 0.010199603624641895, + 0.011449364945292473, + 0.0097360173240304, + 0.0051056439988315105, + 0.009487763047218323, + 0.005498840007930994, + 0.008753092959523201, + 0.005602166522294283, + 0.004914358723908663, + 0.007850616239011288, + 0.008252832107245922, + 0.005517066456377506, + 0.006390056107193232, + 0.0055570234544575214, + 0.01062670350074768, + 0.004230653401464224, + 0.0057337419129908085, + 0.012925795279443264, + 0.00948302261531353, + 0.004341687075793743, + 0.004920573905110359, + 0.0109260194003582, + 0.007958211936056614, + 0.006127413362264633, + 0.01044031884521246, + 0.008419479243457317, + 0.008811275474727154, + 0.009909802116453648, + 0.011266638524830341, + 0.01908726617693901, + 0.021319985389709473, + 0.012731528840959072, + 0.012035301886498928, + 0.017430393025279045, + 0.01594066247344017, + 0.006438495125621557, + 0.01747736893594265, + 0.012130774557590485, + 0.01771366596221924, + 0.0094590550288558, + 0.011584858410060406, + 0.014918728731572628, + 0.013245292007923126, + 0.011248402297496796, + 0.012837878428399563, + 0.06787296384572983, + 0.019233625382184982, + 0.013578475452959538, + 0.020473767071962357, + 0.020130833610892296, + 0.06135497987270355, + 0.006739728152751923, + 0.006325636990368366, + 0.01779324747622013, + 0.008811176754534245, + 0.00600972305983305, + 0.011853551492094994, + 0.007708044722676277, + 0.014872158877551556, + 0.006306973285973072, + 0.010926995426416397, + 0.010535351000726223, + 0.007569848094135523, + 0.0066711995750665665, + 0.008566715754568577, + 0.012003602460026741, + 0.008721674792468548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.04345782846212387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.28933173418045044, + 0.0782102644443512, + 0.13993549346923828, + 0.029967231675982475, + 0.07126478105783463, + 0.0793333575129509, + 0.05995303764939308, + 0.10815633833408356, + 0.08732856065034866, + 0.056519266217947006 + ], + "credit_map": [ + 0.02623744122684002, + 0.02797737717628479, + 0.035237234085798264, + 0.04459227994084358, + 0.031086081638932228, + 0.029683910310268402, + 0.05198900029063225, + 0.04252838343381882, + 0.007388890255242586, + 0.009912746958434582, + 0.009751889854669571, + 0.014288489706814289, + 0.007877826690673828, + 0.007913824170827866, + 0.009532193653285503, + 0.011544402688741684, + 0.011395283974707127, + 0.014943161979317665, + 0.02528911828994751, + 0.016611479222774506, + 0.015959065407514572, + 0.01182269211858511, + 0.022926989942789078, + 0.02098769322037697, + 0.0026750084944069386, + 0.0032260892912745476, + 0.005153598729521036, + 0.0040803211741149426, + 0.002508176490664482, + 0.0036475355736911297, + 0.00367900007404387, + 0.00499750254675746, + 0.0046502877958118916, + 0.006909873802214861, + 0.015858154743909836, + 0.01053214818239212, + 0.0061576226726174355, + 0.005641716066747904, + 0.009363319724798203, + 0.012151653878390789, + 0.005283756647258997, + 0.009062123484909534, + 0.008833604864776134, + 0.018338367342948914, + 0.00978138204663992, + 0.008537848480045795, + 0.01044554729014635, + 0.009050724096596241, + 0.004645511042326689, + 0.0051991683430969715, + 0.011347784660756588, + 0.00983422715216875, + 0.004902970977127552, + 0.007031062617897987, + 0.009157584048807621, + 0.007834730669856071, + 0.006504695862531662, + 0.005917364731431007, + 0.03101743385195732, + 0.008460265584290028, + 0.007994398474693298, + 0.010732605122029781, + 0.009321912191808224, + 0.02820766158401966, + 0.005627371370792389, + 0.007565028965473175, + 0.011558347381651402, + 0.015699537470936775, + 0.008269284851849079, + 0.009590128436684608, + 0.019318226724863052, + 0.009700633585453033, + 0.00350275170058012, + 0.005245857872068882, + 0.011559978127479553, + 0.00769921625033021, + 0.004222482442855835, + 0.007237802725285292, + 0.0071305460296571255, + 0.009920631535351276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.023304030299186707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781643748283386 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781643748283386, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781643748283386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23808661103248596, + 0.17841236293315887, + 0.08687691390514374, + 0.03544824197888374, + 0.06884804368019104, + 0.049320291727781296, + 0.07170943915843964, + 0.05583978071808815, + 0.1038428470492363, + 0.11161541193723679 + ], + "credit_map": [ + 0.022629423066973686, + 0.02192976139485836, + 0.02646440453827381, + 0.03553876280784607, + 0.018313437700271606, + 0.02262006886303425, + 0.05056840181350708, + 0.040022362023591995, + 0.013491286896169186, + 0.020868217572569847, + 0.022626427933573723, + 0.036176249384880066, + 0.017388267442584038, + 0.017820803448557854, + 0.023923134431242943, + 0.026117971166968346, + 0.006145878229290247, + 0.010067821480333805, + 0.01298312284052372, + 0.014901025220751762, + 0.009608032181859016, + 0.010366125032305717, + 0.011523005552589893, + 0.011281905695796013, + 0.0033494429662823677, + 0.0025774987880140543, + 0.00577983632683754, + 0.006697006989270449, + 0.0021924637258052826, + 0.002969746245071292, + 0.006588734220713377, + 0.005293515510857105, + 0.003962436690926552, + 0.007089448627084494, + 0.014536636881530285, + 0.010165579617023468, + 0.006331346929073334, + 0.007263027131557465, + 0.010052911005914211, + 0.009446654468774796, + 0.003001126926392317, + 0.004584961105138063, + 0.0064841266721487045, + 0.009753947146236897, + 0.004926889203488827, + 0.007875055074691772, + 0.005980192217975855, + 0.006713994313031435, + 0.0057249111123383045, + 0.011563955806195736, + 0.00937724206596613, + 0.010440885089337826, + 0.006920762825757265, + 0.008672821335494518, + 0.008953262120485306, + 0.010055598802864552, + 0.005239018704742193, + 0.005035602953284979, + 0.007503085304051638, + 0.008633410558104515, + 0.005829325877130032, + 0.008901044726371765, + 0.008448760025203228, + 0.0062495325691998005, + 0.007675488013774157, + 0.008775060996413231, + 0.01350163109600544, + 0.017338398844003677, + 0.009753822349011898, + 0.011655179783701897, + 0.024552121758461, + 0.010591146536171436, + 0.005186361726373434, + 0.007289589382708073, + 0.03093254193663597, + 0.011955461464822292, + 0.005973338149487972, + 0.00808895193040371, + 0.010265352204442024, + 0.0319238118827343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.009781643748283386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090277012437582 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.004090277012437582, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090277012437582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12357176840305328, + 0.12987935543060303, + 0.05970432609319687, + 0.037129346281290054, + 0.05487004294991493, + 0.12155435979366302, + 0.18306031823158264, + 0.047098107635974884, + 0.1725054383277893, + 0.07062697410583496 + ], + "credit_map": [ + 0.012904674746096134, + 0.010786807164549828, + 0.015867652371525764, + 0.016940144822001457, + 0.013988800346851349, + 0.017683768644928932, + 0.016931520774960518, + 0.018468400463461876, + 0.020684322342276573, + 0.016127394512295723, + 0.014438937418162823, + 0.015769632533192635, + 0.013670277781784534, + 0.01497295405715704, + 0.014610469341278076, + 0.019605373963713646, + 0.004181565251201391, + 0.005143698770552874, + 0.007899213582277298, + 0.008487083949148655, + 0.008459717966616154, + 0.008677649311721325, + 0.009533945471048355, + 0.007321451790630817, + 0.003502273466438055, + 0.003020295174792409, + 0.006141690071672201, + 0.005641039926558733, + 0.003473500721156597, + 0.005503649823367596, + 0.004086238332092762, + 0.005760658532381058, + 0.004075757693499327, + 0.006137312855571508, + 0.01066330075263977, + 0.009319177828729153, + 0.005492888856679201, + 0.005215477664023638, + 0.0064630331471562386, + 0.0075030941516160965, + 0.010974506847560406, + 0.008006490767002106, + 0.011135808192193508, + 0.012683289125561714, + 0.020874379202723503, + 0.029138116165995598, + 0.011047928594052792, + 0.017693841829895973, + 0.01585603877902031, + 0.01321074552834034, + 0.032165173441171646, + 0.025788623839616776, + 0.0134175019338727, + 0.022356418892741203, + 0.027508093044161797, + 0.03275773301720619, + 0.0035387566313147545, + 0.005395067855715752, + 0.007051179185509682, + 0.006094886921346188, + 0.004963168408721685, + 0.008134122006595135, + 0.004754130728542805, + 0.0071667940355837345, + 0.011350784450769424, + 0.015925610437989235, + 0.0216381773352623, + 0.028403205797076225, + 0.016525913029909134, + 0.019832314923405647, + 0.04214119538664818, + 0.016688242554664612, + 0.004853877238929272, + 0.009060717187821865, + 0.010378344915807247, + 0.00990168284624815, + 0.007340427953749895, + 0.01030261255800724, + 0.008158102631568909, + 0.01063120923936367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.004090277012437582 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2613028287887573, + 0.07271520048379898, + 0.037505824118852615, + 0.043646469712257385, + 0.08215171843767166, + 0.15118137001991272, + 0.09216322004795074, + 0.1114717423915863, + 0.06896694004535675, + 0.07889474183320999 + ], + "credit_map": [ + 0.019365185871720314, + 0.02510463446378708, + 0.026973199099302292, + 0.044526975601911545, + 0.024672510102391243, + 0.021399246528744698, + 0.059100013226270676, + 0.04016103968024254, + 0.009278289042413235, + 0.009318741038441658, + 0.00849251076579094, + 0.01001668069511652, + 0.006498078815639019, + 0.008238347247242928, + 0.009955250658094883, + 0.010917294770479202, + 0.00464688939973712, + 0.0036826101131737232, + 0.004479227587580681, + 0.004802734591066837, + 0.004978398326784372, + 0.004694052971899509, + 0.004932008218020201, + 0.005289901979267597, + 0.0039293523877859116, + 0.0031861907336860895, + 0.00593095738440752, + 0.004805042874068022, + 0.004279692191630602, + 0.00564001826569438, + 0.009393504820764065, + 0.0064817131496965885, + 0.006363691762089729, + 0.011866796761751175, + 0.010467026382684708, + 0.01349340658634901, + 0.00805887021124363, + 0.008030587807297707, + 0.011300122365355492, + 0.012571217492222786, + 0.010603148490190506, + 0.013858919031918049, + 0.029878685250878334, + 0.024853063747286797, + 0.011746523901820183, + 0.014447775669395924, + 0.02367793396115303, + 0.022115321829915047, + 0.005497056990861893, + 0.009679105132818222, + 0.015458672307431698, + 0.016813114285469055, + 0.006625700742006302, + 0.008885676972568035, + 0.01576213911175728, + 0.013441747054457664, + 0.006020562723278999, + 0.006175368558615446, + 0.026753589510917664, + 0.011027216911315918, + 0.00797334499657154, + 0.014122680760920048, + 0.013268237002193928, + 0.02613074705004692, + 0.004553122445940971, + 0.006372426170855761, + 0.00968240574002266, + 0.012982064858078957, + 0.006397856865078211, + 0.007246135268360376, + 0.0151373241096735, + 0.0065956017933785915, + 0.006704824976623058, + 0.012655416503548622, + 0.010788694024085999, + 0.009572855196893215, + 0.007041793782263994, + 0.009162538684904575, + 0.014209640212357044, + 0.008758977055549622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0019405120983719826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, set against the backdrop of the water level, captures a stern-faced hunter, his face gaunt and weathered, his long black hair plastered to his scalp, slowly and deliberately immersing himself in the dark, murky water at the heart of a dusky swamp. They are clad in worn, dark brown leather, the material cracked and waterlogged, its tarnished burrs gleaming dimly. The scene is rendered in a somber, painterly style, characterized by dramatic black-and-white contrasts and rich, painterly textures that give the gnarled, moss-covered roots in the background an almost tangible quality. As the hunters sink unflinchingly, water pours over their clenched jaws and thin lips, their posture deliberately stiff. A faint ray of light filters through the dense, oppressive canopy of trees above, capturing the steely determination in their pale, wide-open eyes, a last, sharp glimmer of defiance. Just before the dark, oily water completely obscures their faces, the light fades, leaving only faint ripples spreading across the black void.", + "seed": 281, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_14/index_2290_seed_281_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": -0.01662756083533168, + "vlm_logprob": -0.01662756083533168, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.01662756083533168, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05408124998211861 + ], + "opposite_logprob_tokens": [ + -3.1790812015533447 + ], + "logprob_reward": -0.05408124998211861, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05408124998211861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10772620141506195, + 0.15567557513713837, + 0.08528638631105423, + 0.08530910313129425, + 0.04058236628770828, + 0.06347086280584335, + 0.10356063395738602, + 0.06568820029497147, + 0.1341405212879181, + 0.15856018662452698 + ], + "credit_map": [ + 0.008284316398203373, + 0.010784744285047054, + 0.01750854216516018, + 0.017284229397773743, + 0.011438585817813873, + 0.015167185105383396, + 0.011773757636547089, + 0.015484833158552647, + 0.030279356986284256, + 0.020903216674923897, + 0.015249337069690228, + 0.016325151547789574, + 0.013242769055068493, + 0.016760818660259247, + 0.02342384122312069, + 0.01949106901884079, + 0.007194568403065205, + 0.0066648307256400585, + 0.007903196848928928, + 0.019927652552723885, + 0.010881382040679455, + 0.00971776433289051, + 0.012069967575371265, + 0.010927018709480762, + 0.013216095976531506, + 0.009853589348495007, + 0.0051510208286345005, + 0.005132236052304506, + 0.004430592525750399, + 0.008573380298912525, + 0.01911883056163788, + 0.019833354279398918, + 0.003212700365111232, + 0.003208219539374113, + 0.0055910395458340645, + 0.006278344430029392, + 0.005375836510211229, + 0.002954706083983183, + 0.008789691142737865, + 0.005171827040612698, + 0.005286906845867634, + 0.006715551018714905, + 0.007427228149026632, + 0.008451656438410282, + 0.00842492375522852, + 0.0046152682043612, + 0.014973939396440983, + 0.007575387135148048, + 0.010779326781630516, + 0.011875553987920284, + 0.011559011414647102, + 0.015907807275652885, + 0.010326125659048557, + 0.008177115581929684, + 0.018579410389065742, + 0.016356274485588074, + 0.009620263241231441, + 0.009913030080497265, + 0.010127810761332512, + 0.005825737025588751, + 0.005638606380671263, + 0.0059105828404426575, + 0.00979598518460989, + 0.008856181055307388, + 0.01430382952094078, + 0.018047064542770386, + 0.01724865846335888, + 0.013380292803049088, + 0.018540052697062492, + 0.017627205699682236, + 0.015177243389189243, + 0.019816158339381218, + 0.010300456546247005, + 0.012562083080410957, + 0.02589421719312668, + 0.016530318185687065, + 0.010219979099929333, + 0.013832004740834236, + 0.012823505327105522, + 0.056397613137960434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.05408124998211861 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086884886026382 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.011086884886026382, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086884886026382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.247949481010437, + 0.13688912987709045, + 0.11825168132781982, + 0.05153552070260048, + 0.09213551133871078, + 0.03721242398023605, + 0.10527492314577103, + 0.057832881808280945, + 0.07719754427671432, + 0.07572093605995178 + ], + "credit_map": [ + 0.047870732843875885, + 0.028047526255249977, + 0.031498298048973083, + 0.03018196113407612, + 0.033798422664403915, + 0.023119349032640457, + 0.02259027026593685, + 0.030842917039990425, + 0.016199519857764244, + 0.01327514834702015, + 0.020966773852705956, + 0.023032166063785553, + 0.013630100525915623, + 0.013396997936069965, + 0.016762955114245415, + 0.01962546445429325, + 0.010418280959129333, + 0.01015661284327507, + 0.017880134284496307, + 0.020186301320791245, + 0.014228502288460732, + 0.009778229519724846, + 0.018124204128980637, + 0.017479410395026207, + 0.00366958393715322, + 0.0035361615009605885, + 0.004744304344058037, + 0.004076839424669743, + 0.002851171651855111, + 0.0038222603034228086, + 0.015512979589402676, + 0.013322222046554089, + 0.0070116035640239716, + 0.008036886341869831, + 0.01076705101877451, + 0.011392060667276382, + 0.012962338514626026, + 0.00923977605998516, + 0.022087324410676956, + 0.010638461448252201, + 0.0036715043243020773, + 0.00379648944362998, + 0.004682442173361778, + 0.006242420058697462, + 0.004973460920155048, + 0.003018015995621681, + 0.0056154318153858185, + 0.005212660413235426, + 0.013177234679460526, + 0.011887704022228718, + 0.01401376910507679, + 0.016119956970214844, + 0.011625302955508232, + 0.009893078356981277, + 0.017157936468720436, + 0.011399942450225353, + 0.007496856618672609, + 0.008369027636945248, + 0.007271463517099619, + 0.0071283383294939995, + 0.0064131123945117, + 0.00701211579144001, + 0.006726179737597704, + 0.007415786385536194, + 0.009811620227992535, + 0.010630344040691853, + 0.00843767635524273, + 0.00767506193369627, + 0.011873404495418072, + 0.010752879083156586, + 0.008166457526385784, + 0.009850099682807922, + 0.007329942192882299, + 0.008574200794100761, + 0.010522435419261456, + 0.008139877580106258, + 0.0076557425782084465, + 0.0072267078794538975, + 0.007105414289981127, + 0.01916661486029625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011086884886026382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863473117351532 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.00863473117351532, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863473117351532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2341318279504776, + 0.1312016248703003, + 0.08585043996572495, + 0.02976278029382229, + 0.05572438985109329, + 0.04290037602186203, + 0.2366950511932373, + 0.05015925318002701, + 0.07513441145420074, + 0.05843985453248024 + ], + "credit_map": [ + 0.0509721040725708, + 0.03156586363911629, + 0.023725565522909164, + 0.030695555731654167, + 0.021110115572810173, + 0.021913649514317513, + 0.025015344843268394, + 0.029133634641766548, + 0.023796500638127327, + 0.01777230203151703, + 0.01092016976326704, + 0.016908882185816765, + 0.01325566042214632, + 0.01126212440431118, + 0.02129044570028782, + 0.01599554717540741, + 0.012376009486615658, + 0.008317913860082626, + 0.009174077771604061, + 0.012738135643303394, + 0.008779597468674183, + 0.008795009925961494, + 0.014679985120892525, + 0.01098970603197813, + 0.0029811938293278217, + 0.0034605353139340878, + 0.0034522407222539186, + 0.003995237406343222, + 0.002985806902870536, + 0.004506655968725681, + 0.005194772966206074, + 0.0031863374169915915, + 0.004264737479388714, + 0.004493724554777145, + 0.0079822838306427, + 0.011868509463965893, + 0.00558180408552289, + 0.005757504608482122, + 0.007636440452188253, + 0.008139383047819138, + 0.005841468460857868, + 0.004331847652792931, + 0.004845182877033949, + 0.00635987613350153, + 0.006011508405208588, + 0.0031145645771175623, + 0.007258923724293709, + 0.0051370058208703995, + 0.024675630033016205, + 0.022434892132878304, + 0.030091287568211555, + 0.03616074472665787, + 0.025415964424610138, + 0.02313248999416828, + 0.045312993228435516, + 0.029471050947904587, + 0.0055268569849431515, + 0.004915083292871714, + 0.0059823631308972836, + 0.004837859887629747, + 0.006539344787597656, + 0.006465131416916847, + 0.010142084211111069, + 0.005750528071075678, + 0.008771943859755993, + 0.010184223763644695, + 0.00794487539678812, + 0.008708245120942593, + 0.010595499537885189, + 0.01008180994540453, + 0.009617715142667294, + 0.009230095893144608, + 0.006152720656245947, + 0.005856333300471306, + 0.007832619361579418, + 0.005530995316803455, + 0.006990370806306601, + 0.006545347161591053, + 0.005212008487433195, + 0.014319456182420254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.00863473117351532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088140092790127 + ], + "opposite_logprob_tokens": [ + -5.504087924957275 + ], + "logprob_reward": -0.004088140092790127, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088140092790127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2002062350511551, + 0.1247524693608284, + 0.09397824108600616, + 0.04204784706234932, + 0.05402930825948715, + 0.060514453798532486, + 0.1709710657596588, + 0.07488071173429489, + 0.08033648878335953, + 0.09828320890665054 + ], + "credit_map": [ + 0.03382875770330429, + 0.02925625815987587, + 0.029711561277508736, + 0.02737150900065899, + 0.012664923444390297, + 0.022318962961435318, + 0.017613476142287254, + 0.027440767735242844, + 0.021826090291142464, + 0.015660326927900314, + 0.011731736361980438, + 0.01880616322159767, + 0.013180325739085674, + 0.010894734412431717, + 0.015305602923035622, + 0.017347484827041626, + 0.010277862660586834, + 0.007752277422696352, + 0.015963610261678696, + 0.018541868776082993, + 0.00818569678813219, + 0.008565988391637802, + 0.010023146867752075, + 0.014667782932519913, + 0.008663339540362358, + 0.006573343649506569, + 0.004723707213997841, + 0.004026591777801514, + 0.002899742918089032, + 0.0037030410021543503, + 0.005399338435381651, + 0.00605873903259635, + 0.003919397480785847, + 0.005522351711988449, + 0.007909528911113739, + 0.010187181644141674, + 0.005065169185400009, + 0.005996846593916416, + 0.007738486863672733, + 0.0076903412118554115, + 0.00577376177534461, + 0.0064741685055196285, + 0.007165549788624048, + 0.009273283183574677, + 0.008607223629951477, + 0.005396698601543903, + 0.010440995916724205, + 0.007382767274975777, + 0.016183141618967056, + 0.01889866031706333, + 0.02285362035036087, + 0.023619623854756355, + 0.018252132460474968, + 0.02220843732357025, + 0.02854977734386921, + 0.020405655726790428, + 0.008730335161089897, + 0.009414522908627987, + 0.00946141965687275, + 0.007970594801008701, + 0.007312503177672625, + 0.009168237447738647, + 0.01561895851045847, + 0.007204130757600069, + 0.009597745724022388, + 0.008204704150557518, + 0.009341498836874962, + 0.007467149291187525, + 0.01157237496227026, + 0.013705789111554623, + 0.010685177519917488, + 0.009762044996023178, + 0.010560128837823868, + 0.010196969844400883, + 0.012326402589678764, + 0.01004712749272585, + 0.014324383810162544, + 0.013069003820419312, + 0.00940292701125145, + 0.018356258049607277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.004088140092790127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects depicted as stern-faced hunters with gaunt, weathered faces and long black hair plastered to their scalps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hunters wearing worn, dark brown leather clothing that appears cracked and waterlogged?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dusky swamp with dark, murky water and gnarled, moss-covered roots in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style somber and painterly, featuring dramatic black-and-white contrasts and rich textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246798042207956 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.005246798042207956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246798042207956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1482527107000351, + 0.21632260084152222, + 0.06383495777845383, + 0.05065256729722023, + 0.08357762545347214, + 0.040924716740846634, + 0.14330638945102692, + 0.10164251178503036, + 0.06869592517614365, + 0.08279004693031311 + ], + "credit_map": [ + 0.014361572451889515, + 0.015800300985574722, + 0.016500160098075867, + 0.022103795781731606, + 0.019992752000689507, + 0.02134423516690731, + 0.01720021478831768, + 0.02094966359436512, + 0.0365164615213871, + 0.03406447172164917, + 0.03032873198390007, + 0.0224615428596735, + 0.012440742924809456, + 0.030429305508732796, + 0.02361813187599182, + 0.026463204994797707, + 0.008161132223904133, + 0.006409910507500172, + 0.00657530827447772, + 0.008618488907814026, + 0.009584227576851845, + 0.005536094307899475, + 0.009773624129593372, + 0.009176167659461498, + 0.009192551486194134, + 0.007756845094263554, + 0.004853900521993637, + 0.004011843353509903, + 0.0028902278281748295, + 0.005023355130106211, + 0.008431647904217243, + 0.008492190390825272, + 0.009034586139023304, + 0.011761734262108803, + 0.010386976413428783, + 0.011513748206198215, + 0.007596867624670267, + 0.013667124323546886, + 0.009710561484098434, + 0.009906018152832985, + 0.004440938122570515, + 0.004270236473530531, + 0.005335659720003605, + 0.006533932872116566, + 0.004153979942202568, + 0.0030631800182163715, + 0.0064281583763659, + 0.0066986288875341415, + 0.014122866094112396, + 0.014309258200228214, + 0.017581354826688766, + 0.02683679759502411, + 0.01619119569659233, + 0.012576880864799023, + 0.02523931674659252, + 0.01644871011376381, + 0.01819821447134018, + 0.016999464482069016, + 0.011321420781314373, + 0.009447366930544376, + 0.01102228369563818, + 0.00931103527545929, + 0.014880725182592869, + 0.01046199444681406, + 0.008657768368721008, + 0.007305853068828583, + 0.007829650305211544, + 0.008514286018908024, + 0.008640075102448463, + 0.007330979686230421, + 0.010038872249424458, + 0.010378437116742134, + 0.006994355004280806, + 0.006858296692371368, + 0.012168372981250286, + 0.012682615779340267, + 0.006870551034808159, + 0.009111369028687477, + 0.01024647243320942, + 0.017858009785413742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.005246798042207956 + } + ], + "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/0000025/group_15/index_tensor([5540, 5540])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_15/index_tensor([5540, 5540])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9f17ddfb6165e1c433ccf9e445309d52f2826b97 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_15/index_tensor([5540, 5540])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In vibrant Disney Pixar style, a medium shot, at eye level, remains momentarily static, focusing on a remarkably fluffy ginger cat crouching low on a polished hardwood floor. With thick, soft orange fur punctuated by subtle, light-colored tabby stripes, the cat is a ball of energy. Its large, shimmering emerald eyes, pupils dilated in pure playful mischief, stare at a tall scratching post studded with scissors. Its body tenses for the attack, whiskers jutting forward, the tip of its bushy tail twitching almost imperceptibly. The scene is bathed in warm, soft sunlight, streaming through an unseen window, creating a gentle glow on the cat's fur and casting soft shadows against the cozy, out-of-focus living room backdrop. Just as it raises its front paws, its pearly white claws swiftly and decisively emerging from the soft pads, the camera executes a sudden, sharp jerk, perfectly mimicking the explosive forward thrust of an attack, capturing the peak moment of vibrant, effortless chaos.", + "seed": 8981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_15/index_5540_seed_8981_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": -0.20130316130816936, + "vlm_logprob": -0.20130316130816936, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.20130316130816936, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08760502934455872 + ], + "opposite_logprob_tokens": [ + -2.8376049995422363 + ], + "logprob_reward": -0.08760502934455872, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08760502934455872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28418248891830444, + 0.10789228975772858, + 0.07139147073030472, + 0.04444548487663269, + 0.031759683042764664, + 0.027395352721214294, + 0.15866108238697052, + 0.0705447643995285, + 0.050019849091768265, + 0.15370753407478333 + ], + "credit_map": [ + 0.05485573783516884, + 0.027864966541528702, + 0.030605824664235115, + 0.033577073365449905, + 0.03981264308094978, + 0.028416836634278297, + 0.03732600435614586, + 0.031723398715257645, + 0.015654118731617928, + 0.011375379748642445, + 0.01686607487499714, + 0.01580970175564289, + 0.013917461037635803, + 0.010329882614314556, + 0.012807310558855534, + 0.011132356710731983, + 0.009926254861056805, + 0.0058672381564974785, + 0.00860543642193079, + 0.011172113940119743, + 0.009611991234123707, + 0.006844148971140385, + 0.009097316302359104, + 0.010266969911754131, + 0.006225679535418749, + 0.004589700140058994, + 0.005474220495671034, + 0.008553212508559227, + 0.004587491042912006, + 0.003936055116355419, + 0.004626493901014328, + 0.006452633533626795, + 0.004387125372886658, + 0.003416972467675805, + 0.0034651795867830515, + 0.00691838888451457, + 0.0037469675298780203, + 0.0037751593627035618, + 0.0031312215141952038, + 0.0029186666943132877, + 0.0030377483926713467, + 0.0032859467901289463, + 0.0032466480042785406, + 0.004947388544678688, + 0.00277395430020988, + 0.0028805409092456102, + 0.0032701531890779734, + 0.003952971659600735, + 0.03117191232740879, + 0.017327312380075455, + 0.018435170873999596, + 0.027031971141695976, + 0.015740061178803444, + 0.01662447117269039, + 0.015042147599160671, + 0.017288044095039368, + 0.006982548628002405, + 0.010355882346630096, + 0.008265391923487186, + 0.00978209637105465, + 0.013688392005860806, + 0.007361369673162699, + 0.007907248102128506, + 0.006201831623911858, + 0.006489058490842581, + 0.004758785013109446, + 0.0060366480611264706, + 0.005995888262987137, + 0.006824163720011711, + 0.009872334077954292, + 0.005567543674260378, + 0.004475430119782686, + 0.011728212237358093, + 0.024552740156650543, + 0.010849347338080406, + 0.020516617223620415, + 0.026493389159440994, + 0.02189824916422367, + 0.01753702014684677, + 0.020131953060626984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.08760502934455872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559603899717331 + ], + "opposite_logprob_tokens": [ + -4.387559413909912 + ], + "logprob_reward": -0.012559603899717331, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559603899717331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0678827166557312, + 0.051913246512413025, + 0.035450343042612076, + 0.04769604653120041, + 0.06301411986351013, + 0.05937936529517174, + 0.17379331588745117, + 0.06066703796386719, + 0.08597840368747711, + 0.3542254567146301 + ], + "credit_map": [ + 0.007296811323612928, + 0.006671195849776268, + 0.007980632595717907, + 0.01169437076896429, + 0.010049019940197468, + 0.0092000812292099, + 0.009013649076223373, + 0.0059769535437226295, + 0.00679192878305912, + 0.0059784455224871635, + 0.009079664014279842, + 0.010736501775681973, + 0.004970054607838392, + 0.004039447754621506, + 0.006818860303610563, + 0.0034983395598828793, + 0.0038579681422561407, + 0.003110310761258006, + 0.003686658339574933, + 0.006172073073685169, + 0.005414107348769903, + 0.004237406421452761, + 0.0048794192261993885, + 0.004092393442988396, + 0.01018044725060463, + 0.004105999134480953, + 0.005720728076994419, + 0.0050888219848275185, + 0.004303953144699335, + 0.006674451287835836, + 0.004392111673951149, + 0.007229525595903397, + 0.008307692594826221, + 0.0070213996805250645, + 0.007877619937062263, + 0.01066065113991499, + 0.008169545792043209, + 0.009131224825978279, + 0.0061648692935705185, + 0.005681104492396116, + 0.008512286469340324, + 0.006354240700602531, + 0.009218606166541576, + 0.010394085198640823, + 0.0075819347985088825, + 0.005369945429265499, + 0.0064088464714586735, + 0.005539413075894117, + 0.050948601216077805, + 0.017831958830356598, + 0.020482109859585762, + 0.019601425155997276, + 0.014361769892275333, + 0.018321720883250237, + 0.014600240625441074, + 0.017645468935370445, + 0.011653546243906021, + 0.009670313447713852, + 0.0057218316942453384, + 0.00787895917892456, + 0.008825492113828659, + 0.005071257706731558, + 0.007791026029735804, + 0.00405461061745882, + 0.007376613561064005, + 0.008790060877799988, + 0.014944137074053288, + 0.010164795443415642, + 0.011634483933448792, + 0.014936663210391998, + 0.010677308775484562, + 0.007454330567270517, + 0.029047252610325813, + 0.0443786084651947, + 0.026460908353328705, + 0.04318520054221153, + 0.06328018009662628, + 0.052132755517959595, + 0.04027998819947243, + 0.055460523813962936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.012559603899717331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818566583096981 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818566583096981, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818566583096981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2719862163066864, + 0.14866410195827484, + 0.08506737649440765, + 0.043967340141534805, + 0.017762789502739906, + 0.017937244847416878, + 0.13667909801006317, + 0.04346071183681488, + 0.090803362429142, + 0.14367173612117767 + ], + "credit_map": [ + 0.03542802482843399, + 0.03068644553422928, + 0.03489169478416443, + 0.03961150348186493, + 0.038644228130578995, + 0.02300625666975975, + 0.03978532552719116, + 0.029932767152786255, + 0.02177613601088524, + 0.01505481731146574, + 0.02090880088508129, + 0.02302425540983677, + 0.018421189859509468, + 0.014066096395254135, + 0.01879703626036644, + 0.01661577820777893, + 0.013853463344275951, + 0.007846200838685036, + 0.008747853338718414, + 0.014177152886986732, + 0.012651711702346802, + 0.007404908537864685, + 0.009834567084908485, + 0.010551519691944122, + 0.005781599320471287, + 0.004275698214769363, + 0.006752363871783018, + 0.005930776707828045, + 0.005082831718027592, + 0.006111386232078075, + 0.00455491803586483, + 0.005477764643728733, + 0.0014741651248186827, + 0.0023170060012489557, + 0.0022771451622247696, + 0.0036416358780115843, + 0.002373698865994811, + 0.0020858217030763626, + 0.0020340250339359045, + 0.0015592931304126978, + 0.0014204108156263828, + 0.002351077040657401, + 0.0022479898761957884, + 0.0028021459002047777, + 0.0021615386940538883, + 0.002700141631066799, + 0.0021610476542264223, + 0.0020928934682160616, + 0.033654943108558655, + 0.014411772601306438, + 0.01578257791697979, + 0.024156618863344193, + 0.01136090885847807, + 0.015049133449792862, + 0.011668107472360134, + 0.010595019906759262, + 0.007285760249942541, + 0.007818312384188175, + 0.004187354352325201, + 0.0050977482460439205, + 0.006812628824263811, + 0.004100795369595289, + 0.004744025878608227, + 0.0034140897914767265, + 0.007768533658236265, + 0.009604651480913162, + 0.016127223148941994, + 0.00808685552328825, + 0.0162145234644413, + 0.015453210100531578, + 0.009925534948706627, + 0.00762283056974411, + 0.010316872969269753, + 0.01338494848459959, + 0.01935330219566822, + 0.0176240187138319, + 0.02384105511009693, + 0.02163236401975155, + 0.02393811009824276, + 0.013581063598394394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.01818566583096981 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8261210322380066 + ], + "opposite_logprob_tokens": [ + -0.5761210322380066 + ], + "logprob_reward": -0.8261210322380066, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261210322380066, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15043607354164124, + 0.04366840049624443, + 0.09368429332971573, + 0.1510542631149292, + 0.16457761824131012, + 0.08679910004138947, + 0.07381211221218109, + 0.12000233680009842, + 0.05728136748075485, + 0.05868437513709068 + ], + "credit_map": [ + 0.020744584500789642, + 0.018748607486486435, + 0.01667066663503647, + 0.020078465342521667, + 0.019315045326948166, + 0.02309410646557808, + 0.018472876399755478, + 0.013311728835105896, + 0.005793742835521698, + 0.004361702129244804, + 0.007261461112648249, + 0.007112692575901747, + 0.005032081622630358, + 0.006425604689866304, + 0.004000047687441111, + 0.0036810722667723894, + 0.016081279143691063, + 0.007755526341497898, + 0.009394754655659199, + 0.016177387908101082, + 0.011825776658952236, + 0.009764115326106548, + 0.008864330127835274, + 0.013821135275065899, + 0.02904931642115116, + 0.014385070651769638, + 0.020727060735225677, + 0.019176162779331207, + 0.015787776559591293, + 0.022262781858444214, + 0.011222530156373978, + 0.018443576991558075, + 0.023506315425038338, + 0.017074497416615486, + 0.02169756218791008, + 0.036457546055316925, + 0.01616639643907547, + 0.014936728402972221, + 0.01671740971505642, + 0.018021194264292717, + 0.013036788441240788, + 0.009307900443673134, + 0.010158255696296692, + 0.014743484556674957, + 0.010758419521152973, + 0.010385665111243725, + 0.009160022251307964, + 0.009248577058315277, + 0.01813557557761669, + 0.0071916584856808186, + 0.007717917673289776, + 0.010000044479966164, + 0.006550199817866087, + 0.008101711049675941, + 0.007526251021772623, + 0.008588766679167747, + 0.02293977700173855, + 0.014525093138217926, + 0.012658978812396526, + 0.020082706585526466, + 0.01686723902821541, + 0.01058272086083889, + 0.013804024085402489, + 0.008541813120245934, + 0.006595660001039505, + 0.007396164815872908, + 0.007693165447562933, + 0.006384451873600483, + 0.00805546436458826, + 0.008595730178058147, + 0.006917545571923256, + 0.0056431954726576805, + 0.005244632717221975, + 0.00695409532636404, + 0.006498988717794418, + 0.010514065623283386, + 0.007679566275328398, + 0.00784061849117279, + 0.00682728411629796, + 0.007125134114176035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.8261210322380066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062044475227594376 + ], + "opposite_logprob_tokens": [ + -2.812044382095337 + ], + "logprob_reward": -0.062044475227594376, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062044475227594376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04554812237620354, + 0.018170086666941643, + 0.01430252194404602, + 0.018549172207713127, + 0.032561857253313065, + 0.01329873688519001, + 0.011115598492324352, + 0.03634442761540413, + 0.060390405356884, + 0.7497190833091736 + ], + "credit_map": [ + 0.00952847022563219, + 0.004559486173093319, + 0.005561904516071081, + 0.004847723990678787, + 0.006010795012116432, + 0.006906475871801376, + 0.004159566946327686, + 0.003973697777837515, + 0.00128015480004251, + 0.002075460972264409, + 0.002195279812440276, + 0.003494665026664734, + 0.0022500192280858755, + 0.0014808689011260867, + 0.0031148854177445173, + 0.0022787500638514757, + 0.0024316038470715284, + 0.0012790750479325652, + 0.001994177233427763, + 0.0018819866236299276, + 0.0018554164562374353, + 0.002088097622618079, + 0.0011388466227799654, + 0.0016333182575181127, + 0.0025211332831531763, + 0.002037595259025693, + 0.0020002240780740976, + 0.0023579157423228025, + 0.002407230669632554, + 0.002357626333832741, + 0.0026698133442550898, + 0.002197632333263755, + 0.005664539057761431, + 0.003254011506214738, + 0.003864545375108719, + 0.00463290698826313, + 0.003944504540413618, + 0.004902320448309183, + 0.002821984002366662, + 0.003477043006569147, + 0.001961304107680917, + 0.0015790612669661641, + 0.0013353402027860284, + 0.002181132324039936, + 0.0017984273144975305, + 0.0016019988106563687, + 0.0013558181235566735, + 0.00148565333802253, + 0.003230068599805236, + 0.0009629444102756679, + 0.0011988006299361587, + 0.0014899208908900619, + 0.0013947152765467763, + 0.0011474357452243567, + 0.0008006247808225453, + 0.0008910867618396878, + 0.0039693801663815975, + 0.004678782075643539, + 0.0034248761367052794, + 0.004437339957803488, + 0.007764112204313278, + 0.004738980438560247, + 0.004506302997469902, + 0.0028246508445590734, + 0.003973030019551516, + 0.005393469240516424, + 0.006901324260979891, + 0.007625887636095285, + 0.009331557899713516, + 0.015481785871088505, + 0.005719146225601435, + 0.00596420094370842, + 0.06060533970594406, + 0.0959656685590744, + 0.056464727967977524, + 0.09730485081672668, + 0.13159720599651337, + 0.11193452775478363, + 0.07321833819150925, + 0.12262837588787079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.062044475227594376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In vibrant Disney Pixar style, a medium shot, at eye level, remains momentarily static, focusing on a remarkably fluffy ginger cat crouching low on a polished hardwood floor. With thick, soft orange fur punctuated by subtle, light-colored tabby stripes, the cat is a ball of energy. Its large, shimmering emerald eyes, pupils dilated in pure playful mischief, stare at a tall scratching post studded with scissors. Its body tenses for the attack, whiskers jutting forward, the tip of its bushy tail twitching almost imperceptibly. The scene is bathed in warm, soft sunlight, streaming through an unseen window, creating a gentle glow on the cat's fur and casting soft shadows against the cozy, out-of-focus living room backdrop. Just as it raises its front paws, its pearly white claws swiftly and decisively emerging from the soft pads, the camera executes a sudden, sharp jerk, perfectly mimicking the explosive forward thrust of an attack, capturing the peak moment of vibrant, effortless chaos.", + "seed": 8982, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_15/index_5540_seed_8982_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": -0.23984724543988706, + "vlm_logprob": -0.23984724543988706, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.23984724543988706, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11569506675004959 + ], + "opposite_logprob_tokens": [ + -2.365694999694824 + ], + "logprob_reward": -0.11569506675004959, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11569506675004959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19434812664985657, + 0.16008812189102173, + 0.03898357227444649, + 0.031110301613807678, + 0.21839462220668793, + 0.07028572261333466, + 0.11205112934112549, + 0.057039495557546616, + 0.05354902893304825, + 0.06414982676506042 + ], + "credit_map": [ + 0.018584763631224632, + 0.031005918979644775, + 0.024315139278769493, + 0.03033764846622944, + 0.030816230922937393, + 0.014863637275993824, + 0.029109690338373184, + 0.015315113589167595, + 0.013062617741525173, + 0.019542021676898003, + 0.02140628732740879, + 0.02223806083202362, + 0.02164357714354992, + 0.01436386164277792, + 0.030347082763910294, + 0.01748460717499256, + 0.002752425381913781, + 0.0045534903183579445, + 0.004565673880279064, + 0.006179586984217167, + 0.004681805614382029, + 0.003918924368917942, + 0.004635232500731945, + 0.007696433924138546, + 0.0029641950968652964, + 0.004423517268151045, + 0.0034208688884973526, + 0.007086834870278835, + 0.003954080864787102, + 0.0021541512105613947, + 0.003823370672762394, + 0.003283286467194557, + 0.015535786747932434, + 0.010708481073379517, + 0.01577073521912098, + 0.0815512016415596, + 0.006835691165179014, + 0.009665791876614094, + 0.024503454566001892, + 0.05382348969578743, + 0.004679486155509949, + 0.00431074108928442, + 0.0047287773340940475, + 0.02591424249112606, + 0.002642167964950204, + 0.003094619372859597, + 0.00834602490067482, + 0.016569659113883972, + 0.00800689309835434, + 0.007503831293433905, + 0.01211236510425806, + 0.038480956107378006, + 0.005920261610299349, + 0.00903903879225254, + 0.011865693144500256, + 0.019122092053294182, + 0.004871940240263939, + 0.007111102342605591, + 0.009455605410039425, + 0.008891230449080467, + 0.007092984858900309, + 0.007761629763990641, + 0.005894500762224197, + 0.0059605068527162075, + 0.004921793472021818, + 0.006591863930225372, + 0.007096016313880682, + 0.013605204410851002, + 0.007848860695958138, + 0.002777646528556943, + 0.006491981912404299, + 0.004215663764625788, + 0.004255172796547413, + 0.014957623556256294, + 0.005142581183463335, + 0.008397380821406841, + 0.008530430495738983, + 0.008943874388933182, + 0.008040028624236584, + 0.005882735829800367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.11569506675004959 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603911951184273 + ], + "opposite_logprob_tokens": [ + -3.895603895187378 + ], + "logprob_reward": -0.020603911951184273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603911951184273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12267067283391953, + 0.02863527461886406, + 0.02158319763839245, + 0.016716545447707176, + 0.3139827251434326, + 0.023410409688949585, + 0.21444983780384064, + 0.08602264523506165, + 0.10242976993322372, + 0.07009893655776978 + ], + "credit_map": [ + 0.00897163338959217, + 0.01287032850086689, + 0.0098768575116992, + 0.03279523923993111, + 0.013502833433449268, + 0.010681592859327793, + 0.014899133704602718, + 0.01907305233180523, + 0.002249178010970354, + 0.004113059490919113, + 0.002449645195156336, + 0.0053889574483036995, + 0.003917077090591192, + 0.002440043957903981, + 0.005052519030869007, + 0.0030247934628278017, + 0.0017249081283807755, + 0.0014863602118566632, + 0.0032000818755477667, + 0.005028455518186092, + 0.0014625819167122245, + 0.0016541125951334834, + 0.002581969602033496, + 0.0044447253458201885, + 0.0017030887538567185, + 0.0019200857495889068, + 0.002584070898592472, + 0.002248205477371812, + 0.001790252747014165, + 0.0021605785004794598, + 0.002283163834363222, + 0.0020270980894565582, + 0.022069476544857025, + 0.015481706708669662, + 0.026036549359560013, + 0.11948822438716888, + 0.009630296379327774, + 0.013049342669546604, + 0.03575517609715462, + 0.07247194647789001, + 0.0032356909941881895, + 0.0023218984715640545, + 0.003038609866052866, + 0.0030077763367444277, + 0.0023619304411113262, + 0.0037276200018823147, + 0.0030723465606570244, + 0.0026445353869348764, + 0.018068624660372734, + 0.01637709140777588, + 0.030730916187167168, + 0.05866904929280281, + 0.017420919612050056, + 0.027340179309248924, + 0.021596435457468033, + 0.024246610701084137, + 0.006272129248827696, + 0.010040379129350185, + 0.012123893946409225, + 0.024014988914132118, + 0.008030942641198635, + 0.007519808132201433, + 0.00957564264535904, + 0.00844485405832529, + 0.00913045834749937, + 0.014426798559725285, + 0.011901646852493286, + 0.028449390083551407, + 0.01347584929317236, + 0.0056786490604281425, + 0.012747730128467083, + 0.006619246676564217, + 0.004410325549542904, + 0.014498038217425346, + 0.005672171711921692, + 0.008892456069588661, + 0.007884690538048744, + 0.01282756682485342, + 0.009561331942677498, + 0.006352353375405073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.020603911951184273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369533270597458 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.03369533270597458, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369533270597458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12894992530345917, + 0.11057201772928238, + 0.05933726951479912, + 0.04649514704942703, + 0.15683110058307648, + 0.04973756894469261, + 0.3299444615840912, + 0.03269954398274422, + 0.0326676219701767, + 0.05276532098650932 + ], + "credit_map": [ + 0.010700693354010582, + 0.015247998759150505, + 0.012367184273898602, + 0.03659456968307495, + 0.011591755785048008, + 0.009180814027786255, + 0.012531755492091179, + 0.02073514647781849, + 0.009303731843829155, + 0.015593822114169598, + 0.015112058259546757, + 0.015548614785075188, + 0.015425987541675568, + 0.008489751257002354, + 0.02026285044848919, + 0.010835196822881699, + 0.004883151967078447, + 0.006480199284851551, + 0.007095434702932835, + 0.00971014704555273, + 0.006325217429548502, + 0.007251006551086903, + 0.006548443343490362, + 0.011043667793273926, + 0.004651874769479036, + 0.006368632894009352, + 0.006277601700276136, + 0.0076005421578884125, + 0.005591817200183868, + 0.0039754463359713554, + 0.007598613388836384, + 0.0044306181371212006, + 0.010672157630324364, + 0.0076394472271203995, + 0.014505408704280853, + 0.060258131474256516, + 0.004916470963507891, + 0.006351717282086611, + 0.018069660291075706, + 0.03441811725497246, + 0.003838790813460946, + 0.003341652685776353, + 0.005242204759269953, + 0.014848722144961357, + 0.002054027048870921, + 0.0019717784598469734, + 0.004824045114219189, + 0.01361634861677885, + 0.02514989674091339, + 0.022740988060832024, + 0.036470960825681686, + 0.10625746846199036, + 0.019814899191260338, + 0.032147638499736786, + 0.034849826246500015, + 0.052512768656015396, + 0.0026597946416586637, + 0.0031732614152133465, + 0.0063780369237065315, + 0.005044359248131514, + 0.0038014384917914867, + 0.004610016942024231, + 0.0030460550915449858, + 0.003986576572060585, + 0.0028323731385171413, + 0.004538835026323795, + 0.0028777243569493294, + 0.008030032739043236, + 0.005585265811532736, + 0.0020786758977919817, + 0.004383399151265621, + 0.002341315383091569, + 0.0031790619250386953, + 0.008736269548535347, + 0.0030743489041924477, + 0.007099023554474115, + 0.007916796021163464, + 0.010530578903853893, + 0.007085402030497789, + 0.005143840331584215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.03369533270597458 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9742737412452698 + ], + "opposite_logprob_tokens": [ + -0.4742737412452698 + ], + "logprob_reward": -0.9742737412452698, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742737412452698, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20483610033988953, + 0.13575193285942078, + 0.04279661178588867, + 0.052662428468465805, + 0.05857212841510773, + 0.19416803121566772, + 0.054623935371637344, + 0.052906762808561325, + 0.09151533991098404, + 0.11216668784618378 + ], + "credit_map": [ + 0.01867532730102539, + 0.03982709348201752, + 0.02062145620584488, + 0.033718328922986984, + 0.030463196337223053, + 0.014767928048968315, + 0.022360756993293762, + 0.024402014911174774, + 0.011924211867153645, + 0.017825281247496605, + 0.018858157098293304, + 0.0182833019644022, + 0.01836966909468174, + 0.009796591475605965, + 0.025574028491973877, + 0.015120698139071465, + 0.003711951430886984, + 0.003945330157876015, + 0.005046797916293144, + 0.010950430296361446, + 0.0031478977762162685, + 0.003209067741408944, + 0.005650809500366449, + 0.007134332787245512, + 0.004531704820692539, + 0.006332676857709885, + 0.006900065578520298, + 0.00860752072185278, + 0.004632918629795313, + 0.005726257339119911, + 0.005410728044807911, + 0.010520562529563904, + 0.0061195408925414085, + 0.006538746412843466, + 0.00919574499130249, + 0.010125314816832542, + 0.004990026354789734, + 0.0046602170914411545, + 0.006945353467017412, + 0.00999719463288784, + 0.013789158314466476, + 0.01210719347000122, + 0.011968418955802917, + 0.07025861740112305, + 0.00599270686507225, + 0.008411100134253502, + 0.022731183096766472, + 0.04890967532992363, + 0.003927250392735004, + 0.004104722756892443, + 0.00815082062035799, + 0.01835308037698269, + 0.00259740324690938, + 0.0031044522766023874, + 0.006721941288560629, + 0.007664268370717764, + 0.004838075954467058, + 0.004600669257342815, + 0.00984902586787939, + 0.015028001740574837, + 0.003622326534241438, + 0.004641208332031965, + 0.005648488644510508, + 0.004678966477513313, + 0.00883918721228838, + 0.010264761745929718, + 0.009642798453569412, + 0.030604438856244087, + 0.009155779145658016, + 0.00483354227617383, + 0.011896328069269657, + 0.006278504617512226, + 0.008436254225671291, + 0.021486690267920494, + 0.010400851257145405, + 0.013919730670750141, + 0.016219619661569595, + 0.015230963937938213, + 0.015236740931868553, + 0.011235848069190979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.9742737412452698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054968174546957016 + ], + "opposite_logprob_tokens": [ + -2.9299681186676025 + ], + "logprob_reward": -0.054968174546957016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054968174546957016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07328443974256516, + 0.029728030785918236, + 0.030254554003477097, + 0.020361563190817833, + 0.035180721431970596, + 0.13820107281208038, + 0.07363931834697723, + 0.053873855620622635, + 0.06302978098392487, + 0.48244667053222656 + ], + "credit_map": [ + 0.006134055554866791, + 0.010979541577398777, + 0.008730505593121052, + 0.012084785848855972, + 0.010314451530575752, + 0.007730739191174507, + 0.009977588430047035, + 0.007332770153880119, + 0.0030083446763455868, + 0.003329805564135313, + 0.003201030660420656, + 0.006541242357343435, + 0.003058245638385415, + 0.0022761018481105566, + 0.004214011132717133, + 0.004099250305444002, + 0.002311023185029626, + 0.0027973726391792297, + 0.004264329094439745, + 0.0065539819188416, + 0.003180028637871146, + 0.003180587897077203, + 0.0040853205136954784, + 0.003881912212818861, + 0.0016084186499938369, + 0.002855045022442937, + 0.002543351612985134, + 0.0030846695881336927, + 0.0030063760932534933, + 0.001970049925148487, + 0.0026646233163774014, + 0.0026290290988981724, + 0.002314159879460931, + 0.002539779059588909, + 0.006209223531186581, + 0.012476612813770771, + 0.0019247994059696794, + 0.0012983052292838693, + 0.0037786774337291718, + 0.004639161750674248, + 0.009454354643821716, + 0.00981434341520071, + 0.008735270239412785, + 0.047143418341875076, + 0.004882724955677986, + 0.007663005962967873, + 0.015115756541490555, + 0.03539219871163368, + 0.0054802256636321545, + 0.006020186003297567, + 0.009914305992424488, + 0.0193600095808506, + 0.002809088444337249, + 0.003486738307401538, + 0.006846469361335039, + 0.01972230151295662, + 0.004346002358943224, + 0.004448669496923685, + 0.013187125325202942, + 0.007100663613528013, + 0.002697302959859371, + 0.0026306277140975, + 0.004816899076104164, + 0.014646563678979874, + 0.005920676980167627, + 0.008482208475470543, + 0.006791246589273214, + 0.015302682295441628, + 0.010149406269192696, + 0.0042387954890728, + 0.008023710921406746, + 0.004121057223528624, + 0.02773553505539894, + 0.09849929064512253, + 0.03452629968523979, + 0.05637481436133385, + 0.05773364380002022, + 0.11107076704502106, + 0.05602307990193367, + 0.0404832549393177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.054968174546957016 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In vibrant Disney Pixar style, a medium shot, at eye level, remains momentarily static, focusing on a remarkably fluffy ginger cat crouching low on a polished hardwood floor. With thick, soft orange fur punctuated by subtle, light-colored tabby stripes, the cat is a ball of energy. Its large, shimmering emerald eyes, pupils dilated in pure playful mischief, stare at a tall scratching post studded with scissors. Its body tenses for the attack, whiskers jutting forward, the tip of its bushy tail twitching almost imperceptibly. The scene is bathed in warm, soft sunlight, streaming through an unseen window, creating a gentle glow on the cat's fur and casting soft shadows against the cozy, out-of-focus living room backdrop. Just as it raises its front paws, its pearly white claws swiftly and decisively emerging from the soft pads, the camera executes a sudden, sharp jerk, perfectly mimicking the explosive forward thrust of an attack, capturing the peak moment of vibrant, effortless chaos.", + "seed": 8983, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_15/index_5540_seed_8983_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": -0.16558345444500447, + "vlm_logprob": -0.16558345444500447, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.16558345444500447, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13238829374313354 + ], + "opposite_logprob_tokens": [ + -2.2573883533477783 + ], + "logprob_reward": -0.13238829374313354, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13238829374313354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30765843391418457, + 0.0826968103647232, + 0.04952559620141983, + 0.05681607499718666, + 0.021111425012350082, + 0.05516551062464714, + 0.09739092737436295, + 0.09552774578332901, + 0.08289128541946411, + 0.15121622383594513 + ], + "credit_map": [ + 0.032761070877313614, + 0.020351767539978027, + 0.06278175115585327, + 0.027266399934887886, + 0.03730195760726929, + 0.06786458194255829, + 0.03097175993025303, + 0.028359143063426018, + 0.010677535086870193, + 0.005129546392709017, + 0.020559538155794144, + 0.008517839014530182, + 0.011462783440947533, + 0.008403337560594082, + 0.00942275207489729, + 0.008523485623300076, + 0.00778615428134799, + 0.0036079802084714174, + 0.007855776697397232, + 0.006788310594856739, + 0.007469639182090759, + 0.003950059879571199, + 0.007658611051738262, + 0.004409063141793013, + 0.010848100297152996, + 0.004114362411201, + 0.017090478911995888, + 0.0052005876787006855, + 0.004884764086455107, + 0.004591427743434906, + 0.004361339379101992, + 0.005725016351789236, + 0.002593704964965582, + 0.0019011393887922168, + 0.004200572147965431, + 0.002797333989292383, + 0.0022533880546689034, + 0.001457656268030405, + 0.0030410541221499443, + 0.002866575960069895, + 0.007243295665830374, + 0.004235438071191311, + 0.0074888113886117935, + 0.0075392224825918674, + 0.0052604940719902515, + 0.005359579809010029, + 0.007691947277635336, + 0.010346719063818455, + 0.010683035477995872, + 0.007487979717552662, + 0.011510987766087055, + 0.011121217161417007, + 0.020295407623052597, + 0.006389050744473934, + 0.017355453222990036, + 0.012547794729471207, + 0.0090284189209342, + 0.008314244449138641, + 0.015355907380580902, + 0.008428280241787434, + 0.015394406393170357, + 0.011280873790383339, + 0.021128468215465546, + 0.006597151048481464, + 0.006817888934165239, + 0.010157285258173943, + 0.008133243769407272, + 0.009180180728435516, + 0.0103873610496521, + 0.013946415856480598, + 0.015920614823698997, + 0.008348298259079456, + 0.014591091312468052, + 0.015173094347119331, + 0.01179959625005722, + 0.013710026629269123, + 0.03967912122607231, + 0.03946305438876152, + 0.008375419303774834, + 0.008424809202551842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.13238829374313354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009808912873268127 + ], + "opposite_logprob_tokens": [ + -4.6348090171813965 + ], + "logprob_reward": -0.009808912873268127, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009808912873268127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1883094161748886, + 0.05817444995045662, + 0.015295695513486862, + 0.02679535560309887, + 0.01935552805662155, + 0.024853825569152832, + 0.09438437223434448, + 0.2745472490787506, + 0.1746651977300644, + 0.12361893802881241 + ], + "credit_map": [ + 0.031028077006340027, + 0.010165434330701828, + 0.036140598356723785, + 0.013547452166676521, + 0.011651415377855301, + 0.0511440634727478, + 0.015923988074064255, + 0.018708381801843643, + 0.008787923492491245, + 0.0028463578782975674, + 0.020748060196638107, + 0.0043532270938158035, + 0.003737696446478367, + 0.006761152297258377, + 0.0058163744397461414, + 0.00512365996837616, + 0.0025286891032010317, + 0.0009330434841103852, + 0.0032539062667638063, + 0.0014039548113942146, + 0.001649838057346642, + 0.0019075822783634067, + 0.0015085040358826518, + 0.002110178116708994, + 0.004734336864203215, + 0.001977072563022375, + 0.00509294681251049, + 0.003115086117759347, + 0.0026489344891160727, + 0.002347309375181794, + 0.00317968288436532, + 0.0036999869626015425, + 0.0027052771765738726, + 0.001554720220156014, + 0.002309078350663185, + 0.0029487181454896927, + 0.0025546709075570107, + 0.0015791263431310654, + 0.0030482173897325993, + 0.002655718708410859, + 0.0037232558242976665, + 0.0015038085402920842, + 0.006347374524921179, + 0.002398107433691621, + 0.002184223849326372, + 0.0018688867567107081, + 0.002944311825558543, + 0.003883856814354658, + 0.007683498784899712, + 0.011402563191950321, + 0.01670032925903797, + 0.013843239285051823, + 0.013747100718319416, + 0.005960086360573769, + 0.014472251757979393, + 0.010575303807854652, + 0.017052095383405685, + 0.01538161002099514, + 0.026710672304034233, + 0.022904464974999428, + 0.03564861789345741, + 0.07506939023733139, + 0.06474079936742783, + 0.0170395877212286, + 0.01180973555892706, + 0.01256338506937027, + 0.017692333087325096, + 0.022170118987560272, + 0.020036816596984863, + 0.01844015158712864, + 0.053213994950056076, + 0.018738675862550735, + 0.013280575163662434, + 0.009029273875057697, + 0.008819197304546833, + 0.011866943910717964, + 0.03384911268949509, + 0.03303501009941101, + 0.008094385266304016, + 0.005644436925649643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.009808912873268127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581373944878578 + ], + "opposite_logprob_tokens": [ + -3.8955814838409424 + ], + "logprob_reward": -0.020581373944878578, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581373944878578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1784263253211975, + 0.019118349999189377, + 0.020146293565630913, + 0.019557233899831772, + 0.014003229327499866, + 0.01836097426712513, + 0.15112969279289246, + 0.42235061526298523, + 0.07187008857727051, + 0.08503720909357071 + ], + "credit_map": [ + 0.026497656479477882, + 0.01035555824637413, + 0.033844150602817535, + 0.011742917820811272, + 0.009731476195156574, + 0.055029865354299545, + 0.014506489969789982, + 0.016718195751309395, + 0.0026990221813321114, + 0.0014075618237257004, + 0.0037574004381895065, + 0.0025035953149199486, + 0.0024513909593224525, + 0.0019952936563640833, + 0.0025501966010779142, + 0.001753887627273798, + 0.0028650201857089996, + 0.0013476999010890722, + 0.005735703278332949, + 0.0020625146571546793, + 0.0021490335930138826, + 0.0015864423476159573, + 0.0022215996868908405, + 0.0021782806143164635, + 0.0034051232505589724, + 0.0016320649301633239, + 0.004841297399252653, + 0.0024343407712876797, + 0.002180813578888774, + 0.0012743464903905988, + 0.0018580168252810836, + 0.0019312311196699739, + 0.0022072677966207266, + 0.001357403234578669, + 0.00346830440685153, + 0.001529539586044848, + 0.0014814953319728374, + 0.0010065182577818632, + 0.0013642683625221252, + 0.0015884310705587268, + 0.0017498526722192764, + 0.001225048559717834, + 0.004372118506580591, + 0.0023907709401100874, + 0.0021721920929849148, + 0.001425197347998619, + 0.0028028993401676416, + 0.0022228960879147053, + 0.013733962550759315, + 0.01367108803242445, + 0.03877563029527664, + 0.02001206763088703, + 0.02073242887854576, + 0.008288456127047539, + 0.020629791542887688, + 0.0152862798422575, + 0.0246968325227499, + 0.03204791992902756, + 0.0633893683552742, + 0.04555179178714752, + 0.0642339363694191, + 0.10040433704853058, + 0.06288810819387436, + 0.029138337820768356, + 0.005746439564973116, + 0.009195797145366669, + 0.008246899582445621, + 0.008346136659383774, + 0.008505352772772312, + 0.012562189251184464, + 0.0120243476703763, + 0.007242924068123102, + 0.011869166046380997, + 0.008278250694274902, + 0.007675471715629101, + 0.008922191336750984, + 0.015112999826669693, + 0.013444951735436916, + 0.012073068879544735, + 0.0076611097902059555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.020581373944878578 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761134028434753 + ], + "opposite_logprob_tokens": [ + -0.8261134028434753 + ], + "logprob_reward": -0.5761134028434753, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761134028434753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1588047593832016, + 0.03493787348270416, + 0.02252844348549843, + 0.02202129177749157, + 0.022354139015078545, + 0.015133732929825783, + 0.02446541003882885, + 0.1735471487045288, + 0.03532158583402634, + 0.4908856749534607 + ], + "credit_map": [ + 0.028699452057480812, + 0.009685028344392776, + 0.026803918182849884, + 0.01169317215681076, + 0.011042273603379726, + 0.04409989342093468, + 0.01312330737709999, + 0.0136577058583498, + 0.004745861515402794, + 0.0031002743635326624, + 0.004650040529668331, + 0.003690433455631137, + 0.004447315353900194, + 0.004838764667510986, + 0.0051108491607010365, + 0.004354334436357021, + 0.004083652980625629, + 0.0018626372329890728, + 0.003627987112849951, + 0.0021628886461257935, + 0.0025452261324971914, + 0.0029006621334701777, + 0.0023250742815434933, + 0.0030203142669051886, + 0.004488383885473013, + 0.0015443071024492383, + 0.004491937346756458, + 0.0020535504445433617, + 0.001631515915505588, + 0.0025016730651259422, + 0.0018947157077491283, + 0.003415206912904978, + 0.004328735638409853, + 0.002119927667081356, + 0.003907430917024612, + 0.002610337920486927, + 0.0021752119064331055, + 0.002337338402867317, + 0.002269549062475562, + 0.0026056079659610987, + 0.002888148883357644, + 0.001203550724312663, + 0.0022881950717419386, + 0.0014190968358889222, + 0.001323841279372573, + 0.0020344865042716265, + 0.0013483676593750715, + 0.0026280456222593784, + 0.0029080279637128115, + 0.001697319676168263, + 0.005774452351033688, + 0.00267941621132195, + 0.0028173949103802443, + 0.0027345619164407253, + 0.0025717231910675764, + 0.003282510908320546, + 0.015264267101883888, + 0.008576505817472935, + 0.021334165707230568, + 0.01488819532096386, + 0.014161595143377781, + 0.03265190124511719, + 0.0574536956846714, + 0.009216809645295143, + 0.004157950170338154, + 0.0036554366815835238, + 0.006795031949877739, + 0.004111362621188164, + 0.004405919462442398, + 0.004716335330158472, + 0.0037993709556758404, + 0.003680177265778184, + 0.046306513249874115, + 0.048214141279459, + 0.04777871444821358, + 0.03577536344528198, + 0.1344694197177887, + 0.1333456188440323, + 0.02617914043366909, + 0.01881672628223896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.5761134028434753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902528882026672 + ], + "opposite_logprob_tokens": [ + -2.4640252590179443 + ], + "logprob_reward": -0.08902528882026672, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902528882026672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05790188908576965, + 0.03826605901122093, + 0.019975729286670685, + 0.027349866926670074, + 0.02079513669013977, + 0.015319420956075191, + 0.20344620943069458, + 0.25541800260543823, + 0.1831153929233551, + 0.1784122735261917 + ], + "credit_map": [ + 0.008119182661175728, + 0.004690391011536121, + 0.008240269497036934, + 0.0055874125100672245, + 0.00881551019847393, + 0.00912478007376194, + 0.005612590815871954, + 0.0077117509208619595, + 0.003935002256184816, + 0.00205282773822546, + 0.01207951083779335, + 0.0034975905437022448, + 0.0032046386040747166, + 0.0047248718328773975, + 0.00391796138137579, + 0.004853655118495226, + 0.0024774607736617327, + 0.0012878475245088339, + 0.004469096660614014, + 0.0021594110876321793, + 0.0022513819858431816, + 0.0021084884647279978, + 0.0031672839540988207, + 0.0020547579042613506, + 0.0029430424328893423, + 0.0032358565367758274, + 0.002929295413196087, + 0.004806801211088896, + 0.0036314907483756542, + 0.0016375972190871835, + 0.004964324180036783, + 0.003201462095603347, + 0.003295156406238675, + 0.0016157726058736444, + 0.003855637274682522, + 0.0023622249718755484, + 0.002231848193332553, + 0.001580777927301824, + 0.0034212779719382524, + 0.0024324406404048204, + 0.0028555099852383137, + 0.0013360088923946023, + 0.00207281531766057, + 0.0016874833963811398, + 0.0019085545791313052, + 0.0013646831503137946, + 0.002168122446164489, + 0.0019262435380369425, + 0.018678177148103714, + 0.021803295239806175, + 0.03736143186688423, + 0.026816509664058685, + 0.03421808034181595, + 0.011510897427797318, + 0.03313889726996422, + 0.01991892419755459, + 0.015885522589087486, + 0.01766262575984001, + 0.020995359867811203, + 0.03104824759066105, + 0.03415394946932793, + 0.057364244014024734, + 0.05944795906543732, + 0.018860090523958206, + 0.012723990716040134, + 0.01983807235956192, + 0.01792459562420845, + 0.022402716800570488, + 0.022805457934737206, + 0.037101101130247116, + 0.024396944791078568, + 0.02592252753674984, + 0.011900199577212334, + 0.03565008565783501, + 0.012430468574166298, + 0.016367552801966667, + 0.025939809158444405, + 0.02831280790269375, + 0.027660733088850975, + 0.020150620490312576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.08902528882026672 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In vibrant Disney Pixar style, a medium shot, at eye level, remains momentarily static, focusing on a remarkably fluffy ginger cat crouching low on a polished hardwood floor. With thick, soft orange fur punctuated by subtle, light-colored tabby stripes, the cat is a ball of energy. Its large, shimmering emerald eyes, pupils dilated in pure playful mischief, stare at a tall scratching post studded with scissors. Its body tenses for the attack, whiskers jutting forward, the tip of its bushy tail twitching almost imperceptibly. The scene is bathed in warm, soft sunlight, streaming through an unseen window, creating a gentle glow on the cat's fur and casting soft shadows against the cozy, out-of-focus living room backdrop. Just as it raises its front paws, its pearly white claws swiftly and decisively emerging from the soft pads, the camera executes a sudden, sharp jerk, perfectly mimicking the explosive forward thrust of an attack, capturing the peak moment of vibrant, effortless chaos.", + "seed": 8984, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_15/index_5540_seed_8984_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": -0.2598386902362108, + "vlm_logprob": -0.2598386902362108, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.2598386902362108, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09665928035974503 + ], + "opposite_logprob_tokens": [ + -2.5966591835021973 + ], + "logprob_reward": -0.09665928035974503, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09665928035974503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2964993715286255, + 0.07695630192756653, + 0.05490757152438164, + 0.057837922126054764, + 0.048334214836359024, + 0.06661363691091537, + 0.05629730597138405, + 0.030003225430846214, + 0.1776314675807953, + 0.13491903245449066 + ], + "credit_map": [ + 0.03054003231227398, + 0.022252686321735382, + 0.03038071282207966, + 0.03480621799826622, + 0.033370621502399445, + 0.035028472542762756, + 0.04197410121560097, + 0.06814651936292648, + 0.008617894724011421, + 0.0062703657895326614, + 0.008633351884782314, + 0.009305530227720737, + 0.008436132222414017, + 0.009073000401258469, + 0.009743118658661842, + 0.016876909881830215, + 0.006396263837814331, + 0.005021126475185156, + 0.0063135139644145966, + 0.007678802125155926, + 0.004184509161859751, + 0.005758598912507296, + 0.006687438115477562, + 0.012867319397628307, + 0.007139644119888544, + 0.005493397358804941, + 0.005411104764789343, + 0.005831824149936438, + 0.004577220417559147, + 0.005437149200588465, + 0.009400707669556141, + 0.014546875841915607, + 0.004987821914255619, + 0.0027526156045496464, + 0.003885724116116762, + 0.004095828160643578, + 0.0054174247197806835, + 0.006910756696015596, + 0.010831449180841446, + 0.009452592581510544, + 0.005057189147919416, + 0.006051936186850071, + 0.0037275836803019047, + 0.005880292970687151, + 0.005088100675493479, + 0.007285456173121929, + 0.016095604747533798, + 0.01742747612297535, + 0.004289950709789991, + 0.004536906722933054, + 0.003998677246272564, + 0.005063749849796295, + 0.005279141943901777, + 0.007674910593777895, + 0.013012039475142956, + 0.012441933155059814, + 0.002827709075063467, + 0.00295205507427454, + 0.002245009411126375, + 0.005206274334341288, + 0.00319340405985713, + 0.0026295373681932688, + 0.004745420068502426, + 0.006203815340995789, + 0.013989018276333809, + 0.02420005388557911, + 0.014019008725881577, + 0.023271802812814713, + 0.04484923556447029, + 0.01816786453127861, + 0.025749458000063896, + 0.013385026715695858, + 0.014964121393859386, + 0.024059763178229332, + 0.008980697020888329, + 0.014081773348152637, + 0.026671476662158966, + 0.026667021214962006, + 0.00902564451098442, + 0.010468541644513607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.09665928035974503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981995977461338 + ], + "opposite_logprob_tokens": [ + -3.529819965362549 + ], + "logprob_reward": -0.02981995977461338, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981995977461338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10932659357786179, + 0.0591689758002758, + 0.03110300749540329, + 0.04514305293560028, + 0.12152784317731857, + 0.0779513344168663, + 0.042838238179683685, + 0.04486757889389992, + 0.28493931889533997, + 0.18313409388065338 + ], + "credit_map": [ + 0.01521015353500843, + 0.008727235719561577, + 0.009479966945946217, + 0.012560268864035606, + 0.008763282559812069, + 0.014760678634047508, + 0.01748749054968357, + 0.022337503731250763, + 0.008243706077337265, + 0.005000824574381113, + 0.005141218658536673, + 0.0065650977194309235, + 0.004796078894287348, + 0.005237736739218235, + 0.0095546068623662, + 0.014629703015089035, + 0.004011490847915411, + 0.002575524849817157, + 0.0026954207569360733, + 0.003722868161275983, + 0.0031322252470999956, + 0.004113207571208477, + 0.00559481605887413, + 0.005257452372461557, + 0.006213457323610783, + 0.004091298673301935, + 0.003515003016218543, + 0.0044993869960308075, + 0.005203956738114357, + 0.0040135011076927185, + 0.008685961365699768, + 0.00892048142850399, + 0.011662231758236885, + 0.011080959811806679, + 0.007371912244707346, + 0.015659255906939507, + 0.010905832052230835, + 0.010783352889120579, + 0.033904630690813065, + 0.020159658044576645, + 0.005990548525005579, + 0.008044908754527569, + 0.00609626853838563, + 0.009443203918635845, + 0.008126351051032543, + 0.012051048688590527, + 0.015545065514743328, + 0.012653929181396961, + 0.0056806872598826885, + 0.003277194919064641, + 0.004202937241643667, + 0.00570380175486207, + 0.0045829336158931255, + 0.004832657054066658, + 0.005264467094093561, + 0.00929355714470148, + 0.004236625507473946, + 0.0055564516223967075, + 0.0030900142155587673, + 0.005691311322152615, + 0.0034069810062646866, + 0.004408850334584713, + 0.005885130260139704, + 0.012592213228344917, + 0.019504815340042114, + 0.0243679191917181, + 0.01569221168756485, + 0.052681706845760345, + 0.06172395125031471, + 0.042254529893398285, + 0.0487072579562664, + 0.020006919279694557, + 0.014788510277867317, + 0.026924578472971916, + 0.015494748018682003, + 0.029941372573375702, + 0.031651683151721954, + 0.03966295346617699, + 0.012649274431169033, + 0.012020958587527275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.02981995977461338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809615224599838 + ], + "opposite_logprob_tokens": [ + -3.2880961894989014 + ], + "logprob_reward": -0.03809615224599838, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809615224599838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26648715138435364, + 0.12325198203325272, + 0.09789326041936874, + 0.06549736112356186, + 0.044265300035476685, + 0.03190474212169647, + 0.04822968691587448, + 0.032444123178720474, + 0.1861458420753479, + 0.1038806140422821 + ], + "credit_map": [ + 0.014509275555610657, + 0.02252865768969059, + 0.021078497171401978, + 0.04104522988200188, + 0.0421287938952446, + 0.04133562371134758, + 0.03488374128937721, + 0.0489773154258728, + 0.007691035978496075, + 0.009752782993018627, + 0.009992478415369987, + 0.016995742917060852, + 0.013668813742697239, + 0.012078475207090378, + 0.018190616741776466, + 0.03488202765583992, + 0.007178262807428837, + 0.007135638501495123, + 0.01037613395601511, + 0.013251434080302715, + 0.00849165953695774, + 0.011541158892214298, + 0.012037853710353374, + 0.027881108224391937, + 0.00661799730733037, + 0.006522235926240683, + 0.006628618109971285, + 0.008950870484113693, + 0.006269935518503189, + 0.007614496164023876, + 0.007863862439990044, + 0.015029341913759708, + 0.006459059659391642, + 0.00561983697116375, + 0.004427353851497173, + 0.006523049436509609, + 0.0035341631155461073, + 0.00379460328258574, + 0.0046905833296477795, + 0.009216648526489735, + 0.004089892841875553, + 0.0031609879806637764, + 0.0026757160667330027, + 0.004807644989341497, + 0.00377231789752841, + 0.0030242903158068657, + 0.003711236175149679, + 0.006662653759121895, + 0.006633343640714884, + 0.006330601405352354, + 0.004961353726685047, + 0.005529745947569609, + 0.005162801593542099, + 0.007066467311233282, + 0.005761828739196062, + 0.006783541291952133, + 0.002953421324491501, + 0.005183487199246883, + 0.0026155465748161077, + 0.005994373932480812, + 0.0030550064984709024, + 0.0027836610097438097, + 0.004344637040048838, + 0.005513985175639391, + 0.0127702746540308, + 0.01995275542140007, + 0.011291472241282463, + 0.037296898663043976, + 0.043667953461408615, + 0.023584920912981033, + 0.023335859179496765, + 0.014245712198317051, + 0.007278259843587875, + 0.01710107922554016, + 0.007252129260450602, + 0.01207528356462717, + 0.02108466811478138, + 0.02000565268099308, + 0.01114623062312603, + 0.00793729443103075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.03809615224599838 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9742594957351685 + ], + "opposite_logprob_tokens": [ + -0.47425949573516846 + ], + "logprob_reward": -0.9742594957351685, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742594957351685, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19294197857379913, + 0.09168531745672226, + 0.07244127988815308, + 0.05028286203742027, + 0.07914692908525467, + 0.07894664257764816, + 0.07190213352441788, + 0.05403522774577141, + 0.17076478898525238, + 0.13785283267498016 + ], + "credit_map": [ + 0.028792252764105797, + 0.017424969002604485, + 0.023680411279201508, + 0.021375644952058792, + 0.02053373120725155, + 0.026597658172249794, + 0.024663599207997322, + 0.02987370826303959, + 0.014062750153243542, + 0.011606024578213692, + 0.00983812753111124, + 0.01021071057766676, + 0.008427531458437443, + 0.011722554452717304, + 0.013243012130260468, + 0.012574599124491215, + 0.007682851981371641, + 0.0072886100970208645, + 0.007026757579296827, + 0.011150314472615719, + 0.007785542402416468, + 0.009359177201986313, + 0.01190407294780016, + 0.01024395227432251, + 0.006493444554507732, + 0.005338903982192278, + 0.006056973244994879, + 0.0062872073613107204, + 0.005655518267303705, + 0.005910235922783613, + 0.005934522952884436, + 0.00860605575144291, + 0.00777658075094223, + 0.005396479740738869, + 0.008817194029688835, + 0.01137393806129694, + 0.007513134274631739, + 0.01138625480234623, + 0.01409548707306385, + 0.012787856161594391, + 0.010158060118556023, + 0.0099366819486022, + 0.006303634028881788, + 0.009861051104962826, + 0.007607326842844486, + 0.010731199756264687, + 0.013257650658488274, + 0.01109103299677372, + 0.008223889395594597, + 0.00744697405025363, + 0.006690577138215303, + 0.008833772502839565, + 0.006863266229629517, + 0.012860818766057491, + 0.008735623210668564, + 0.012247213162481785, + 0.004941650200635195, + 0.006470943335443735, + 0.004309650044888258, + 0.0068150474689900875, + 0.004819714464247227, + 0.00597741873934865, + 0.0063913436606526375, + 0.01430946309119463, + 0.010217188857495785, + 0.01650240644812584, + 0.015174067579209805, + 0.019945651292800903, + 0.04666890203952789, + 0.021955937147140503, + 0.02745971456170082, + 0.01284092292189598, + 0.015490247868001461, + 0.018567785620689392, + 0.011726408265531063, + 0.01776723936200142, + 0.022271517664194107, + 0.03152919188141823, + 0.009509257040917873, + 0.010991186834871769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.9742594957351685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with vibrant Disney Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot as a medium shot at eye level that remains momentarily static before moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fluffy ginger cat with light-colored tabby stripes and emerald eyes visible crouching on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall scratching post studded with scissors visible in front of the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sudden, sharp jerk mimicking an explosive forward thrust as the cat raises its paws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035856306552887 + ], + "opposite_logprob_tokens": [ + -1.9103585481643677 + ], + "logprob_reward": -0.16035856306552887, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035856306552887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08238156139850616, + 0.04445092752575874, + 0.039268456399440765, + 0.04261288419365883, + 0.04268123209476471, + 0.03443628177046776, + 0.03940451890230179, + 0.05952254310250282, + 0.32483214139938354, + 0.29040950536727905 + ], + "credit_map": [ + 0.006509872153401375, + 0.007111788261681795, + 0.007147613912820816, + 0.011141984723508358, + 0.00791857112199068, + 0.015028993599116802, + 0.010011361911892891, + 0.01751137338578701, + 0.0033222706988453865, + 0.0049908156506717205, + 0.003145440947264433, + 0.006685043219476938, + 0.004288658499717712, + 0.007895085029304028, + 0.005362939555197954, + 0.008760669268667698, + 0.002762438729405403, + 0.002275297185406089, + 0.002966690808534622, + 0.005602260120213032, + 0.003640038426965475, + 0.006070851814001799, + 0.006895592901855707, + 0.009055282920598984, + 0.003494864096865058, + 0.003155138809233904, + 0.002416888950392604, + 0.005723919253796339, + 0.004852555226534605, + 0.0057927388697862625, + 0.008107665926218033, + 0.009069111198186874, + 0.005519823636859655, + 0.004566344898194075, + 0.003988149110227823, + 0.005834283772855997, + 0.003223630366846919, + 0.0049827382899820805, + 0.009379968047142029, + 0.005186294671148062, + 0.0035416234750300646, + 0.002978429663926363, + 0.002569019328802824, + 0.005509689450263977, + 0.004913962446153164, + 0.005320005584508181, + 0.005109325982630253, + 0.004494223278015852, + 0.006177715957164764, + 0.0034244880080223083, + 0.0033870136830955744, + 0.004658163990825415, + 0.0033724659588187933, + 0.005347284954041243, + 0.005096037872135639, + 0.007941345684230328, + 0.005345988553017378, + 0.006474335212260485, + 0.003779903519898653, + 0.007033795118331909, + 0.006007638294249773, + 0.005797382444143295, + 0.008741204626858234, + 0.01634228229522705, + 0.018166648223996162, + 0.03373153880238533, + 0.015559676103293896, + 0.06069974601268768, + 0.08030480146408081, + 0.03415143862366676, + 0.05893854796886444, + 0.023279724642634392, + 0.019353704527020454, + 0.042894113808870316, + 0.031071340665221214, + 0.032210271805524826, + 0.04650922864675522, + 0.06774023920297623, + 0.022512119263410568, + 0.028118440881371498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.16035856306552887 + } + ], + "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/0000025/group_2/index_tensor([4482, 4482])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_2/index_tensor([4482, 4482])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..af728cabaabbcb2b24b6c62807b24e8c6b0d0238 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_2/index_tensor([4482, 4482])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "photorealistic 4K high-fidelity video begins with a static, locked-on close-up, maintaining sharp, unwavering focus on two subjects perfectly centered in the frame. In the foreground, a plump, ripe red tomato rests on a light-grained wooden cutting board; its vibrant scarlet skin is taut, subtly lustrous, and punctuated by tiny, natural imperfections that attest to its authenticity. The board itself displays faint, crisscrossing knife marks from frequent use, its matte texture contrasting sharply with the tomato's smooth surface. A small, partially obscured Pikachu plush toy lies directly behind the tomato, its soft, furry, bright yellow fabric providing textural contrast; its signature long ears, tipped with black, and a round, black button-eye, peek out from behind the fruit. The scene is set in a well-lit kitchen, bathed in a bright, even, diffuse, neutral white light that casts soft, subtle shadows and meticulously defines each object's form. The composition emphasizes surreal physical textures—from the waxy skin of a tomato to the fibrous weave of a plum—and their gentle interplay of shadows across the wooden surface, all contrasted against a clean, neutral white background to create a quiet, focused, and visually rich still life.", + "seed": 4896, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_2/index_4482_seed_4896_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": -0.010528165940195322, + "vlm_logprob": -0.010528165940195322, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.010528165940195322, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0039606234058737755 + ], + "opposite_logprob_tokens": [ + -8.003960609436035 + ], + "logprob_reward": -0.0039606234058737755, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.0039606234058737755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16003407537937164, + 0.05526186525821686, + 0.08149141818284988, + 0.06215914711356163, + 0.09530295431613922, + 0.06241418421268463, + 0.06456299871206284, + 0.034594036638736725, + 0.11845672875642776, + 0.2657226324081421 + ], + "credit_map": [ + 0.01736617274582386, + 0.026635268703103065, + 0.01098366267979145, + 0.01836402714252472, + 0.01963655836880207, + 0.01886530965566635, + 0.018978985026478767, + 0.02920408546924591, + 0.0065728952176868916, + 0.008002050220966339, + 0.005155827850103378, + 0.006676197052001953, + 0.006210441701114178, + 0.006801092531532049, + 0.005366134457290173, + 0.01047721691429615, + 0.015426094643771648, + 0.008516556583344936, + 0.009550056420266628, + 0.009240420535206795, + 0.006141343619674444, + 0.011155744083225727, + 0.007458958309143782, + 0.014002237468957901, + 0.007607898209244013, + 0.008338372223079205, + 0.005714790429919958, + 0.007980234920978546, + 0.011191715486347675, + 0.00597082544118166, + 0.007959580048918724, + 0.007395721040666103, + 0.01153961569070816, + 0.009582949802279472, + 0.011960797943174839, + 0.011757578700780869, + 0.01133895106613636, + 0.010371558368206024, + 0.01047336682677269, + 0.018278131261467934, + 0.008515525609254837, + 0.006253369618207216, + 0.004803032148629427, + 0.0073792231269180775, + 0.0057741194032132626, + 0.007377623114734888, + 0.008206182159483433, + 0.01410510204732418, + 0.013429898768663406, + 0.00909414142370224, + 0.008320347405970097, + 0.005340504460036755, + 0.004049980081617832, + 0.006148470100015402, + 0.00649205083027482, + 0.011687601916491985, + 0.005119715817272663, + 0.0037600677460432053, + 0.004303542897105217, + 0.0038403449580073357, + 0.003035626607015729, + 0.0040902611799538136, + 0.0036871794145554304, + 0.00675729475915432, + 0.01270280685275793, + 0.008562029339373112, + 0.007153863087296486, + 0.01187906414270401, + 0.0073150270618498325, + 0.020226160064339638, + 0.01599295064806938, + 0.03462481498718262, + 0.019120298326015472, + 0.013344107195734978, + 0.012624756433069706, + 0.038089215755462646, + 0.01676560565829277, + 0.05441634729504585, + 0.054107192903757095, + 0.05725507810711861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0039606234058737755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659076876938343 + ], + "opposite_logprob_tokens": [ + -4.7586588859558105 + ], + "logprob_reward": -0.008659076876938343, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008659076876938343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09654086083173752, + 0.15330815315246582, + 0.1303863376379013, + 0.08534353971481323, + 0.09025012701749802, + 0.037775397300720215, + 0.029903516173362732, + 0.13213679194450378, + 0.03713083267211914, + 0.20722441375255585 + ], + "credit_map": [ + 0.021104728803038597, + 0.011736787855625153, + 0.011078975163400173, + 0.01135323103517294, + 0.008200723677873611, + 0.009901325218379498, + 0.00837499089539051, + 0.014790095388889313, + 0.03678435832262039, + 0.0212759617716074, + 0.01946878433227539, + 0.011151980608701706, + 0.010535599663853645, + 0.017912661656737328, + 0.010685082525014877, + 0.02549372985959053, + 0.026309965178370476, + 0.017769914120435715, + 0.01729855313897133, + 0.0097123421728611, + 0.005925661418586969, + 0.020151738077402115, + 0.008900823071599007, + 0.024317340925335884, + 0.019911564886569977, + 0.009956846944987774, + 0.01118127629160881, + 0.007117378059774637, + 0.0066977692767977715, + 0.010377469472587109, + 0.007602430414408445, + 0.012498804368078709, + 0.016726188361644745, + 0.0078094047494232655, + 0.00848417729139328, + 0.009685631841421127, + 0.007214158773422241, + 0.011112382635474205, + 0.00921088270843029, + 0.020007299259305, + 0.006309323012828827, + 0.004381181672215462, + 0.0061996327713131905, + 0.004241399001330137, + 0.003204069100320339, + 0.003019997151568532, + 0.004628969822078943, + 0.005790825933218002, + 0.0057372599840164185, + 0.003709290875121951, + 0.003146701492369175, + 0.0030949441716074944, + 0.0023887816350907087, + 0.0029672705568373203, + 0.003967661876231432, + 0.00489160418510437, + 0.0292508602142334, + 0.012004011310636997, + 0.014511197805404663, + 0.009470447897911072, + 0.009565088897943497, + 0.01239454559981823, + 0.012035297229886055, + 0.03290534019470215, + 0.005390864331275225, + 0.001896355301141739, + 0.004858443047851324, + 0.0027854982763528824, + 0.002437364077195525, + 0.00676976703107357, + 0.004429833497852087, + 0.008562705479562283, + 0.017815856263041496, + 0.010014696046710014, + 0.013286327943205833, + 0.022074950858950615, + 0.01158897764980793, + 0.04773860424757004, + 0.03452253341674805, + 0.050182465463876724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.008659076876938343 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581839606165886 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581839606165886, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581839606165886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10244324058294296, + 0.03109046071767807, + 0.032934028655290604, + 0.08447515964508057, + 0.3196571171283722, + 0.2823989987373352, + 0.05123523622751236, + 0.01176995225250721, + 0.009984705597162247, + 0.07401108741760254 + ], + "credit_map": [ + 0.006788774859160185, + 0.013817875646054745, + 0.007623080164194107, + 0.01128584984689951, + 0.01461118459701538, + 0.013327859342098236, + 0.011506358161568642, + 0.023482276126742363, + 0.003046884899958968, + 0.00429902970790863, + 0.0023598100524395704, + 0.003924897871911526, + 0.003924065735191107, + 0.00366508262231946, + 0.00392857426777482, + 0.005942119285464287, + 0.004039213992655277, + 0.006515030283480883, + 0.0028305102605372667, + 0.004031541757285595, + 0.004022078588604927, + 0.003943357151001692, + 0.003381098620593548, + 0.004171202890574932, + 0.01766437664628029, + 0.010086643509566784, + 0.011088410392403603, + 0.010130110196769238, + 0.007194128818809986, + 0.008612784557044506, + 0.008682147599756718, + 0.011016561649739742, + 0.04133646562695503, + 0.03126008063554764, + 0.03520914167165756, + 0.041287168860435486, + 0.02473260462284088, + 0.038937076926231384, + 0.037227559834718704, + 0.0696670338511467, + 0.04310130700469017, + 0.02731107920408249, + 0.026694830507040024, + 0.03175238519906998, + 0.024567147716879845, + 0.034147389233112335, + 0.0327962189912796, + 0.062028635293245316, + 0.00867961160838604, + 0.005869904067367315, + 0.006223931442946196, + 0.0054143560118973255, + 0.0035491078160703182, + 0.004638655111193657, + 0.0056503345258533955, + 0.011209341697394848, + 0.0013470277190208435, + 0.0014032362960278988, + 0.001469178474508226, + 0.0014829020947217941, + 0.0011219207663089037, + 0.0012102484470233321, + 0.0013307236367836595, + 0.0024047151673585176, + 0.0014278696617111564, + 0.000662614475004375, + 0.0010009155375882983, + 0.0006859353743493557, + 0.0006416490650735795, + 0.0010129481088370085, + 0.0009246402769349515, + 0.0036281344946473837, + 0.006222838535904884, + 0.0031475576106458902, + 0.003362414427101612, + 0.007202270440757275, + 0.00425673509016633, + 0.0191997941583395, + 0.011473865248262882, + 0.01914561353623867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.020581839606165886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248221103101969 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005248221103101969, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248221103101969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08379114419221878, + 0.034165214747190475, + 0.043028514832258224, + 0.10417120158672333, + 0.3448236286640167, + 0.29332154989242554, + 0.022831333801150322, + 0.014662125147879124, + 0.014784777536988258, + 0.044420529156923294 + ], + "credit_map": [ + 0.009374372661113739, + 0.011373409070074558, + 0.007782378233969212, + 0.010401356033980846, + 0.011207450181245804, + 0.009165746159851551, + 0.01177875604480505, + 0.01270767580717802, + 0.004347084555774927, + 0.003786127083003521, + 0.003958322573453188, + 0.0038603630382567644, + 0.005417310632765293, + 0.003673038212582469, + 0.0043001663871109486, + 0.004822800401598215, + 0.006377229001373053, + 0.00726248137652874, + 0.004390196409076452, + 0.00492831040173769, + 0.004804089665412903, + 0.0052199047058820724, + 0.004122646525502205, + 0.005923656281083822, + 0.020446253940463066, + 0.013347551226615906, + 0.013620632700622082, + 0.01193773839622736, + 0.010052639059722424, + 0.010746940039098263, + 0.010570998303592205, + 0.013448457233607769, + 0.04320244863629341, + 0.03317756950855255, + 0.03641024976968765, + 0.04527098685503006, + 0.027362581342458725, + 0.042259782552719116, + 0.040691688656806946, + 0.07644833624362946, + 0.03936353325843811, + 0.03152599185705185, + 0.02601691149175167, + 0.034363534301519394, + 0.024220969527959824, + 0.036111533641815186, + 0.03482316434383392, + 0.06689591705799103, + 0.00402161804959178, + 0.002440617885440588, + 0.0023819084744900465, + 0.002095115603879094, + 0.0016770721413195133, + 0.00206109881401062, + 0.0022978363558650017, + 0.005856065079569817, + 0.0016089414712041616, + 0.0014576022513210773, + 0.001733821234665811, + 0.0015636024763807654, + 0.0012448620982468128, + 0.0013771161902695894, + 0.001824319246225059, + 0.0038518602959811687, + 0.0021228918340057135, + 0.001982753863558173, + 0.0017962501151487231, + 0.0012472299858927727, + 0.0011538575636222959, + 0.0019601762760430574, + 0.0018736848141998053, + 0.0026479330845177174, + 0.005039926618337631, + 0.002342323074117303, + 0.0038193485233932734, + 0.008791874162852764, + 0.0025160640943795443, + 0.0027260046917945147, + 0.012278134003281593, + 0.006906856782734394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.005248221103101969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191068708896637 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191068708896637, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191068708896637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03676338866353035, + 0.01039393525570631, + 0.016657356172800064, + 0.010512909851968288, + 0.02761245332658291, + 0.011944815516471863, + 0.01444901805371046, + 0.025280291214585304, + 0.018587898463010788, + 0.8277978897094727 + ], + "credit_map": [ + 0.0040807039476931095, + 0.004391686525195837, + 0.002340053441002965, + 0.004414124879986048, + 0.004733244888484478, + 0.005116108804941177, + 0.0042435904033482075, + 0.007443878799676895, + 0.001116485451348126, + 0.0011728135868906975, + 0.0014624865725636482, + 0.0014131878269836307, + 0.0016179142985492945, + 0.0010052949655801058, + 0.0012437497498467565, + 0.0013620032696053386, + 0.003135672304779291, + 0.0024572189431637526, + 0.001934438361786306, + 0.0013986830599606037, + 0.001078232889994979, + 0.002242014743387699, + 0.0013545048423111439, + 0.0030565934721380472, + 0.001321909250691533, + 0.0019759475253522396, + 0.0009663398959673941, + 0.0011168473865836859, + 0.000950286746956408, + 0.0011959252879023552, + 0.0009592447313480079, + 0.0020264098420739174, + 0.00410185381770134, + 0.0028380979783833027, + 0.0034558719489723444, + 0.0028180801309645176, + 0.002222175942733884, + 0.0033852618653327227, + 0.002737188246101141, + 0.006053924094885588, + 0.0027324019465595484, + 0.001312572043389082, + 0.0012446356704458594, + 0.0012009311467409134, + 0.0006587643874809146, + 0.0012443700106814504, + 0.0011556977406144142, + 0.0023954445496201515, + 0.0012164993677288294, + 0.0016207778826355934, + 0.0011318827746436, + 0.0018449979834258556, + 0.0017150057246908545, + 0.0016209406312555075, + 0.0024756337516009808, + 0.0028232818003743887, + 0.00572962174192071, + 0.0023584463633596897, + 0.0028799904976040125, + 0.0018842583522200584, + 0.0019339328864589334, + 0.0023580726701766253, + 0.002414212329313159, + 0.005721759051084518, + 0.0022911045234650373, + 0.001538991928100586, + 0.0019845152273774147, + 0.001356725231744349, + 0.0011492632329463959, + 0.0022860607132315636, + 0.0018746013520285487, + 0.006106638349592686, + 0.05993567407131195, + 0.03385251387953758, + 0.03291041776537895, + 0.17096258699893951, + 0.04038181155920029, + 0.10208488255739212, + 0.24536001682281494, + 0.1423100382089615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.014191068708896637 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "photorealistic 4K high-fidelity video begins with a static, locked-on close-up, maintaining sharp, unwavering focus on two subjects perfectly centered in the frame. In the foreground, a plump, ripe red tomato rests on a light-grained wooden cutting board; its vibrant scarlet skin is taut, subtly lustrous, and punctuated by tiny, natural imperfections that attest to its authenticity. The board itself displays faint, crisscrossing knife marks from frequent use, its matte texture contrasting sharply with the tomato's smooth surface. A small, partially obscured Pikachu plush toy lies directly behind the tomato, its soft, furry, bright yellow fabric providing textural contrast; its signature long ears, tipped with black, and a round, black button-eye, peek out from behind the fruit. The scene is set in a well-lit kitchen, bathed in a bright, even, diffuse, neutral white light that casts soft, subtle shadows and meticulously defines each object's form. The composition emphasizes surreal physical textures—from the waxy skin of a tomato to the fibrous weave of a plum—and their gentle interplay of shadows across the wooden surface, all contrasted against a clean, neutral white background to create a quiet, focused, and visually rich still life.", + "seed": 4897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_2/index_4482_seed_4897_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": -0.008919007214717566, + "vlm_logprob": -0.008919007214717566, + "vlm_yes_no_margin": 5.65, + "vlm_token_credit": -0.008919007214717566, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005698269698768854 + ], + "opposite_logprob_tokens": [ + -7.755698204040527 + ], + "logprob_reward": -0.005698269698768854, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.005698269698768854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08211459219455719, + 0.11301147937774658, + 0.12259957194328308, + 0.03896626830101013, + 0.044908326119184494, + 0.0932253748178482, + 0.2503525912761688, + 0.13584394752979279, + 0.0833219364285469, + 0.03565588593482971 + ], + "credit_map": [ + 0.006037411745637655, + 0.00940645206719637, + 0.010874219238758087, + 0.011578108184039593, + 0.011203733272850513, + 0.007602829951792955, + 0.014183567836880684, + 0.011228279210627079, + 0.006373309995979071, + 0.008358643390238285, + 0.01601988449692726, + 0.019460024312138557, + 0.008312908932566643, + 0.007877720519900322, + 0.024090178310871124, + 0.022518813610076904, + 0.006452072877436876, + 0.010562494397163391, + 0.009100944735109806, + 0.02440737932920456, + 0.009762763977050781, + 0.00958648044615984, + 0.03266662731766701, + 0.020060822367668152, + 0.005955217406153679, + 0.0034209166187793016, + 0.005694946739822626, + 0.0038804211653769016, + 0.003986269235610962, + 0.005716342478990555, + 0.005573454778641462, + 0.004738700110465288, + 0.006759424228221178, + 0.0043596005998551846, + 0.007779167033731937, + 0.006524655036628246, + 0.004810047335922718, + 0.0039013116620481014, + 0.005745156202465296, + 0.00502896960824728, + 0.008262574672698975, + 0.006108404602855444, + 0.01621183753013611, + 0.012765203602612019, + 0.008936144411563873, + 0.01348795648664236, + 0.0128556527197361, + 0.014597611501812935, + 0.01690049283206463, + 0.013574119657278061, + 0.037342533469200134, + 0.04440611973404884, + 0.01310877874493599, + 0.0187541451305151, + 0.0553305484354496, + 0.05093584582209587, + 0.011383209377527237, + 0.006549193058162928, + 0.016604194417595863, + 0.025813352316617966, + 0.0087759243324399, + 0.008928597904741764, + 0.03208789974451065, + 0.02570159360766411, + 0.0045312028378248215, + 0.0033741884399205446, + 0.01091020181775093, + 0.014822030439972878, + 0.0049967169761657715, + 0.005694808904081583, + 0.019655445590615273, + 0.01933733932673931, + 0.0026027513667941093, + 0.002751367399469018, + 0.0030339029617607594, + 0.006037742365151644, + 0.00491477083414793, + 0.003226520726457238, + 0.007804191205650568, + 0.005284642800688744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.005698269698768854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011892396723851562 + ], + "opposite_logprob_tokens": [ + -6.751189231872559 + ], + "logprob_reward": -0.0011892396723851562, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011892396723851562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04885251820087433, + 0.10297061502933502, + 0.05073011666536331, + 0.04983406141400337, + 0.032326579093933105, + 0.09381263703107834, + 0.23306584358215332, + 0.3301689624786377, + 0.03926340490579605, + 0.01897531934082508 + ], + "credit_map": [ + 0.005772025790065527, + 0.0036837647203356028, + 0.012707033194601536, + 0.005509324371814728, + 0.004194004461169243, + 0.0033456594683229923, + 0.004643239080905914, + 0.008997471071779728, + 0.007300894241780043, + 0.01034346129745245, + 0.010799768380820751, + 0.01634950004518032, + 0.009615377523005009, + 0.009256516583263874, + 0.021686535328626633, + 0.017618563026189804, + 0.005833837669342756, + 0.0069903414696455, + 0.00826833862811327, + 0.005399878602474928, + 0.006405307911336422, + 0.005631304811686277, + 0.0061506303027272224, + 0.006050475873053074, + 0.004552232101559639, + 0.006070589646697044, + 0.00879347138106823, + 0.005322164855897427, + 0.005484764464199543, + 0.005850363522768021, + 0.005643726326525211, + 0.008116749115288258, + 0.0032350076362490654, + 0.0030897788237780333, + 0.006047023460268974, + 0.0031082225032150745, + 0.002990155015140772, + 0.0036601428873836994, + 0.003358288202434778, + 0.006837960798293352, + 0.008938174694776535, + 0.009609989821910858, + 0.009170706383883953, + 0.014143571257591248, + 0.008324910886585712, + 0.009190143086016178, + 0.02076900564134121, + 0.013666135258972645, + 0.019724024459719658, + 0.014407250098884106, + 0.027487633749842644, + 0.04155946895480156, + 0.014579640701413155, + 0.01951492577791214, + 0.05297113582491875, + 0.04282176494598389, + 0.013067612424492836, + 0.014861227944493294, + 0.04630574584007263, + 0.058687031269073486, + 0.017755283042788506, + 0.01353679783642292, + 0.08017867803573608, + 0.08577658981084824, + 0.001979788765311241, + 0.002686311025172472, + 0.004783087410032749, + 0.006975584663450718, + 0.0029230390209704638, + 0.0029320400208234787, + 0.009268556721508503, + 0.007714997977018356, + 0.0012204587692394853, + 0.002078892895951867, + 0.002893578726798296, + 0.0026120529510080814, + 0.0021326628047972918, + 0.0025473572313785553, + 0.0027087819762527943, + 0.0027815341018140316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.0011892396723851562 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294245824217796 + ], + "opposite_logprob_tokens": [ + -3.77329421043396 + ], + "logprob_reward": -0.023294245824217796, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294245824217796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052060287445783615, + 0.025744590908288956, + 0.020530231297016144, + 0.028128568083047867, + 0.03258504718542099, + 0.07646603137254715, + 0.4744173288345337, + 0.2744031250476837, + 0.009029638022184372, + 0.00663516391068697 + ], + "credit_map": [ + 0.005897935479879379, + 0.006064127665013075, + 0.00783323310315609, + 0.008473574183881283, + 0.005425196141004562, + 0.0043813674710690975, + 0.00805540569126606, + 0.0059294505044817924, + 0.004200222436338663, + 0.0030218265019357204, + 0.00409533130005002, + 0.0032155185472220182, + 0.0025165907572954893, + 0.003091118996962905, + 0.0025755774695426226, + 0.003028403501957655, + 0.0011978816473856568, + 0.001782106701284647, + 0.0016606077551841736, + 0.0037638498470187187, + 0.0018893915694206953, + 0.0015812600031495094, + 0.0050490256398916245, + 0.0036061082500964403, + 0.003385842777788639, + 0.00298981717787683, + 0.0032393469009548426, + 0.0030794385820627213, + 0.002610016381368041, + 0.002926604123786092, + 0.004593957215547562, + 0.005303543526679277, + 0.0028331398498266935, + 0.0018820887198671699, + 0.008321871981024742, + 0.004104472231119871, + 0.0020728979725390673, + 0.002194576431065798, + 0.003751155221834779, + 0.007424843031913042, + 0.008960353210568428, + 0.00790481548756361, + 0.010954913683235645, + 0.008999524638056755, + 0.008654426783323288, + 0.011261731386184692, + 0.01169720571488142, + 0.008033058606088161, + 0.026587843894958496, + 0.019759660586714745, + 0.05806112661957741, + 0.09183818846940994, + 0.024967163801193237, + 0.029283203184604645, + 0.12144488841295242, + 0.10247524827718735, + 0.019578145816922188, + 0.011112368665635586, + 0.042209383100271225, + 0.047604694962501526, + 0.015175077132880688, + 0.018520137295126915, + 0.05917130783200264, + 0.0610320121049881, + 0.00042989279609173536, + 0.0005702091730199754, + 0.001533181988634169, + 0.0013842858606949449, + 0.0006355774239636958, + 0.0006014087703078985, + 0.0017457776702940464, + 0.0021293037571012974, + 0.00040156158502213657, + 0.0005785708199255168, + 0.0012928724754601717, + 0.0008998223347589374, + 0.0006545555079355836, + 0.0006972532719373703, + 0.0009271292365156114, + 0.0011833987664431334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.023294245824217796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633998032659292 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004633998032659292, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633998032659292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08431516587734222, + 0.027728745713829994, + 0.018903136253356934, + 0.034463733434677124, + 0.023241205140948296, + 0.06380370259284973, + 0.44906553626060486, + 0.2658403515815735, + 0.024016350507736206, + 0.008622104302048683 + ], + "credit_map": [ + 0.006270779296755791, + 0.007843241095542908, + 0.013844361528754234, + 0.01112344954162836, + 0.007899724878370762, + 0.009534839540719986, + 0.012243451550602913, + 0.01555531844496727, + 0.003608238883316517, + 0.0024743610993027687, + 0.00330825662240386, + 0.004487290512770414, + 0.00281133153475821, + 0.0032121664844453335, + 0.004663397092372179, + 0.0031637016218155622, + 0.001399010419845581, + 0.0016215733485296369, + 0.0019727565813809633, + 0.0033242113422602415, + 0.0015880911378189921, + 0.0014240314485505223, + 0.0043258266523480415, + 0.003247633809223771, + 0.00227643852122128, + 0.003735313657671213, + 0.005737396888434887, + 0.004317129496484995, + 0.003109946846961975, + 0.003245953470468521, + 0.005200526211410761, + 0.006841027177870274, + 0.002028822433203459, + 0.0020203168969601393, + 0.003939456306397915, + 0.0031057146843522787, + 0.0018115995917469263, + 0.002266917610540986, + 0.003749782918021083, + 0.004318594001233578, + 0.006851006764918566, + 0.005730315577238798, + 0.00962273869663477, + 0.00747270742431283, + 0.007729281205683947, + 0.010423592291772366, + 0.009348863735795021, + 0.006625194568186998, + 0.02568514086306095, + 0.018795615062117577, + 0.05240117385983467, + 0.08811789005994797, + 0.023739976808428764, + 0.027543680742383003, + 0.11697563529014587, + 0.09580639749765396, + 0.019480912014842033, + 0.010679309256374836, + 0.03946060314774513, + 0.04723108187317848, + 0.014551004394888878, + 0.017933810129761696, + 0.058695655316114426, + 0.05780796706676483, + 0.0012895498657599092, + 0.0019077074248343706, + 0.0030022086575627327, + 0.004052675794810057, + 0.002127512823790312, + 0.0022072463762015104, + 0.005132692866027355, + 0.004296757280826569, + 0.000579503714106977, + 0.0010044407099485397, + 0.0017167915357276797, + 0.0008524000877514482, + 0.0010309044737368822, + 0.0009393757791258395, + 0.0008608583593741059, + 0.0016378298169001937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.004633998032659292 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779282845556736 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779282845556736, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779282845556736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18557938933372498, + 0.06035313755273819, + 0.040848542004823685, + 0.11116884648799896, + 0.081085205078125, + 0.06956694275140762, + 0.20293515920639038, + 0.1291544884443283, + 0.0897531807422638, + 0.029555141925811768 + ], + "credit_map": [ + 0.012314887717366219, + 0.011137524619698524, + 0.030076919123530388, + 0.02802341990172863, + 0.01136583462357521, + 0.010912518016994, + 0.03592119738459587, + 0.04582709074020386, + 0.004303302615880966, + 0.007386656012386084, + 0.010187219828367233, + 0.0063742585480213165, + 0.008860645815730095, + 0.009902559220790863, + 0.006107910070568323, + 0.007230585906654596, + 0.0032300730235874653, + 0.005365123972296715, + 0.00529223307967186, + 0.004373172298073769, + 0.006237371359020472, + 0.007888014428317547, + 0.003882448421791196, + 0.004580103326588869, + 0.005850554443895817, + 0.006606339477002621, + 0.020743470638990402, + 0.01737217605113983, + 0.0075749666430056095, + 0.008172402158379555, + 0.020050041377544403, + 0.02479889616370201, + 0.00546836294233799, + 0.009408040903508663, + 0.013105368241667747, + 0.012400799430906773, + 0.007923958823084831, + 0.007210279814898968, + 0.015047233551740646, + 0.01052115485072136, + 0.005506682675331831, + 0.004804383032023907, + 0.012480136938393116, + 0.007259408012032509, + 0.008575765416026115, + 0.011628489010035992, + 0.007059132214635611, + 0.012252950109541416, + 0.009875481016933918, + 0.008138389326632023, + 0.029163038358092308, + 0.0370827317237854, + 0.010693756863474846, + 0.011019899509847164, + 0.04883043095469475, + 0.048131439834833145, + 0.008085032925009727, + 0.006151793524622917, + 0.012776979245245457, + 0.025245996192097664, + 0.008976282551884651, + 0.010619784705340862, + 0.03271894156932831, + 0.024579672142863274, + 0.006073955446481705, + 0.010654188692569733, + 0.015409953892230988, + 0.009967798367142677, + 0.009876844473183155, + 0.011972913518548012, + 0.011431632563471794, + 0.01436589565128088, + 0.002233823761343956, + 0.004311990924179554, + 0.004075018223375082, + 0.0033805284183472395, + 0.0039188857190310955, + 0.0047071753069758415, + 0.003638577414676547, + 0.0032891398295760155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.009779282845556736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "photorealistic 4K high-fidelity video begins with a static, locked-on close-up, maintaining sharp, unwavering focus on two subjects perfectly centered in the frame. In the foreground, a plump, ripe red tomato rests on a light-grained wooden cutting board; its vibrant scarlet skin is taut, subtly lustrous, and punctuated by tiny, natural imperfections that attest to its authenticity. The board itself displays faint, crisscrossing knife marks from frequent use, its matte texture contrasting sharply with the tomato's smooth surface. A small, partially obscured Pikachu plush toy lies directly behind the tomato, its soft, furry, bright yellow fabric providing textural contrast; its signature long ears, tipped with black, and a round, black button-eye, peek out from behind the fruit. The scene is set in a well-lit kitchen, bathed in a bright, even, diffuse, neutral white light that casts soft, subtle shadows and meticulously defines each object's form. The composition emphasizes surreal physical textures—from the waxy skin of a tomato to the fibrous weave of a plum—and their gentle interplay of shadows across the wooden surface, all contrasted against a clean, neutral white background to create a quiet, focused, and visually rich still life.", + "seed": 4898, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_2/index_4482_seed_4898_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": -0.007886783697176724, + "vlm_logprob": -0.007886783697176724, + "vlm_yes_no_margin": 5.7, + "vlm_token_credit": -0.007886783697176724, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005666739773005247 + ], + "opposite_logprob_tokens": [ + -5.880666732788086 + ], + "logprob_reward": -0.005666739773005247, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.005666739773005247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2908419072628021, + 0.15645407140254974, + 0.03469647839665413, + 0.05012056604027748, + 0.04012082889676094, + 0.022927431389689445, + 0.18629857897758484, + 0.05321584641933441, + 0.13120803236961365, + 0.03411625698208809 + ], + "credit_map": [ + 0.018153095617890358, + 0.0208121445029974, + 0.08928985148668289, + 0.017894377931952477, + 0.02697291411459446, + 0.019678587093949318, + 0.012096748687326908, + 0.08594419062137604, + 0.009633626788854599, + 0.011526438407599926, + 0.04934084415435791, + 0.010586750693619251, + 0.010768359526991844, + 0.010987162590026855, + 0.006313835736364126, + 0.04729706048965454, + 0.0021472920197993517, + 0.0016927614342421293, + 0.01078145019710064, + 0.0029433509334921837, + 0.002467629499733448, + 0.002153827575966716, + 0.0017080280231311917, + 0.010802141390740871, + 0.0045730010606348515, + 0.005268837325274944, + 0.011039652861654758, + 0.005299755372107029, + 0.006257407367229462, + 0.006063998676836491, + 0.0029987075831741095, + 0.00861920416355133, + 0.0018295500194653869, + 0.002310859737917781, + 0.01187979243695736, + 0.004064889624714851, + 0.0031650704331696033, + 0.0025863174814730883, + 0.0021871598437428474, + 0.012097193859517574, + 0.0015816886443644762, + 0.002592083066701889, + 0.004907106515020132, + 0.003062436357140541, + 0.002392159542068839, + 0.002643048297613859, + 0.0013734950916841626, + 0.004375413525849581, + 0.014286571182310581, + 0.012954742647707462, + 0.05623483657836914, + 0.015376071445643902, + 0.011532755568623543, + 0.01756223291158676, + 0.008494668640196323, + 0.04985671862959862, + 0.004368399735540152, + 0.005326505750417709, + 0.013961756601929665, + 0.005431639030575752, + 0.004875912331044674, + 0.005348152946680784, + 0.0031354192178696394, + 0.01076805591583252, + 0.008594643324613571, + 0.012053069658577442, + 0.033342309296131134, + 0.009845949709415436, + 0.012350056320428848, + 0.012888148427009583, + 0.007029934786260128, + 0.03510390222072601, + 0.0026640300638973713, + 0.0029767388477921486, + 0.00995005201548338, + 0.0027510449290275574, + 0.0025119443889707327, + 0.003113759681582451, + 0.0015891174552962184, + 0.008559567853808403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.005666739773005247 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005720409681089222 + ], + "opposite_logprob_tokens": [ + -7.500572204589844 + ], + "logprob_reward": -0.0005720409681089222, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005720409681089222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24885323643684387, + 0.09513214975595474, + 0.17090709507465363, + 0.09782381355762482, + 0.13142409920692444, + 0.07939475029706955, + 0.043011292815208435, + 0.03523114696145058, + 0.08827348798513412, + 0.009949013590812683 + ], + "credit_map": [ + 0.019629506394267082, + 0.024959534406661987, + 0.06369973719120026, + 0.02224006876349449, + 0.028457844629883766, + 0.027259461581707, + 0.013682826422154903, + 0.048924244940280914, + 0.006137380842119455, + 0.009721428155899048, + 0.026348434388637543, + 0.007635628804564476, + 0.0079492861405015, + 0.01010726299136877, + 0.0043684509582817554, + 0.022864272817969322, + 0.01066265907138586, + 0.009593246504664421, + 0.05842387676239014, + 0.011607865802943707, + 0.00847629550844431, + 0.00994946714490652, + 0.006303749047219753, + 0.055889926850795746, + 0.004823054186999798, + 0.004963899031281471, + 0.03467680886387825, + 0.0066002835519611835, + 0.004832216538488865, + 0.004453347064554691, + 0.003585644531995058, + 0.0338885672390461, + 0.006807105615735054, + 0.005860522389411926, + 0.04623590409755707, + 0.009153672493994236, + 0.006820398848503828, + 0.005907632876187563, + 0.0048366268165409565, + 0.04580222815275192, + 0.004903733730316162, + 0.00389407342299819, + 0.0270648542791605, + 0.005703640636056662, + 0.004290350712835789, + 0.003748912364244461, + 0.0029658740386366844, + 0.02682330831885338, + 0.0026310475077480078, + 0.003008484374731779, + 0.011741893365979195, + 0.003990957047790289, + 0.003994547296315432, + 0.0031185317784547806, + 0.0024121215101331472, + 0.012113711796700954, + 0.0015078127617016435, + 0.0023650401271879673, + 0.009874598123133183, + 0.003953106235712767, + 0.0033670850098133087, + 0.0018908511847257614, + 0.0019679998513311148, + 0.010304652154445648, + 0.005146725568920374, + 0.006048058625310659, + 0.026398401707410812, + 0.006671090144664049, + 0.005636660382151604, + 0.006867506075650454, + 0.004141710232943296, + 0.02736334130167961, + 0.0008797079208306968, + 0.0012104392517358065, + 0.0013189358869567513, + 0.0017749974504113197, + 0.0018312912434339523, + 0.000807770062237978, + 0.0008749055559746921, + 0.00125096645206213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0005720409681089222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636094018816948 + ], + "opposite_logprob_tokens": [ + -3.6513609886169434 + ], + "logprob_reward": -0.02636094018816948, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636094018816948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20136354863643646, + 0.1092258170247078, + 0.1254175901412964, + 0.05311649292707443, + 0.0656590461730957, + 0.06363967061042786, + 0.21669240295886993, + 0.059155095368623734, + 0.0886114165186882, + 0.01711897738277912 + ], + "credit_map": [ + 0.012474138289690018, + 0.013945789076387882, + 0.06318603456020355, + 0.016832921653985977, + 0.016123013570904732, + 0.01056660432368517, + 0.00891369953751564, + 0.05932135879993439, + 0.007452573627233505, + 0.00786894466727972, + 0.032827094197273254, + 0.009168863296508789, + 0.007025747559964657, + 0.008887577801942825, + 0.004882193636149168, + 0.03111281991004944, + 0.006615424528717995, + 0.00616149278357625, + 0.04080617055296898, + 0.010356074199080467, + 0.008563565090298653, + 0.0067541226744651794, + 0.005414783954620361, + 0.04074596241116524, + 0.003248624736443162, + 0.0056467303074896336, + 0.01549591589719057, + 0.004803923889994621, + 0.0039599002338945866, + 0.0032893039751797915, + 0.002261610934510827, + 0.014410486444830894, + 0.0028575651813298464, + 0.0036918979603797197, + 0.020690754055976868, + 0.005753911565989256, + 0.004445980302989483, + 0.003448988776654005, + 0.0028675447683781385, + 0.021902402862906456, + 0.0028623160906136036, + 0.004037149250507355, + 0.019254153594374657, + 0.0045245494693517685, + 0.006561276502907276, + 0.003518300596624613, + 0.0031300655100494623, + 0.019751863554120064, + 0.009886665269732475, + 0.014538563787937164, + 0.06356306374073029, + 0.01810322143137455, + 0.020526055246591568, + 0.01259035523980856, + 0.01130940206348896, + 0.06617506593465805, + 0.004232608247548342, + 0.0028204312548041344, + 0.019368568435311317, + 0.005031963344663382, + 0.003130721626803279, + 0.0034252654295414686, + 0.0024801481049507856, + 0.018665390089154243, + 0.0073614115826785564, + 0.010500011034309864, + 0.013466229662299156, + 0.012285321019589901, + 0.013743836432695389, + 0.010130437090992928, + 0.007043140009045601, + 0.014081022702157497, + 0.001576693612150848, + 0.0015212688595056534, + 0.0033537764102220535, + 0.0017446601996198297, + 0.0023183003067970276, + 0.0021804601419717073, + 0.001154201803728938, + 0.003269616048783064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.02636094018816948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021967582870274782 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021967582870274782, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021967582870274782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12158196419477463, + 0.07934550195932388, + 0.0875922292470932, + 0.1350279450416565, + 0.10229963064193726, + 0.06552807241678238, + 0.07798531651496887, + 0.07751864939928055, + 0.120453380048275, + 0.13266728818416595 + ], + "credit_map": [ + 0.011187094263732433, + 0.012834856286644936, + 0.019732598215341568, + 0.015146900899708271, + 0.018889596685767174, + 0.01315477304160595, + 0.009605003520846367, + 0.02103114314377308, + 0.00629423838108778, + 0.007524380460381508, + 0.018052229657769203, + 0.010617931373417377, + 0.0072264885529875755, + 0.007044002413749695, + 0.005132007412612438, + 0.017454221844673157, + 0.004973180592060089, + 0.007265559863299131, + 0.020259598270058632, + 0.011024897918105125, + 0.010474936105310917, + 0.00661025196313858, + 0.005770896561443806, + 0.021212909370660782, + 0.008578157052397728, + 0.01118761021643877, + 0.04211842268705368, + 0.009400441311299801, + 0.008835562504827976, + 0.008341155014932156, + 0.0059247068129479885, + 0.040641900151968, + 0.005313101224601269, + 0.005768713541328907, + 0.03477954491972923, + 0.00714529026299715, + 0.005665096919983625, + 0.004922183696180582, + 0.003913846332579851, + 0.03479185327887535, + 0.0046909828670322895, + 0.006516351830214262, + 0.01497605536133051, + 0.007646957878023386, + 0.0070641799829900265, + 0.007187801878899336, + 0.003850026521831751, + 0.01359571609646082, + 0.00787986908107996, + 0.00718340277671814, + 0.015394885092973709, + 0.010375653393566608, + 0.01148510817438364, + 0.009531908668577671, + 0.006123205181211233, + 0.010011283680796623, + 0.004776990041136742, + 0.004374825861304998, + 0.02451525256037712, + 0.006138433702290058, + 0.004830449819564819, + 0.0054884981364011765, + 0.0034164173994213343, + 0.023977776989340782, + 0.008504833094775677, + 0.01084903534501791, + 0.0328022725880146, + 0.011321721598505974, + 0.00989046785980463, + 0.010345213115215302, + 0.0065202838741242886, + 0.03021954745054245, + 0.008695870637893677, + 0.009385500103235245, + 0.04193082079291344, + 0.010900769382715225, + 0.006849040277302265, + 0.009133981540799141, + 0.0054695941507816315, + 0.04030170291662216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0021967582870274782 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637439269572496 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637439269572496, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637439269572496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16104111075401306, + 0.1045728549361229, + 0.03503444790840149, + 0.06658695638179779, + 0.09186305105686188, + 0.09707040339708328, + 0.07722741365432739, + 0.06573151051998138, + 0.20687855780124664, + 0.09399369359016418 + ], + "credit_map": [ + 0.009356045164167881, + 0.009408106096088886, + 0.05240839719772339, + 0.012088489718735218, + 0.011522713117301464, + 0.007887826301157475, + 0.006219906732439995, + 0.052149638533592224, + 0.00744142709299922, + 0.0074397786520421505, + 0.03359829634428024, + 0.007210462354123592, + 0.007363807875663042, + 0.00536366505548358, + 0.004323557950556278, + 0.03183186426758766, + 0.0021200012415647507, + 0.0042052376084029675, + 0.008796323090791702, + 0.0033237352035939693, + 0.004263854119926691, + 0.001979242544621229, + 0.002076496835798025, + 0.008269556798040867, + 0.0038601031992584467, + 0.005901374388486147, + 0.017814049497246742, + 0.005634042900055647, + 0.00629469146952033, + 0.004385516047477722, + 0.004101207945495844, + 0.018595976755023003, + 0.005231635645031929, + 0.004002440720796585, + 0.03319159150123596, + 0.00624866085126996, + 0.00354509218595922, + 0.004197191912680864, + 0.0032682695891708136, + 0.03217817842960358, + 0.006136320065706968, + 0.005126455333083868, + 0.032649677246809006, + 0.007502316031605005, + 0.004722630605101585, + 0.005436555482447147, + 0.003740181215107441, + 0.03175627440214157, + 0.004743705037981272, + 0.006499429699033499, + 0.016930703073740005, + 0.008406898938119411, + 0.010991296730935574, + 0.00647749425843358, + 0.005313124507665634, + 0.0178647693246603, + 0.004275596234947443, + 0.003034766297787428, + 0.02236790582537651, + 0.004783072508871555, + 0.0029099888633936644, + 0.003973748534917831, + 0.002854915801435709, + 0.021531522274017334, + 0.01371260266751051, + 0.011480703018605709, + 0.06609727442264557, + 0.015720661729574203, + 0.011348591186106205, + 0.013976016081869602, + 0.008622915484011173, + 0.06591980904340744, + 0.005481223110109568, + 0.005153409205377102, + 0.03153510019183159, + 0.007231925148516893, + 0.004717899486422539, + 0.005310146603733301, + 0.0036288192495703697, + 0.030935188755393028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.004637439269572496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "photorealistic 4K high-fidelity video begins with a static, locked-on close-up, maintaining sharp, unwavering focus on two subjects perfectly centered in the frame. In the foreground, a plump, ripe red tomato rests on a light-grained wooden cutting board; its vibrant scarlet skin is taut, subtly lustrous, and punctuated by tiny, natural imperfections that attest to its authenticity. The board itself displays faint, crisscrossing knife marks from frequent use, its matte texture contrasting sharply with the tomato's smooth surface. A small, partially obscured Pikachu plush toy lies directly behind the tomato, its soft, furry, bright yellow fabric providing textural contrast; its signature long ears, tipped with black, and a round, black button-eye, peek out from behind the fruit. The scene is set in a well-lit kitchen, bathed in a bright, even, diffuse, neutral white light that casts soft, subtle shadows and meticulously defines each object's form. The composition emphasizes surreal physical textures—from the waxy skin of a tomato to the fibrous weave of a plum—and their gentle interplay of shadows across the wooden surface, all contrasted against a clean, neutral white background to create a quiet, focused, and visually rich still life.", + "seed": 4899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_2/index_4482_seed_4899_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": -0.008034891600254923, + "vlm_logprob": -0.008034891600254923, + "vlm_yes_no_margin": 5.9, + "vlm_token_credit": -0.008034891600254923, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006255096290260553 + ], + "opposite_logprob_tokens": [ + -8.131255149841309 + ], + "logprob_reward": -0.006255096290260553, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.006255096290260553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30773410201072693, + 0.07217870652675629, + 0.0459868498146534, + 0.039554037153720856, + 0.0468168631196022, + 0.12306295335292816, + 0.10911434888839722, + 0.10327985882759094, + 0.07944707572460175, + 0.07282523065805435 + ], + "credit_map": [ + 0.05907686427235603, + 0.04573863744735718, + 0.04324386268854141, + 0.020491451025009155, + 0.01977980136871338, + 0.04690723493695259, + 0.016152355819940567, + 0.05634387582540512, + 0.009909757412970066, + 0.012601067312061787, + 0.008853708393871784, + 0.005191449075937271, + 0.005772411823272705, + 0.012231994420289993, + 0.0036584599874913692, + 0.013959857635200024, + 0.005147804971784353, + 0.004993578884750605, + 0.005979891400784254, + 0.005313256289809942, + 0.0038301830645650625, + 0.007741271052509546, + 0.004106405656784773, + 0.008874460123479366, + 0.0056403642520308495, + 0.004830567631870508, + 0.0050471010617911816, + 0.004737053532153368, + 0.0025493258144706488, + 0.0065153539180755615, + 0.002558924723416567, + 0.007675346918404102, + 0.006456031464040279, + 0.00623695133253932, + 0.006225916091352701, + 0.005584882106631994, + 0.00449339346960187, + 0.007195251528173685, + 0.0033682272769510746, + 0.007256209850311279, + 0.019160814583301544, + 0.014260831288993359, + 0.011831504292786121, + 0.0097974743694067, + 0.010526266880333424, + 0.022488342598080635, + 0.0062682670541107655, + 0.028729453682899475, + 0.014092257246375084, + 0.01534211914986372, + 0.009799540042877197, + 0.01130681112408638, + 0.011493988335132599, + 0.01707383058965206, + 0.008625227957963943, + 0.021380571648478508, + 0.015927571803331375, + 0.012158476747572422, + 0.011681915260851383, + 0.008864177390933037, + 0.009725959040224552, + 0.01642511412501335, + 0.006979017052799463, + 0.02151763066649437, + 0.01289766188710928, + 0.011967729777097702, + 0.010167610831558704, + 0.006822046358138323, + 0.007183959241956472, + 0.01242632046341896, + 0.005402373615652323, + 0.012579379603266716, + 0.015478462912142277, + 0.0094657177105546, + 0.011036170646548271, + 0.005611460190266371, + 0.007046028506010771, + 0.010726360604166985, + 0.0031841190066188574, + 0.010276912711560726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.006255096290260553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008266131044365466 + ], + "opposite_logprob_tokens": [ + -7.125826835632324 + ], + "logprob_reward": -0.0008266131044365466, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008266131044365466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18149684369564056, + 0.10385531187057495, + 0.10170981287956238, + 0.10819685459136963, + 0.04318929463624954, + 0.11990346759557724, + 0.13939711451530457, + 0.11438657343387604, + 0.05809975787997246, + 0.029764903709292412 + ], + "credit_map": [ + 0.030380014330148697, + 0.026761777698993683, + 0.019831741228699684, + 0.015370503067970276, + 0.01919289119541645, + 0.02983488142490387, + 0.01297692209482193, + 0.027148138731718063, + 0.017694532871246338, + 0.015688862651586533, + 0.011371267028152943, + 0.010206437669694424, + 0.01046668365597725, + 0.015584754757583141, + 0.0086669297888875, + 0.014175851829349995, + 0.01443460676819086, + 0.01825699955224991, + 0.016036417335271835, + 0.01081047486513853, + 0.007385972887277603, + 0.015735769644379616, + 0.006394215859472752, + 0.01265537366271019, + 0.020479736849665642, + 0.01324924174696207, + 0.013561688363552094, + 0.006458270363509655, + 0.00796778965741396, + 0.019440360367298126, + 0.0037382757291197777, + 0.02330150455236435, + 0.008799279108643532, + 0.006158438045531511, + 0.006296463310718536, + 0.0033392617478966713, + 0.0035919577348977327, + 0.00640520453453064, + 0.002041097031906247, + 0.006557593587785959, + 0.022165510803461075, + 0.016925234347581863, + 0.012772766873240471, + 0.00780845619738102, + 0.008325531147420406, + 0.021562421694397926, + 0.004879151936620474, + 0.025464406237006187, + 0.01972617581486702, + 0.015504796989262104, + 0.02002270147204399, + 0.01594686508178711, + 0.01248121727257967, + 0.020021948963403702, + 0.011003836989402771, + 0.024689601734280586, + 0.019019151106476784, + 0.014657221734523773, + 0.013857577927410603, + 0.011172541417181492, + 0.010264323092997074, + 0.01641080342233181, + 0.008985386230051517, + 0.020019572228193283, + 0.008719881996512413, + 0.0055271633900702, + 0.006600312888622284, + 0.005520826205611229, + 0.004968046676367521, + 0.010510300286114216, + 0.0032589833717793226, + 0.012994242832064629, + 0.004847629461437464, + 0.004376590251922607, + 0.005295770708471537, + 0.002748701022937894, + 0.002565244445577264, + 0.004278365522623062, + 0.0018808540189638734, + 0.0037717544473707676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0008266131044365466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608303189277649 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.01608303189277649, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608303189277649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0976780578494072, + 0.028325797989964485, + 0.047290973365306854, + 0.05392645299434662, + 0.04215908423066139, + 0.19120390713214874, + 0.313698410987854, + 0.19737288355827332, + 0.02099785767495632, + 0.007346541155129671 + ], + "credit_map": [ + 0.008695250377058983, + 0.006887182593345642, + 0.00893772765994072, + 0.009196439757943153, + 0.012174406088888645, + 0.01876308023929596, + 0.009619351476430893, + 0.023404622450470924, + 0.0034468837548047304, + 0.00334040611051023, + 0.003582018194720149, + 0.0027370648458600044, + 0.0027415684890002012, + 0.004946786444634199, + 0.002383202314376831, + 0.005147869233042002, + 0.007626427337527275, + 0.004686873406171799, + 0.006458359770476818, + 0.004314263351261616, + 0.004526824224740267, + 0.007379351183772087, + 0.0034774558153003454, + 0.008821415714919567, + 0.009452742524445057, + 0.009296038188040257, + 0.007396921515464783, + 0.003494501579552889, + 0.003910231404006481, + 0.009149845689535141, + 0.002136407420039177, + 0.009089763276278973, + 0.004300074651837349, + 0.0037385395262390375, + 0.008708775043487549, + 0.0042745862156152725, + 0.0029620046261698008, + 0.007214648183435202, + 0.002266010269522667, + 0.008694443851709366, + 0.03292188420891762, + 0.02250736579298973, + 0.02252396196126938, + 0.016221094876527786, + 0.01490987278521061, + 0.0310690775513649, + 0.0108489403501153, + 0.040201712399721146, + 0.04558740183711052, + 0.03603360429406166, + 0.04115922003984451, + 0.03313485160470009, + 0.02980431728065014, + 0.045388732105493546, + 0.02497260831296444, + 0.05761765316128731, + 0.031034082174301147, + 0.02458544261753559, + 0.021461011841893196, + 0.01937469281256199, + 0.019037755206227303, + 0.029133465141057968, + 0.016710443422198296, + 0.03603599965572357, + 0.0036137118004262447, + 0.0025176252238452435, + 0.0026422564405947924, + 0.0030389311723411083, + 0.0014562319265678525, + 0.002901434199884534, + 0.0017322726780548692, + 0.0030953942332416773, + 0.0011713375570252538, + 0.001133193727582693, + 0.001044902135618031, + 0.0007141211535781622, + 0.000674108974635601, + 0.0011478548403829336, + 0.00044273422099649906, + 0.0010182886617258191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.01608303189277649 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028174726758152246 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028174726758152246, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028174726758152246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24104809761047363, + 0.0962134450674057, + 0.03165410831570625, + 0.07459703832864761, + 0.12419985234737396, + 0.0983998104929924, + 0.15370599925518036, + 0.08705732971429825, + 0.07037585973739624, + 0.02274850569665432 + ], + "credit_map": [ + 0.04176051542162895, + 0.03130027651786804, + 0.03593342378735542, + 0.018263760954141617, + 0.01368759199976921, + 0.039601024240255356, + 0.011415861546993256, + 0.04908564314246178, + 0.013808364979922771, + 0.018748348578810692, + 0.01683688722550869, + 0.006755002774298191, + 0.0053771198727190495, + 0.015337411314249039, + 0.004523556213825941, + 0.014826747588813305, + 0.003511980641633272, + 0.0033901482820510864, + 0.003821999067440629, + 0.0036485837772488594, + 0.0031723103020340204, + 0.005643955431878567, + 0.0022113195154815912, + 0.006253810599446297, + 0.013348981738090515, + 0.012774686329066753, + 0.011886585503816605, + 0.006076309829950333, + 0.006003943271934986, + 0.01134320255368948, + 0.004217592068016529, + 0.008945737965404987, + 0.017455097287893295, + 0.01887771673500538, + 0.01971057988703251, + 0.01319847721606493, + 0.009994464926421642, + 0.018641386181116104, + 0.007885739207267761, + 0.018436400219798088, + 0.01825355552136898, + 0.014464231207966805, + 0.01194053515791893, + 0.006165714003145695, + 0.00622432678937912, + 0.01738486811518669, + 0.004203655757009983, + 0.019762922078371048, + 0.02473517879843712, + 0.019960716366767883, + 0.02325691096484661, + 0.01762392930686474, + 0.01192569825798273, + 0.02157883159816265, + 0.011267634108662605, + 0.023357100784778595, + 0.013659080490469933, + 0.010206951759755611, + 0.01110934466123581, + 0.008082089945673943, + 0.008106792345643044, + 0.012868206948041916, + 0.006898997817188501, + 0.016125861555337906, + 0.012173057533800602, + 0.013226532377302647, + 0.010829921811819077, + 0.00522605748847127, + 0.004583442118018866, + 0.011701682582497597, + 0.0030832067131996155, + 0.009551960043609142, + 0.004210028797388077, + 0.0026457286439836025, + 0.003020287025719881, + 0.00234447093680501, + 0.0017057678196579218, + 0.003330821171402931, + 0.0013001327170059085, + 0.004191266838461161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0028174726758152246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, ripe red tomato visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Pikachu plush toy with bright yellow fabric and black-tipped ears visible behind the tomato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a well-lit kitchen with bright, even, diffuse neutral white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a clean, neutral white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a locked-on close-up framing the two subjects perfectly centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1722698211669922, + 0.07548298686742783, + 0.10975681990385056, + 0.11196956038475037, + 0.10767880827188492, + 0.10037782788276672, + 0.08584503084421158, + 0.06050742045044899, + 0.14287957549095154, + 0.03323215991258621 + ], + "credit_map": [ + 0.03048389032483101, + 0.018947234377264977, + 0.019973210990428925, + 0.012786703184247017, + 0.01425179559737444, + 0.028655124828219414, + 0.009444340132176876, + 0.037727512419223785, + 0.013056013733148575, + 0.009800886735320091, + 0.008574116043746471, + 0.005705136340111494, + 0.0066346656531095505, + 0.012453294359147549, + 0.0041871885769069195, + 0.015071682631969452, + 0.01854800619184971, + 0.022880665957927704, + 0.02300149016082287, + 0.0076747313141822815, + 0.006102383136749268, + 0.01787697896361351, + 0.0043769655749201775, + 0.009295597672462463, + 0.02339170128107071, + 0.018657106906175613, + 0.015678754076361656, + 0.007446241565048695, + 0.008771411143243313, + 0.01852191984653473, + 0.0049621835350990295, + 0.014540230855345726, + 0.02260836772620678, + 0.019380837678909302, + 0.01417431142181158, + 0.007612434681504965, + 0.0066117229871451855, + 0.017710592597723007, + 0.004130449146032333, + 0.015450096689164639, + 0.01644950546324253, + 0.019409680739045143, + 0.013765926472842693, + 0.007521471474319696, + 0.006896892562508583, + 0.017367927357554436, + 0.004181262571364641, + 0.01478516310453415, + 0.015264215879142284, + 0.01772470399737358, + 0.01650361344218254, + 0.006031710654497147, + 0.004825070034712553, + 0.014599758200347424, + 0.0035504410043358803, + 0.007345518097281456, + 0.011287346482276917, + 0.01243559829890728, + 0.011529963463544846, + 0.003992781508713961, + 0.0031343535520136356, + 0.01017866563051939, + 0.002245501382276416, + 0.0057032061740756035, + 0.027959764003753662, + 0.026525020599365234, + 0.022042080760002136, + 0.010593870654702187, + 0.008251998573541641, + 0.023914743214845657, + 0.005692365113645792, + 0.017899727448821068, + 0.003317528869956732, + 0.0037017473950982094, + 0.0034693425986915827, + 0.004966989625245333, + 0.0033719418570399284, + 0.005342213436961174, + 0.0030895969830453396, + 0.0059727998450398445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.014192244037985802 + } + ], + "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/0000025/group_3/index_tensor([8693, 8693])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_3/index_tensor([8693, 8693])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..74bbceebfd41d09e348c8e92a28d53c2a6923813 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_3/index_tensor([8693, 8693])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, static shot captured from a high angle captures a lone figure, seemingly dwarfed against a vast, sparsely planted, gray, rocky wasteland. The golden afterglow of the setting sun sweeps across the landscape from a very low angle, elongating the figure's silhouette into a long, solitary shadow. The figure wears a dark brown, weathered hooded windbreaker, and the weight of a hiking backpack weighs heavily on his shoulders. He pauses, as if gathering his strength, then slowly bends his knees, lowering his center of gravity. His movement is steady and weighted, ultimately resting in a half-crouch on the cracked earth. One dark-gloved hand reaches down, seemingly examining something on the ground. The dominant tones of the image are desolate gray and ochre, tinged with a warm, melancholic golden hue by the setting sun, creating an epic atmosphere of desolation and tranquility, profoundly conveying the insignificance and resilience of individuals in the face of the harshness of nature.", + "seed": 5712, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_3/index_8693_seed_5712_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": -0.02079082587733865, + "vlm_logprob": -0.02079082587733865, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02079082587733865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043284207582473755 + ], + "opposite_logprob_tokens": [ + -3.2932841777801514 + ], + "logprob_reward": -0.043284207582473755, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.043284207582473755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0947972759604454, + 0.03916211426258087, + 0.03771796450018883, + 0.04207778722047806, + 0.607250988483429, + 0.046734701842069626, + 0.027456797659397125, + 0.017457682639360428, + 0.03196442872285843, + 0.05538024753332138 + ], + "credit_map": [ + 0.006891865748912096, + 0.011342660523951054, + 0.018939606845378876, + 0.03021177090704441, + 0.005604411941021681, + 0.008384606800973415, + 0.005776962731033564, + 0.0076453895308077335, + 0.00384068232960999, + 0.004567551426589489, + 0.006372200325131416, + 0.010147941298782825, + 0.003299441421404481, + 0.004083671607077122, + 0.002538467524573207, + 0.004312159027904272, + 0.002634360920637846, + 0.005415065214037895, + 0.006981263402849436, + 0.00863322988152504, + 0.0022504222579300404, + 0.003704605856910348, + 0.002265087328851223, + 0.005833927076309919, + 0.0037248937878757715, + 0.0055681513622403145, + 0.008771686814725399, + 0.009549456648528576, + 0.0036915335804224014, + 0.003963034134358168, + 0.0032921992242336273, + 0.0035168330650776625, + 0.027534373104572296, + 0.018512744456529617, + 0.2510055601596832, + 0.18015320599079132, + 0.010238442569971085, + 0.05575352534651756, + 0.04623918980360031, + 0.01781398616731167, + 0.004730438347905874, + 0.004964965395629406, + 0.004814659710973501, + 0.0060160658322274685, + 0.005948472768068314, + 0.008157262578606606, + 0.005338404793292284, + 0.006764434278011322, + 0.006687782239168882, + 0.0038285632617771626, + 0.0028930241242051125, + 0.002642341423779726, + 0.0021300255320966244, + 0.005340837873518467, + 0.0016251954948529601, + 0.0023090271279215813, + 0.0032597959507256746, + 0.002441957825794816, + 0.00207356084138155, + 0.0015414803056046367, + 0.0018414977239444852, + 0.002792168641462922, + 0.0015569871757179499, + 0.00195023522246629, + 0.004516112618148327, + 0.0026186362374573946, + 0.004319474101066589, + 0.0037863943725824356, + 0.0032871365547180176, + 0.004529823083430529, + 0.003520683152601123, + 0.005386169534176588, + 0.01092439703643322, + 0.004802134819328785, + 0.008629969321191311, + 0.0031581842340528965, + 0.0031006229110062122, + 0.012532321736216545, + 0.003550565103068948, + 0.008682050742208958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.043284207582473755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560899369418621 + ], + "opposite_logprob_tokens": [ + -4.387560844421387 + ], + "logprob_reward": -0.012560899369418621, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560899369418621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0913165956735611, + 0.03298293054103851, + 0.04655242711305618, + 0.2158999890089035, + 0.3552996516227722, + 0.050368230789899826, + 0.05295901000499725, + 0.03839118406176567, + 0.07158900052309036, + 0.04464099556207657 + ], + "credit_map": [ + 0.004822819959372282, + 0.007685570511966944, + 0.01927713118493557, + 0.02670392021536827, + 0.009793723933398724, + 0.0055966442450881, + 0.008411954157054424, + 0.009024827741086483, + 0.002669592620804906, + 0.0026352673303335905, + 0.005164538510143757, + 0.007017059251666069, + 0.004536159336566925, + 0.004019934684038162, + 0.0030816341750323772, + 0.003858742071315646, + 0.0027239706832915545, + 0.005077491980046034, + 0.0074967313557863235, + 0.010331249795854092, + 0.005170435179024935, + 0.00360029935836792, + 0.005257888697087765, + 0.006894361227750778, + 0.009398031048476696, + 0.00870560109615326, + 0.08348463475704193, + 0.08289293199777603, + 0.006136201787739992, + 0.005911149084568024, + 0.010278568603098392, + 0.009092869237065315, + 0.015693537890911102, + 0.012935321778059006, + 0.14401207864284515, + 0.10434762388467789, + 0.007021732162684202, + 0.03419971838593483, + 0.02612150087952614, + 0.010968147777020931, + 0.004142952151596546, + 0.006741625722497702, + 0.005075195804238319, + 0.008906354196369648, + 0.007602576166391373, + 0.00515547301620245, + 0.006890268996357918, + 0.005853783339262009, + 0.004524391144514084, + 0.0069240061566233635, + 0.007273529190570116, + 0.009512742049992085, + 0.005821623839437962, + 0.005163746420294046, + 0.005648562218993902, + 0.008090409450232983, + 0.008532887324690819, + 0.004173454828560352, + 0.0044095623306930065, + 0.0032897351775318384, + 0.0034280302934348583, + 0.008067237213253975, + 0.003281658049672842, + 0.0032086188439279795, + 0.008325819857418537, + 0.005663126707077026, + 0.010027594864368439, + 0.008139846846461296, + 0.008282041177153587, + 0.010376554913818836, + 0.008022282272577286, + 0.012751733884215355, + 0.006695061456412077, + 0.003341678762808442, + 0.0054659005254507065, + 0.005081078037619591, + 0.005326634272933006, + 0.007676764391362667, + 0.004846926312893629, + 0.006206950172781944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.012560899369418621 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099181387573481 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099181387573481, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099181387573481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04939740523695946, + 0.01660938747227192, + 0.031612150371074677, + 0.22073908150196075, + 0.44113820791244507, + 0.04555748403072357, + 0.04230312630534172, + 0.09754475951194763, + 0.02940688468515873, + 0.025691471993923187 + ], + "credit_map": [ + 0.0030185584910213947, + 0.0046966602094471455, + 0.010043170303106308, + 0.014560762792825699, + 0.004246775526553392, + 0.002754759043455124, + 0.005601389799267054, + 0.004475334659218788, + 0.0010878435568884015, + 0.0014563254080712795, + 0.0027236538007855415, + 0.003958429209887981, + 0.001878974144347012, + 0.0013960537035018206, + 0.0021619752515107393, + 0.001946133910678327, + 0.001999111147597432, + 0.002868414158001542, + 0.005249944515526295, + 0.005712603218853474, + 0.0039504398591816425, + 0.002707497449591756, + 0.004002729430794716, + 0.0051214140839874744, + 0.009459172375500202, + 0.0279682744294405, + 0.06806140393018723, + 0.08368580788373947, + 0.005440401844680309, + 0.010017995722591877, + 0.008653713390231133, + 0.0074523393996059895, + 0.019690876826643944, + 0.014644347131252289, + 0.17901399731636047, + 0.12332791090011597, + 0.009193487465381622, + 0.047560106962919235, + 0.033413808792829514, + 0.014293668791651726, + 0.005233990494161844, + 0.006011850666254759, + 0.004996119067072868, + 0.0036949748173356056, + 0.007170709315687418, + 0.005533667281270027, + 0.006247658748179674, + 0.006668516434729099, + 0.003997781313955784, + 0.005760685540735722, + 0.004758196417242289, + 0.0062807356007397175, + 0.004507963079959154, + 0.003292808076366782, + 0.004631286021322012, + 0.009073673747479916, + 0.01558948028832674, + 0.008321181870996952, + 0.011132391169667244, + 0.009439514018595219, + 0.010254441760480404, + 0.017388857901096344, + 0.008737674914300442, + 0.01668122410774231, + 0.005668441765010357, + 0.0028530326671898365, + 0.004528141114860773, + 0.003441490465775132, + 0.0029417939949780703, + 0.0027430839836597443, + 0.003135567530989647, + 0.004095335025340319, + 0.0027232440188527107, + 0.0027515594847500324, + 0.0043014236725866795, + 0.0024319987278431654, + 0.0030056277755647898, + 0.003415794111788273, + 0.003103204071521759, + 0.003958621993660927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.004099181387573481 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298132486641407 + ], + "opposite_logprob_tokens": [ + -3.529813289642334 + ], + "logprob_reward": -0.0298132486641407, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298132486641407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11145444959402084, + 0.04172448068857193, + 0.04074229672551155, + 0.20907026529312134, + 0.17552706599235535, + 0.07539468258619308, + 0.0673757791519165, + 0.09124474227428436, + 0.15555089712142944, + 0.03191527724266052 + ], + "credit_map": [ + 0.005982471164315939, + 0.00923672690987587, + 0.019322169944643974, + 0.022169779986143112, + 0.015822719782590866, + 0.006636549718677998, + 0.013538827188313007, + 0.018745196983218193, + 0.002185796620324254, + 0.003433321136981249, + 0.008187436498701572, + 0.012728308327496052, + 0.0033408780582249165, + 0.003289477899670601, + 0.0035469657741487026, + 0.005012294743210077, + 0.0027389151509851217, + 0.005774750839918852, + 0.007084679789841175, + 0.006797045934945345, + 0.004172805696725845, + 0.004489172715693712, + 0.003030696650967002, + 0.0066542294807732105, + 0.009273869916796684, + 0.02414431981742382, + 0.06524623930454254, + 0.07708706706762314, + 0.005860643461346626, + 0.009737929329276085, + 0.00859930831938982, + 0.009120894595980644, + 0.008451358415186405, + 0.00779118062928319, + 0.06740613281726837, + 0.04155746102333069, + 0.005336700472980738, + 0.025598948821425438, + 0.013881571590900421, + 0.005503703840076923, + 0.007109326776117086, + 0.00853680819272995, + 0.008082440122961998, + 0.007502892054617405, + 0.010965205729007721, + 0.011322920210659504, + 0.009252410382032394, + 0.012622681446373463, + 0.0077772182412445545, + 0.006145985797047615, + 0.00744773680344224, + 0.007327268831431866, + 0.008944081142544746, + 0.006962626241147518, + 0.008224059827625751, + 0.014546805992722511, + 0.013318384066224098, + 0.008394732140004635, + 0.011835233308374882, + 0.008373935706913471, + 0.01157090999186039, + 0.017766444012522697, + 0.008324304595589638, + 0.011660791002213955, + 0.01921098120510578, + 0.008453339338302612, + 0.02584398351609707, + 0.02060084603726864, + 0.011790950782597065, + 0.018245652318000793, + 0.018989667296409607, + 0.032415490597486496, + 0.003330365987494588, + 0.003322882577776909, + 0.004100274760276079, + 0.00444081611931324, + 0.004433640278875828, + 0.0046140109188854694, + 0.0036783446557819843, + 0.003994944039732218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.0298132486641407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196592383086681 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196592383086681, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196592383086681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06951885670423508, + 0.03196660056710243, + 0.021303217858076096, + 0.26451432704925537, + 0.27904990315437317, + 0.1066613644361496, + 0.035471655428409576, + 0.11747254431247711, + 0.054694097489118576, + 0.019347388297319412 + ], + "credit_map": [ + 0.005718410015106201, + 0.008315656334161758, + 0.012729251757264137, + 0.02133314311504364, + 0.003321118885651231, + 0.007183133624494076, + 0.004067010246217251, + 0.00685113063082099, + 0.003080734983086586, + 0.004430032800883055, + 0.005361291579902172, + 0.007457314990460873, + 0.0027730981819331646, + 0.003306945553049445, + 0.0019468587124720216, + 0.0036103210877627134, + 0.001915035885758698, + 0.003909206949174404, + 0.002948835026472807, + 0.003281797282397747, + 0.0022292097564786673, + 0.002440967131406069, + 0.0014485068386420608, + 0.003129658754914999, + 0.011017957702279091, + 0.025466276332736015, + 0.0995858907699585, + 0.09227831661701202, + 0.005460274871438742, + 0.009752810932695866, + 0.011211858130991459, + 0.009740937501192093, + 0.012560793198645115, + 0.009214991703629494, + 0.11803359538316727, + 0.09047196060419083, + 0.005552371963858604, + 0.014167080633342266, + 0.02105521783232689, + 0.007993906736373901, + 0.014776414260268211, + 0.011262436397373676, + 0.010460993275046349, + 0.013786766678094864, + 0.007786807604134083, + 0.025276822969317436, + 0.008000987581908703, + 0.015310130082070827, + 0.00550839863717556, + 0.004200023133307695, + 0.003921751398593187, + 0.0026611967477947474, + 0.0029415623284876347, + 0.008528435602784157, + 0.0025637769140303135, + 0.005146510899066925, + 0.021391766145825386, + 0.011061569675803185, + 0.01397618930786848, + 0.010164729319512844, + 0.010331584140658379, + 0.02532111294567585, + 0.008153271861374378, + 0.017072321847081184, + 0.006116108503192663, + 0.0033566069323569536, + 0.008069072850048542, + 0.007083627860993147, + 0.005147264339029789, + 0.0060988194309175014, + 0.007631787098944187, + 0.011190812103450298, + 0.002594804158434272, + 0.0025232622865587473, + 0.0029564700089395046, + 0.0020447904244065285, + 0.0020921044051647186, + 0.0019827166106551886, + 0.0018250729190185666, + 0.003328167600557208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.014196592383086681 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, static shot captured from a high angle captures a lone figure, seemingly dwarfed against a vast, sparsely planted, gray, rocky wasteland. The golden afterglow of the setting sun sweeps across the landscape from a very low angle, elongating the figure's silhouette into a long, solitary shadow. The figure wears a dark brown, weathered hooded windbreaker, and the weight of a hiking backpack weighs heavily on his shoulders. He pauses, as if gathering his strength, then slowly bends his knees, lowering his center of gravity. His movement is steady and weighted, ultimately resting in a half-crouch on the cracked earth. One dark-gloved hand reaches down, seemingly examining something on the ground. The dominant tones of the image are desolate gray and ochre, tinged with a warm, melancholic golden hue by the setting sun, creating an epic atmosphere of desolation and tranquility, profoundly conveying the insignificance and resilience of individuals in the face of the harshness of nature.", + "seed": 5713, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_3/index_8693_seed_5713_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": -0.03405087050050497, + "vlm_logprob": -0.03405087050050497, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03405087050050497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06662493944168091 + ], + "opposite_logprob_tokens": [ + -2.816624879837036 + ], + "logprob_reward": -0.06662493944168091, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06662493944168091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06500200927257538, + 0.15118418633937836, + 0.11465194821357727, + 0.2332874834537506, + 0.1351851373910904, + 0.04973221570253372, + 0.043725136667490005, + 0.022031426429748535, + 0.049820203334093094, + 0.13538019359111786 + ], + "credit_map": [ + 0.007548173889517784, + 0.008860242553055286, + 0.009141364134848118, + 0.006478354800492525, + 0.006163043901324272, + 0.007893622852861881, + 0.007064726669341326, + 0.011852487921714783, + 0.01628112606704235, + 0.015418031252920628, + 0.030339866876602173, + 0.015769125893712044, + 0.013157250359654427, + 0.022410418838262558, + 0.013214454986155033, + 0.024593938142061234, + 0.01430428121238947, + 0.009625920094549656, + 0.016239585354924202, + 0.015614937990903854, + 0.010475564748048782, + 0.022040266543626785, + 0.01184881292283535, + 0.014502584002912045, + 0.025517122820019722, + 0.017866289243102074, + 0.03829362615942955, + 0.017593245953321457, + 0.022670743986964226, + 0.04794324189424515, + 0.019344042986631393, + 0.04405917972326279, + 0.012503184378147125, + 0.009857247583568096, + 0.021677419543266296, + 0.01773081161081791, + 0.010755837894976139, + 0.020216884091496468, + 0.015807684510946274, + 0.026636065915226936, + 0.004361446481198072, + 0.005400737281888723, + 0.006661008112132549, + 0.006317206658422947, + 0.0049268873408436775, + 0.007072682958096266, + 0.005540875717997551, + 0.009451379999518394, + 0.004551958292722702, + 0.005171008873730898, + 0.007062459364533424, + 0.006142937578260899, + 0.004627816844731569, + 0.0033602092880755663, + 0.006812951993197203, + 0.005995794665068388, + 0.0027662974316626787, + 0.0028873297851532698, + 0.0029892069287598133, + 0.0023770860861986876, + 0.0024149916134774685, + 0.0022545717656612396, + 0.0024787229485809803, + 0.0038632212672382593, + 0.003336291527375579, + 0.002898386912420392, + 0.01479512918740511, + 0.0037124191876500845, + 0.0039011649787425995, + 0.004290510434657335, + 0.004025442060083151, + 0.012860867194831371, + 0.01366208866238594, + 0.01335348654538393, + 0.019046824425458908, + 0.011268191039562225, + 0.013899742625653744, + 0.03636171296238899, + 0.011461185291409492, + 0.01632697507739067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.06662493944168091 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820462942123413 + ], + "opposite_logprob_tokens": [ + -4.018204689025879 + ], + "logprob_reward": -0.01820462942123413, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820462942123413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07582662999629974, + 0.14699608087539673, + 0.07788500934839249, + 0.07133965194225311, + 0.11483129858970642, + 0.0278970617800951, + 0.07472604513168335, + 0.05011344328522682, + 0.17302924394607544, + 0.18735550343990326 + ], + "credit_map": [ + 0.007845993153750896, + 0.009692567400634289, + 0.011097034439444542, + 0.008594658225774765, + 0.007027381099760532, + 0.009092948399484158, + 0.00930892862379551, + 0.013167114928364754, + 0.012901630252599716, + 0.01453583873808384, + 0.031056677922606468, + 0.01987639255821705, + 0.01136529166251421, + 0.015580635517835617, + 0.014796876348555088, + 0.026882730424404144, + 0.008821072056889534, + 0.008894041180610657, + 0.010807434096932411, + 0.012959854677319527, + 0.00586944492533803, + 0.009555095806717873, + 0.008764585480093956, + 0.012213480658829212, + 0.008201702497899532, + 0.00674587395042181, + 0.011137932538986206, + 0.006734202150255442, + 0.00733180670067668, + 0.013320374302566051, + 0.005583825055509806, + 0.012283924967050552, + 0.011317577213048935, + 0.007589015178382397, + 0.0185052789747715, + 0.014219752512872219, + 0.00764784449711442, + 0.01780155673623085, + 0.014343773946166039, + 0.023406503722071648, + 0.0029120587278157473, + 0.0029662768356502056, + 0.004442770499736071, + 0.0035181737039238214, + 0.002838490530848503, + 0.0033992633689194918, + 0.003294094232842326, + 0.004525934811681509, + 0.007654023822396994, + 0.00864822044968605, + 0.008896560408174992, + 0.007998980581760406, + 0.006993161980062723, + 0.0074718790128827095, + 0.012038241140544415, + 0.015024976804852486, + 0.004846390802413225, + 0.005439793691039085, + 0.008149368688464165, + 0.005452461075037718, + 0.005635937675833702, + 0.005483626388013363, + 0.006228716112673283, + 0.008877148851752281, + 0.008404611609876156, + 0.008855319581925869, + 0.05109088867902756, + 0.014636664651334286, + 0.013049326837062836, + 0.017059218138456345, + 0.013653488829731941, + 0.046279728412628174, + 0.018401125445961952, + 0.016935253515839577, + 0.025990145280957222, + 0.01794453151524067, + 0.01761307381093502, + 0.048844750970602036, + 0.0173908993601799, + 0.024235716089606285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.01820462942123413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648676797747612 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648676797747612, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648676797747612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04169364646077156, + 0.047767218202352524, + 0.04789360612630844, + 0.03695359081029892, + 0.06510249525308609, + 0.038428883999586105, + 0.23042522370815277, + 0.08345472812652588, + 0.230631485581398, + 0.17764915525913239 + ], + "credit_map": [ + 0.0044457572512328625, + 0.005224339663982391, + 0.005767854396253824, + 0.004434035625308752, + 0.0042404793202877045, + 0.004935980774462223, + 0.0052351681515574455, + 0.007410029415041208, + 0.0048264325596392155, + 0.004252872429788113, + 0.00836468767374754, + 0.0066445618867874146, + 0.003889553714543581, + 0.006785945501178503, + 0.0055703213438391685, + 0.0074328407645225525, + 0.00581315765157342, + 0.004202412907034159, + 0.006841059774160385, + 0.008628764189779758, + 0.003310150932520628, + 0.0067551820538938046, + 0.005446716211736202, + 0.006896157283335924, + 0.0038081782404333353, + 0.004546965938061476, + 0.0059643336571753025, + 0.004691327456384897, + 0.0034870679955929518, + 0.005423924885690212, + 0.004250206518918276, + 0.004781587515026331, + 0.004775513429194689, + 0.00515720946714282, + 0.010181005112826824, + 0.009616739116609097, + 0.005183279514312744, + 0.008645490743219852, + 0.009414526633918285, + 0.012128731235861778, + 0.00285342033021152, + 0.0034697067458182573, + 0.00626306002959609, + 0.006533259525895119, + 0.003415594808757305, + 0.0032307703513652086, + 0.006797241047024727, + 0.00586582999676466, + 0.01177448220551014, + 0.02232065424323082, + 0.03190574049949646, + 0.04127766937017441, + 0.023736756294965744, + 0.014488760381937027, + 0.05143062397837639, + 0.03349050134420395, + 0.006907755509018898, + 0.007371086627244949, + 0.013479703105986118, + 0.011520497500896454, + 0.009040314704179764, + 0.006145792081952095, + 0.017995506525039673, + 0.010994073934853077, + 0.009671847335994244, + 0.008059537038207054, + 0.08166126161813736, + 0.017522074282169342, + 0.008460789918899536, + 0.01315908320248127, + 0.016888398677110672, + 0.07520848512649536, + 0.014893833547830582, + 0.01320957113057375, + 0.02692572958767414, + 0.023014195263385773, + 0.014072371646761894, + 0.03378940373659134, + 0.022853828966617584, + 0.028890199959278107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.008648676797747612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003837823867798 + ], + "opposite_logprob_tokens": [ + -2.695038318634033 + ], + "logprob_reward": -0.07003837823867798, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003837823867798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037466663867235184, + 0.07781875878572464, + 0.0562090165913105, + 0.08716627210378647, + 0.1961164027452469, + 0.027190109714865685, + 0.07048841565847397, + 0.04927181452512741, + 0.2668073773384094, + 0.13146516680717468 + ], + "credit_map": [ + 0.0028809739742428064, + 0.0055541107431054115, + 0.005096236243844032, + 0.0047372616827487946, + 0.005067494232207537, + 0.004068507347255945, + 0.004173364955931902, + 0.005888714920729399, + 0.007228674367070198, + 0.007747076451778412, + 0.018006233498454094, + 0.008442546240985394, + 0.005817870143800974, + 0.009185240603983402, + 0.006652158685028553, + 0.014738956466317177, + 0.007238323334604502, + 0.006130869500339031, + 0.007710296660661697, + 0.008955107070505619, + 0.00460014445707202, + 0.00810074433684349, + 0.00583604397252202, + 0.007637485396116972, + 0.00862374622374773, + 0.00896333809942007, + 0.013204831629991531, + 0.010333174839615822, + 0.007099573500454426, + 0.012852869927883148, + 0.00942423939704895, + 0.016664497554302216, + 0.021196305751800537, + 0.012669658288359642, + 0.032450538128614426, + 0.020591216161847115, + 0.01282290369272232, + 0.031208330765366554, + 0.023377060890197754, + 0.0418003685772419, + 0.0028848499059677124, + 0.003655867651104927, + 0.004274892620742321, + 0.003311112755909562, + 0.0022333383094519377, + 0.0037267461884766817, + 0.0027723493985831738, + 0.004330951254814863, + 0.010994216427206993, + 0.010458779521286488, + 0.008421504870057106, + 0.007820434868335724, + 0.009575257077813148, + 0.0054216571152210236, + 0.008183418773114681, + 0.009613141417503357, + 0.004078899044543505, + 0.004173655528575182, + 0.007097131572663784, + 0.007120430003851652, + 0.005914515350013971, + 0.0059670922346413136, + 0.008483633399009705, + 0.006436449941247702, + 0.011129473336040974, + 0.0065423594787716866, + 0.10227872431278229, + 0.018201956525444984, + 0.007294865325093269, + 0.008738404139876366, + 0.018626978620886803, + 0.0939946249127388, + 0.013955180533230305, + 0.011582287028431892, + 0.018322885036468506, + 0.010550313629209995, + 0.012839964590966702, + 0.03813069313764572, + 0.010037951171398163, + 0.016045883297920227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.07003837823867798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737728603184223 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737728603184223, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737728603184223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11240322887897491, + 0.157785564661026, + 0.04122495651245117, + 0.18878892064094543, + 0.1069045215845108, + 0.03597160801291466, + 0.04796219617128372, + 0.0771295428276062, + 0.08740349113941193, + 0.14442603290081024 + ], + "credit_map": [ + 0.012658793479204178, + 0.01073254644870758, + 0.01730060763657093, + 0.012053617276251316, + 0.008361008018255234, + 0.016834497451782227, + 0.011355707421898842, + 0.023106444627046585, + 0.017985418438911438, + 0.014287438243627548, + 0.03470291197299957, + 0.012941744178533554, + 0.00991938915103674, + 0.019483646377921104, + 0.013114487752318382, + 0.035350508987903595, + 0.004356300458312035, + 0.003616946516558528, + 0.007593472022563219, + 0.005084095988422632, + 0.0038399670738726854, + 0.00496637960895896, + 0.0046524242497980595, + 0.007115368731319904, + 0.02003585174679756, + 0.013594618998467922, + 0.0305185504257679, + 0.015583012253046036, + 0.01715926080942154, + 0.038585834205150604, + 0.017628196626901627, + 0.035683587193489075, + 0.015956535935401917, + 0.009482989087700844, + 0.01679404079914093, + 0.009444162249565125, + 0.00778539152815938, + 0.017398636788129807, + 0.009987642988562584, + 0.02005511149764061, + 0.004468634724617004, + 0.003819352714344859, + 0.005245069973170757, + 0.003596619237214327, + 0.0034488970413804054, + 0.005473956000059843, + 0.0038504907861351967, + 0.006068585440516472, + 0.003712120931595564, + 0.005792245734483004, + 0.005841545760631561, + 0.006147812586277723, + 0.007041958626359701, + 0.0050535802729427814, + 0.007818094454705715, + 0.006554830353707075, + 0.006437115836888552, + 0.007051342166960239, + 0.011452335864305496, + 0.00805025827139616, + 0.008732899092137814, + 0.010182926431298256, + 0.011135471053421497, + 0.014087183400988579, + 0.005117946770042181, + 0.004123331047594547, + 0.02913169376552105, + 0.0061033726669847965, + 0.004889672622084618, + 0.00464610056951642, + 0.0060228463262319565, + 0.027368515729904175, + 0.010248370468616486, + 0.00994171854108572, + 0.02557401917874813, + 0.017773132771253586, + 0.011897827498614788, + 0.02419857308268547, + 0.017977623268961906, + 0.02681475132703781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.006737728603184223 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, static shot captured from a high angle captures a lone figure, seemingly dwarfed against a vast, sparsely planted, gray, rocky wasteland. The golden afterglow of the setting sun sweeps across the landscape from a very low angle, elongating the figure's silhouette into a long, solitary shadow. The figure wears a dark brown, weathered hooded windbreaker, and the weight of a hiking backpack weighs heavily on his shoulders. He pauses, as if gathering his strength, then slowly bends his knees, lowering his center of gravity. His movement is steady and weighted, ultimately resting in a half-crouch on the cracked earth. One dark-gloved hand reaches down, seemingly examining something on the ground. The dominant tones of the image are desolate gray and ochre, tinged with a warm, melancholic golden hue by the setting sun, creating an epic atmosphere of desolation and tranquility, profoundly conveying the insignificance and resilience of individuals in the face of the harshness of nature.", + "seed": 5714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_3/index_8693_seed_5714_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": -0.02134949816390872, + "vlm_logprob": -0.02134949816390872, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02134949816390872, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01573210395872593 + ], + "opposite_logprob_tokens": [ + -5.390732288360596 + ], + "logprob_reward": -0.01573210395872593, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01573210395872593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09268258512020111, + 0.05829761549830437, + 0.07243935763835907, + 0.09303488582372665, + 0.31174618005752563, + 0.05354036018252373, + 0.0648164302110672, + 0.09435401856899261, + 0.09017401933670044, + 0.068914495408535 + ], + "credit_map": [ + 0.015478238463401794, + 0.00972752459347248, + 0.015976734459400177, + 0.006873672362416983, + 0.010713881812989712, + 0.011389707215130329, + 0.00754990428686142, + 0.01497292798012495, + 0.010412164032459259, + 0.010967781767249107, + 0.00920991599559784, + 0.004057115409523249, + 0.007907379418611526, + 0.007116200868040323, + 0.003661793889477849, + 0.004965271335095167, + 0.006964644882827997, + 0.016059424728155136, + 0.008387383073568344, + 0.006587672512978315, + 0.009487991221249104, + 0.008074076846241951, + 0.008475777693092823, + 0.0084023866802454, + 0.01033773459494114, + 0.015746960416436195, + 0.011191872879862785, + 0.009511117823421955, + 0.014836495742201805, + 0.009578459896147251, + 0.010590697638690472, + 0.011241565458476543, + 0.029403692111372948, + 0.04817532002925873, + 0.048309698700904846, + 0.036445893347263336, + 0.03972778096795082, + 0.03436838462948799, + 0.03393520042300224, + 0.04138026759028435, + 0.007749058306217194, + 0.00850722100585699, + 0.006714289542287588, + 0.0048168618232011795, + 0.00560386385768652, + 0.006537813227623701, + 0.006331346929073334, + 0.0072799064218997955, + 0.007260700222104788, + 0.00947400089353323, + 0.009214621968567371, + 0.009744343347847462, + 0.008171796798706055, + 0.006665871478617191, + 0.008021700195968151, + 0.0062634083442389965, + 0.021068794652819633, + 0.011242511682212353, + 0.012396097183227539, + 0.008608079515397549, + 0.008603863418102264, + 0.011118520982563496, + 0.0088728629052639, + 0.012443293817341328, + 0.013586641289293766, + 0.008400180377066135, + 0.012462177313864231, + 0.008967647328972816, + 0.010355046018958092, + 0.010385262779891491, + 0.012347519397735596, + 0.013669558800756931, + 0.011150500737130642, + 0.010475809685885906, + 0.009954309090971947, + 0.005845991894602776, + 0.008861585520207882, + 0.009195275604724884, + 0.005392150953412056, + 0.008038878440856934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.01573210395872593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599007606506348 + ], + "opposite_logprob_tokens": [ + -3.895598888397217 + ], + "logprob_reward": -0.020599007606506348, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599007606506348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19073684513568878, + 0.14246568083763123, + 0.09786516427993774, + 0.1540478616952896, + 0.09389250725507736, + 0.058857448399066925, + 0.03840065002441406, + 0.05182281881570816, + 0.09698987752199173, + 0.07492116093635559 + ], + "credit_map": [ + 0.02409346215426922, + 0.037574101239442825, + 0.018024178221821785, + 0.024440262466669083, + 0.03512364998459816, + 0.020472398027777672, + 0.014827568084001541, + 0.016181213781237602, + 0.021042807027697563, + 0.03764071688055992, + 0.010145708918571472, + 0.014231162145733833, + 0.023628996685147285, + 0.017533816397190094, + 0.006876557134091854, + 0.011365912854671478, + 0.010662592947483063, + 0.018837125971913338, + 0.012642555870115757, + 0.010085822083055973, + 0.01379561610519886, + 0.009809262119233608, + 0.010868764482438564, + 0.01116341631859541, + 0.0171158816665411, + 0.021132037043571472, + 0.02179548889398575, + 0.022050706669688225, + 0.019704272970557213, + 0.01736156828701496, + 0.01984640210866928, + 0.015041487291455269, + 0.011184279806911945, + 0.01621740683913231, + 0.007700835354626179, + 0.012546026147902012, + 0.01666371151804924, + 0.00917976163327694, + 0.013824161142110825, + 0.006576321087777615, + 0.006332607474178076, + 0.010928357020020485, + 0.004852446727454662, + 0.0072886948473751545, + 0.009787562303245068, + 0.006872590631246567, + 0.007808963768184185, + 0.004986221436411142, + 0.004294156562536955, + 0.005719358567148447, + 0.006331916898488998, + 0.004337103571742773, + 0.004785983357578516, + 0.005021465476602316, + 0.003694116370752454, + 0.004216543398797512, + 0.011877148412168026, + 0.006699864752590656, + 0.006251292768865824, + 0.0048250495456159115, + 0.007104924414306879, + 0.005926910322159529, + 0.005216515157371759, + 0.003921110183000565, + 0.01240610796958208, + 0.015791447833180428, + 0.011541389860212803, + 0.009240387938916683, + 0.012944645248353481, + 0.009736760519444942, + 0.013805997557938099, + 0.011523127555847168, + 0.01022722851485014, + 0.010775325819849968, + 0.008039002306759357, + 0.007294712588191032, + 0.013658535666763783, + 0.007997633889317513, + 0.01067242119461298, + 0.006256298627704382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.020599007606506348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059547401033341885 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.0059547401033341885, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059547401033341885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06663664430379868, + 0.03975009173154831, + 0.08775608986616135, + 0.132186621427536, + 0.4177820384502411, + 0.06283912807703018, + 0.022386472672224045, + 0.07897105067968369, + 0.056306738406419754, + 0.03538507595658302 + ], + "credit_map": [ + 0.008260312490165234, + 0.011392293497920036, + 0.006664929445832968, + 0.011238808743655682, + 0.01201245654374361, + 0.007542253006249666, + 0.004624575842171907, + 0.004901018459349871, + 0.004399191588163376, + 0.008713304065167904, + 0.005474484991282225, + 0.004084080923348665, + 0.006719553843140602, + 0.0045468430034816265, + 0.00255560502409935, + 0.0032570299226790667, + 0.008611130528151989, + 0.016262710094451904, + 0.009229842573404312, + 0.009919367730617523, + 0.014698692597448826, + 0.00931901391595602, + 0.012556776404380798, + 0.007158547639846802, + 0.013272439129650593, + 0.027234021574258804, + 0.01706465147435665, + 0.020122099667787552, + 0.014711146242916584, + 0.012670738622546196, + 0.01595095731317997, + 0.011160559020936489, + 0.04057487100362778, + 0.06484270840883255, + 0.05741220340132713, + 0.04743262380361557, + 0.05733828619122505, + 0.04485010728240013, + 0.05478782579302788, + 0.05054343491792679, + 0.006784795317798853, + 0.011135313659906387, + 0.006329614669084549, + 0.006068820133805275, + 0.009544750675559044, + 0.006143446546047926, + 0.011259742081165314, + 0.005572641734033823, + 0.0023727722000330687, + 0.0034663889091461897, + 0.0026460401713848114, + 0.0034337425604462624, + 0.0029636870604008436, + 0.002239110879600048, + 0.003284117439761758, + 0.0019806153140962124, + 0.009301621466875076, + 0.009200669825077057, + 0.0066230883821845055, + 0.006925287656486034, + 0.01883971504867077, + 0.007624991703778505, + 0.012446332722902298, + 0.008009339682757854, + 0.005839212331920862, + 0.006602402776479721, + 0.00804127473384142, + 0.007786630652844906, + 0.00586547190323472, + 0.006351441144943237, + 0.008767693303525448, + 0.007052609696984291, + 0.005125490948557854, + 0.005441318266093731, + 0.003969122190028429, + 0.003656096290796995, + 0.005928275175392628, + 0.003506316337734461, + 0.004694889299571514, + 0.0030635674484074116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.0059547401033341885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811439871788025 + ], + "opposite_logprob_tokens": [ + -3.288114309310913 + ], + "logprob_reward": -0.03811439871788025, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811439871788025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09490475803613663, + 0.03998708352446556, + 0.05977875366806984, + 0.12795193493366241, + 0.3962251842021942, + 0.06310609728097916, + 0.026783308014273643, + 0.04230525344610214, + 0.1037733405828476, + 0.04518428072333336 + ], + "credit_map": [ + 0.014137452468276024, + 0.016090838238596916, + 0.00968267023563385, + 0.011476145125925541, + 0.017238065600395203, + 0.010838890448212624, + 0.007931660860776901, + 0.00750903831794858, + 0.0053313360549509525, + 0.009820334613323212, + 0.005377257242798805, + 0.004090387839823961, + 0.004556535743176937, + 0.003997975029051304, + 0.002666809828951955, + 0.004146449267864227, + 0.005862982012331486, + 0.011888793669641018, + 0.008609894663095474, + 0.006145702674984932, + 0.008543089963495731, + 0.0069187236949801445, + 0.005488754250109196, + 0.0063208118081092834, + 0.01336139626801014, + 0.027317224070429802, + 0.014231757260859013, + 0.015597007237374783, + 0.018979867920279503, + 0.01287227962166071, + 0.01629372499883175, + 0.009298667311668396, + 0.03856423869729042, + 0.06772931665182114, + 0.053360626101493835, + 0.04929274693131447, + 0.05087251588702202, + 0.04075399413704872, + 0.05095343291759491, + 0.04469829797744751, + 0.005563064478337765, + 0.013957066461443901, + 0.005018041469156742, + 0.006924312561750412, + 0.009877199307084084, + 0.006257789209485054, + 0.010335608385503292, + 0.005173012148588896, + 0.0029181838035583496, + 0.0035889216233044863, + 0.003505836008116603, + 0.003217380726709962, + 0.0033098256681114435, + 0.0031451056711375713, + 0.0038153885398060083, + 0.0032826659735292196, + 0.006490564439445734, + 0.006541100330650806, + 0.003957599401473999, + 0.005259322002530098, + 0.00778123177587986, + 0.00432949885725975, + 0.004258717410266399, + 0.0036872162017971277, + 0.00944875180721283, + 0.014345798641443253, + 0.01024867594242096, + 0.016687538474798203, + 0.01464633084833622, + 0.010540438815951347, + 0.018395090475678444, + 0.009460723027586937, + 0.006475221831351519, + 0.009283105842769146, + 0.0053139436058700085, + 0.00476743932813406, + 0.004862945526838303, + 0.004978510085493326, + 0.004591582342982292, + 0.004911537282168865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.03811439871788025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347240433096886 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347240433096886, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347240433096886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13096311688423157, + 0.04815148189663887, + 0.06445609033107758, + 0.038576189428567886, + 0.25284263491630554, + 0.06114823371171951, + 0.08602912724018097, + 0.23650433123111725, + 0.03167112544178963, + 0.04965770244598389 + ], + "credit_map": [ + 0.019061073660850525, + 0.023123018443584442, + 0.019651692360639572, + 0.01093976479023695, + 0.019748248159885406, + 0.016359947621822357, + 0.006905324757099152, + 0.015174029394984245, + 0.008200837299227715, + 0.008879506029188633, + 0.006196631584316492, + 0.0038402106147259474, + 0.00570504879578948, + 0.006139273289591074, + 0.002685509854927659, + 0.006504458840936422, + 0.009950832463800907, + 0.007208438124507666, + 0.013780246488749981, + 0.0048557547852396965, + 0.005996942054480314, + 0.007472328841686249, + 0.00493681849911809, + 0.010254728607833385, + 0.004609480034559965, + 0.0043991026468575, + 0.006400160025805235, + 0.0033155782148241997, + 0.0062240008264780045, + 0.004172057379037142, + 0.004563338588923216, + 0.004892469383776188, + 0.0231173112988472, + 0.03743923082947731, + 0.0404210239648819, + 0.028987007215619087, + 0.030136924237012863, + 0.028529701754450798, + 0.0266125351190567, + 0.0375988744199276, + 0.00833342969417572, + 0.009357881732285023, + 0.008881522342562675, + 0.0054525709711015224, + 0.0074117123149335384, + 0.006927104201167822, + 0.006673748139292002, + 0.008110257796943188, + 0.011340870521962643, + 0.01415416318923235, + 0.011437063105404377, + 0.007611376233398914, + 0.008657023310661316, + 0.009749866090714931, + 0.006495635490864515, + 0.016583118587732315, + 0.05226144939661026, + 0.02818186953663826, + 0.03275996074080467, + 0.022472083568572998, + 0.020396869629621506, + 0.02926289476454258, + 0.018342340365052223, + 0.03282685950398445, + 0.003992303274571896, + 0.003585302270948887, + 0.004277007654309273, + 0.004023456014692783, + 0.004377379082143307, + 0.003632801119238138, + 0.004525127820670605, + 0.0032577451784163713, + 0.006839352194219828, + 0.004984201863408089, + 0.012815376743674278, + 0.004106991458684206, + 0.0036951389629393816, + 0.0064636957831680775, + 0.0025624551344662905, + 0.00819048285484314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.026347240433096886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, static shot captured from a high angle captures a lone figure, seemingly dwarfed against a vast, sparsely planted, gray, rocky wasteland. The golden afterglow of the setting sun sweeps across the landscape from a very low angle, elongating the figure's silhouette into a long, solitary shadow. The figure wears a dark brown, weathered hooded windbreaker, and the weight of a hiking backpack weighs heavily on his shoulders. He pauses, as if gathering his strength, then slowly bends his knees, lowering his center of gravity. His movement is steady and weighted, ultimately resting in a half-crouch on the cracked earth. One dark-gloved hand reaches down, seemingly examining something on the ground. The dominant tones of the image are desolate gray and ochre, tinged with a warm, melancholic golden hue by the setting sun, creating an epic atmosphere of desolation and tranquility, profoundly conveying the insignificance and resilience of individuals in the face of the harshness of nature.", + "seed": 5715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_3/index_8693_seed_5715_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": -0.11891976995393634, + "vlm_logprob": -0.11891976995393634, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.11891976995393634, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.527961015701294 + ], + "opposite_logprob_tokens": [ + -0.902961015701294 + ], + "logprob_reward": -0.527961015701294, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.527961015701294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2737468183040619, + 0.04171968251466751, + 0.016945628449320793, + 0.026954088360071182, + 0.20920589566230774, + 0.16959883272647858, + 0.11091910302639008, + 0.04539085552096367, + 0.05228658393025398, + 0.05323246493935585 + ], + "credit_map": [ + 0.037052787840366364, + 0.0181774590164423, + 0.09487873315811157, + 0.02329973876476288, + 0.017079535871744156, + 0.031264349818229675, + 0.02536918967962265, + 0.026625044643878937, + 0.005604987498372793, + 0.002483865013346076, + 0.014998726546764374, + 0.004292335826903582, + 0.002143171848729253, + 0.004971255082637072, + 0.003176437458023429, + 0.0040489062666893005, + 0.001231907051987946, + 0.0017159259878098965, + 0.00365841225720942, + 0.001429298659786582, + 0.0014964326983317733, + 0.0034895280841737986, + 0.0019215308129787445, + 0.0020025961566716433, + 0.002487655496224761, + 0.002060405444353819, + 0.00424334779381752, + 0.0031394765246659517, + 0.0028085336089134216, + 0.0034635672345757484, + 0.004401894751936197, + 0.004349212162196636, + 0.013210763223469257, + 0.010569754056632519, + 0.07868849486112595, + 0.01516387052834034, + 0.012536670081317425, + 0.021906014531850815, + 0.019808366894721985, + 0.037321969866752625, + 0.0104865413159132, + 0.011116101406514645, + 0.035563819110393524, + 0.027804818004369736, + 0.01867026276886463, + 0.01571359485387802, + 0.027611184865236282, + 0.022632522508502007, + 0.012496961280703545, + 0.010507564060389996, + 0.022185785695910454, + 0.014177422970533371, + 0.013593760319054127, + 0.007678284775465727, + 0.015175729058682919, + 0.015103600919246674, + 0.0036719380877912045, + 0.007713481318205595, + 0.00842090230435133, + 0.005013476125895977, + 0.00523802638053894, + 0.00427446560934186, + 0.006175736431032419, + 0.0048828329890966415, + 0.006479084026068449, + 0.008709393441677094, + 0.007412515114992857, + 0.006528185680508614, + 0.005706704221665859, + 0.00875072181224823, + 0.004094760399311781, + 0.0046052271500229836, + 0.004275370854884386, + 0.005898898001760244, + 0.007299286779016256, + 0.010186241939663887, + 0.0061645242385566235, + 0.008774175308644772, + 0.005038441624492407, + 0.005595532711595297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.527961015701294 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011097848415374756 + ], + "opposite_logprob_tokens": [ + -4.5110979080200195 + ], + "logprob_reward": -0.011097848415374756, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097848415374756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18604019284248352, + 0.04466509446501732, + 0.12202012538909912, + 0.044255152344703674, + 0.17209504544734955, + 0.06967764347791672, + 0.12624721229076385, + 0.053964510560035706, + 0.03396117314696312, + 0.14707376062870026 + ], + "credit_map": [ + 0.015692977234721184, + 0.018230920657515526, + 0.03229576349258423, + 0.02227587066590786, + 0.023465776816010475, + 0.02452382817864418, + 0.02898198552429676, + 0.02057308703660965, + 0.0027527841739356518, + 0.002154238987714052, + 0.011664994060993195, + 0.0066681369207799435, + 0.005567799787968397, + 0.006012753117829561, + 0.006011521443724632, + 0.0038328729569911957, + 0.019835546612739563, + 0.010101351886987686, + 0.03707163780927658, + 0.01160145178437233, + 0.008068865165114403, + 0.01204164419323206, + 0.011472691781818867, + 0.011826936155557632, + 0.00368282338604331, + 0.003872031345963478, + 0.00973509717732668, + 0.005133886355906725, + 0.005851315334439278, + 0.004027560818940401, + 0.006551352795213461, + 0.005401087924838066, + 0.010827221907675266, + 0.00863535888493061, + 0.06776585429906845, + 0.011960195377469063, + 0.007884438149631023, + 0.016305241733789444, + 0.015525506809353828, + 0.0331912599503994, + 0.004106359556317329, + 0.004549380857497454, + 0.01670796424150467, + 0.010643820278346539, + 0.007773080840706825, + 0.006199919618666172, + 0.010158304125070572, + 0.009538816288113594, + 0.013209470547735691, + 0.010512850247323513, + 0.022358758375048637, + 0.016632309183478355, + 0.016168367117643356, + 0.009668692015111446, + 0.020681416615843773, + 0.0170153696089983, + 0.003916285932064056, + 0.004436034243553877, + 0.009372306056320667, + 0.00545860081911087, + 0.009877486154437065, + 0.00565511267632246, + 0.009756980463862419, + 0.005491707939654589, + 0.004018859006464481, + 0.004964517895132303, + 0.005065930541604757, + 0.004711462184786797, + 0.004331373609602451, + 0.004518602974712849, + 0.0030313676688820124, + 0.0033190608955919743, + 0.010623851791024208, + 0.014002741314470768, + 0.01943392865359783, + 0.027613123878836632, + 0.020859181880950928, + 0.02067074365913868, + 0.015558660961687565, + 0.01831154152750969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.011097848415374756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371020406484604 + ], + "opposite_logprob_tokens": [ + -3.408710241317749 + ], + "logprob_reward": -0.03371020406484604, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371020406484604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1269884556531906, + 0.05285089462995529, + 0.036225516349077225, + 0.046745430678129196, + 0.0960937887430191, + 0.12474807351827621, + 0.09213068336248398, + 0.14626666903495789, + 0.0560024231672287, + 0.22194796800613403 + ], + "credit_map": [ + 0.022335171699523926, + 0.01418670266866684, + 0.03723050653934479, + 0.009237165562808514, + 0.00827811285853386, + 0.012372935190796852, + 0.008544168435037136, + 0.014803702011704445, + 0.0048392280004918575, + 0.0030770401936024427, + 0.01892571710050106, + 0.0058974321000278, + 0.003594429697841406, + 0.006780984811484814, + 0.004431119654327631, + 0.005304950289428234, + 0.0038731477688997984, + 0.0026493393816053867, + 0.010876546613872051, + 0.0038364226929843426, + 0.002361216116696596, + 0.006356022320687771, + 0.0033265736419707537, + 0.0029462508391588926, + 0.0034237646032124758, + 0.0032101324759423733, + 0.008474433794617653, + 0.006738177966326475, + 0.006624160800129175, + 0.005135832820087671, + 0.007062617223709822, + 0.006076316349208355, + 0.005738206673413515, + 0.0040658013895154, + 0.03400030359625816, + 0.00829845666885376, + 0.006518105044960976, + 0.01230583991855383, + 0.009301183745265007, + 0.015865905210375786, + 0.007952435873448849, + 0.00881615374237299, + 0.021191785112023354, + 0.020057374611496925, + 0.015826784074306488, + 0.012543940916657448, + 0.02125726081430912, + 0.017102353274822235, + 0.006310744676738977, + 0.007471317891031504, + 0.011018849909305573, + 0.014079468324780464, + 0.012951378710567951, + 0.009584709070622921, + 0.019646549597382545, + 0.011067679151892662, + 0.01048949547111988, + 0.018995152786374092, + 0.024645939469337463, + 0.013737361878156662, + 0.02338750846683979, + 0.016478415578603745, + 0.023906316608190536, + 0.014626492746174335, + 0.005726504139602184, + 0.008306395262479782, + 0.008907987736165524, + 0.007267820183187723, + 0.007065487094223499, + 0.0061686537228524685, + 0.005354869179427624, + 0.007204712834209204, + 0.013260029256343842, + 0.014421562664210796, + 0.02943367511034012, + 0.021936235949397087, + 0.030472531914711, + 0.02488696575164795, + 0.03476593643426895, + 0.05277106538414955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.03371020406484604 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419964898377657 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.01419964898377657, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419964898377657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0914405956864357, + 0.024883095175027847, + 0.036059945821762085, + 0.026239024475216866, + 0.0992351695895195, + 0.031434401869773865, + 0.06603579968214035, + 0.04425188899040222, + 0.019668037071824074, + 0.5607520341873169 + ], + "credit_map": [ + 0.008412583731114864, + 0.007884609512984753, + 0.02424265630543232, + 0.009063320234417915, + 0.008603048510849476, + 0.01170180644840002, + 0.01113047543913126, + 0.010402088984847069, + 0.001721115317195654, + 0.0012023255694657564, + 0.007432170212268829, + 0.0031958548352122307, + 0.0023520670365542173, + 0.0036120417062193155, + 0.003224362153559923, + 0.002143159508705139, + 0.0062482464127242565, + 0.003133525839075446, + 0.012244710698723793, + 0.0030137996654957533, + 0.00139233423396945, + 0.004124247469007969, + 0.003011722583323717, + 0.0028913600835949183, + 0.0029995625372976065, + 0.0032233973033726215, + 0.0040108924731612206, + 0.0026530481409281492, + 0.002067727502435446, + 0.0032709180377423763, + 0.003468656912446022, + 0.004544822499155998, + 0.00630528898909688, + 0.0048421588726341724, + 0.03795890882611275, + 0.007245301268994808, + 0.005194793920964003, + 0.010294863022863865, + 0.00906816404312849, + 0.018325693905353546, + 0.0025401536840945482, + 0.002822678769007325, + 0.006197687238454819, + 0.0046659498475492, + 0.0030058640986680984, + 0.00394806545227766, + 0.005001566372811794, + 0.003252436174079776, + 0.006332184188067913, + 0.005241746082901955, + 0.009820089675486088, + 0.010211201384663582, + 0.009883860126137733, + 0.0049174209125339985, + 0.011199387721717358, + 0.008429909124970436, + 0.0032764351926743984, + 0.003880178788676858, + 0.007100650575011969, + 0.005362383555620909, + 0.007098776288330555, + 0.004894869867712259, + 0.00798691064119339, + 0.004651683382689953, + 0.002623284701257944, + 0.00321603543125093, + 0.0026206783950328827, + 0.0027875539381057024, + 0.0025560848880559206, + 0.002362256171181798, + 0.0016110592987388372, + 0.0018910838989540935, + 0.0369102917611599, + 0.04196105897426605, + 0.07465573400259018, + 0.1007690280675888, + 0.0662861317396164, + 0.06963413953781128, + 0.07300728559494019, + 0.09752834588289261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.01419964898377657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high angle showing a lone figure in a vast, gray, rocky wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark brown, weathered hooded windbreaker and carrying a hiking backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure perform the action of bending knees to lower their center of gravity and rest in a half-crouch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dark-gloved hand reach down to examine something on the cracked earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a low-angle setting sun, creating long shadows and a desolate atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630132604390383 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630132604390383, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630132604390383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2648102343082428, + 0.02627963200211525, + 0.032339729368686676, + 0.038613300770521164, + 0.24139344692230225, + 0.08466020971536636, + 0.14837515354156494, + 0.033320434391498566, + 0.027282794937491417, + 0.10292509198188782 + ], + "credit_map": [ + 0.03732539713382721, + 0.015466541983187199, + 0.09176391363143921, + 0.02116801030933857, + 0.015195892192423344, + 0.026423456147313118, + 0.025038346648216248, + 0.03242865204811096, + 0.00267160520888865, + 0.0026352806016802788, + 0.0042316545732319355, + 0.0032015310134738684, + 0.003489862196147442, + 0.0026063367258757353, + 0.003464057110249996, + 0.003979304805397987, + 0.0036823758855462074, + 0.004437878727912903, + 0.00619836151599884, + 0.0033183638006448746, + 0.0028364164754748344, + 0.005489503499120474, + 0.0030022517312318087, + 0.0033745774999260902, + 0.003772979136556387, + 0.003267995547503233, + 0.006455531343817711, + 0.00457494892179966, + 0.0037711551412940025, + 0.004441010765731335, + 0.0064834244549274445, + 0.005846251733601093, + 0.015985969454050064, + 0.012987890280783176, + 0.08980825543403625, + 0.017650635913014412, + 0.013552430085837841, + 0.025463536381721497, + 0.023368626832962036, + 0.042576078325510025, + 0.005789618473500013, + 0.004939212463796139, + 0.024147123098373413, + 0.012063424102962017, + 0.007659922819584608, + 0.007170791272073984, + 0.010875780135393143, + 0.012014325708150864, + 0.0164638664573431, + 0.013635578565299511, + 0.027083219960331917, + 0.019275322556495667, + 0.019730309024453163, + 0.010223865509033203, + 0.02224329300224781, + 0.019719690084457397, + 0.002979450160637498, + 0.004071693867444992, + 0.003999365493655205, + 0.004578464664518833, + 0.005098342429846525, + 0.003380548907443881, + 0.005384932272136211, + 0.003827634733170271, + 0.003394140861928463, + 0.0033778164070099592, + 0.0042160567827522755, + 0.0033158590085804462, + 0.004176783841103315, + 0.0032064032275229692, + 0.0030990783125162125, + 0.0024966546334326267, + 0.0080701420083642, + 0.008395631797611713, + 0.01264842227101326, + 0.016787689179182053, + 0.01623038575053215, + 0.011099719442427158, + 0.013853047043085098, + 0.015840040519833565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.007630132604390383 + } + ], + "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/0000025/group_4/index_tensor([6191, 6191])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_4/index_tensor([6191, 6191])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f3d6964b406c41d5563b5f753da7adb321b97b3a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_4/index_tensor([6191, 6191])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, close-up tracking shot, panning steadily to the right, closely follows a mottled mural painted on a damp, water-stained, rough concrete wall deep within an abandoned subway tunnel. The mural features a simplified human figure, outlined in faded ochre and charcoal black. Its form is distorted and elongated, its surface riddled with fine cracks and patches of flaking. A harsh, cold, white searchlight, emanating from outside the frame, strikes the wall at a sharp angle, piercing through the thick, white simulated smoke that slowly swirls in the air. The light not only highlights the wall's every uneven texture, but also accentuates the mural's texture. Under the gaze of the camera, this human-shaped pattern, which should have been still, is strangely moving autonomously in an incoherent, twitching manner similar to stop-motion animation. It slides horizontally along the wall, and its paint edges seem to dissolve and re-solidify in front, driving tiny concrete fragments and dried paint powder to fall. The whole picture, with its cold tones, creates a strong sense of supernatural suspense and an unsettling claustrophobic atmosphere.", + "seed": 10834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_4/index_6191_seed_10834_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": -0.023115601483732463, + "vlm_logprob": -0.023115601483732463, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.023115601483732463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038374047726392746 + ], + "opposite_logprob_tokens": [ + -3.5383739471435547 + ], + "logprob_reward": -0.038374047726392746, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.038374047726392746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2280668020248413, + 0.053202640265226364, + 0.16955743730068207, + 0.18554991483688354, + 0.03241793066263199, + 0.032190460711717606, + 0.03690499812364578, + 0.15756234526634216, + 0.034033242613077164, + 0.07051421701908112 + ], + "credit_map": [ + 0.013146085664629936, + 0.025837404653429985, + 0.02315376326441765, + 0.021939747035503387, + 0.02559526450932026, + 0.05162752792239189, + 0.01819177344441414, + 0.048575252294540405, + 0.003489293158054352, + 0.009380380623042583, + 0.00535599747672677, + 0.006918181665241718, + 0.006784556433558464, + 0.008270295336842537, + 0.006894922815263271, + 0.006109009496867657, + 0.01355438120663166, + 0.02109295316040516, + 0.019183071330189705, + 0.025984052568674088, + 0.025618990883231163, + 0.025085635483264923, + 0.016706407070159912, + 0.02233194001019001, + 0.011393643915653229, + 0.020852016285061836, + 0.016883207485079765, + 0.020692095160484314, + 0.026706503704190254, + 0.04128812253475189, + 0.022422173991799355, + 0.025312162935733795, + 0.0025966602843254805, + 0.004548937547951937, + 0.004696846008300781, + 0.004229661077260971, + 0.003535797353833914, + 0.004120638128370047, + 0.003874729387462139, + 0.004814662039279938, + 0.0030228402465581894, + 0.004078065976500511, + 0.0031503094360232353, + 0.0044759404845535755, + 0.005008370615541935, + 0.0047796498984098434, + 0.004162365570664406, + 0.003512915223836899, + 0.004075430799275637, + 0.004895057994872332, + 0.005035145673900843, + 0.004445867147296667, + 0.004820633213967085, + 0.006185318809002638, + 0.0031114162411540747, + 0.004336124286055565, + 0.00914105586707592, + 0.02732263132929802, + 0.01766023226082325, + 0.023313703015446663, + 0.009929723106324673, + 0.011642924509942532, + 0.031377390027046204, + 0.0271746926009655, + 0.0023872738238424063, + 0.006632887292653322, + 0.0036123408935964108, + 0.00498975208029151, + 0.004189120605587959, + 0.004104964435100555, + 0.005436429753899574, + 0.0026804713997989893, + 0.006425668019801378, + 0.0061041200533509254, + 0.01238611713051796, + 0.013745173811912537, + 0.010206460021436214, + 0.006063341163098812, + 0.008723906241357327, + 0.0068594287149608135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.038374047726392746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00283946399576962 + ], + "opposite_logprob_tokens": [ + -5.8778395652771 + ], + "logprob_reward": -0.00283946399576962, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00283946399576962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2288394272327423, + 0.033511582762002945, + 0.0654653012752533, + 0.11499302089214325, + 0.0381358303129673, + 0.038867246359586716, + 0.03288143873214722, + 0.35573115944862366, + 0.0517876036465168, + 0.039787326008081436 + ], + "credit_map": [ + 0.01362857036292553, + 0.03095509670674801, + 0.022551734000444412, + 0.02889171428978443, + 0.034424763172864914, + 0.041549891233444214, + 0.02062896080315113, + 0.036208730190992355, + 0.002764908131211996, + 0.0060630119405686855, + 0.0036489125341176987, + 0.004927535075694323, + 0.004025258123874664, + 0.003673763247206807, + 0.004625319037586451, + 0.0037828809581696987, + 0.004269195720553398, + 0.010856599546968937, + 0.008064121939241886, + 0.009520952589809895, + 0.008775941096246243, + 0.008381613530218601, + 0.007764890324324369, + 0.007831992581486702, + 0.008823457174003124, + 0.011326397769153118, + 0.014368153177201748, + 0.012434185482561588, + 0.02487899921834469, + 0.016647877171635628, + 0.014599226415157318, + 0.011914748698472977, + 0.0030754816252738237, + 0.004920390900224447, + 0.005444741807878017, + 0.004948490299284458, + 0.004163981880992651, + 0.005837243981659412, + 0.004604808054864407, + 0.005140696186572313, + 0.0027336280327290297, + 0.0038348452653735876, + 0.003463219152763486, + 0.008819052018225193, + 0.004527179524302483, + 0.004843648057430983, + 0.003925258293747902, + 0.006720423698425293, + 0.004229831043630838, + 0.004174819216132164, + 0.004726584535092115, + 0.005207793787121773, + 0.003348889062181115, + 0.004834753926843405, + 0.0030214288271963596, + 0.0033373436890542507, + 0.016508178785443306, + 0.025111867114901543, + 0.014119461178779602, + 0.03273697569966316, + 0.015829063951969147, + 0.01459635142236948, + 0.12330876290798187, + 0.1135205402970314, + 0.004034959711134434, + 0.010516039095818996, + 0.00799095630645752, + 0.00817212462425232, + 0.006528220605105162, + 0.004750112537294626, + 0.005546357482671738, + 0.0042488412000238895, + 0.003525966312736273, + 0.003443819470703602, + 0.0068633053451776505, + 0.008010421879589558, + 0.005061695352196693, + 0.003511573886498809, + 0.005111767444759607, + 0.004258783534169197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.00283946399576962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015210260171443224 + ], + "opposite_logprob_tokens": [ + -6.501521110534668 + ], + "logprob_reward": -0.0015210260171443224, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015210260171443224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12655697762966156, + 0.034521300345659256, + 0.1145615503191948, + 0.09774893522262573, + 0.041450437158346176, + 0.044867515563964844, + 0.04646918550133705, + 0.346539169549942, + 0.10228672623634338, + 0.04499824717640877 + ], + "credit_map": [ + 0.00847394485026598, + 0.01696263812482357, + 0.01578516513109207, + 0.01613163948059082, + 0.022483566775918007, + 0.018510716035962105, + 0.013383422046899796, + 0.014825872145593166, + 0.003692034399136901, + 0.003651689738035202, + 0.004907938651740551, + 0.004345830529928207, + 0.0027887122705578804, + 0.006517614703625441, + 0.004346336703747511, + 0.0042711421847343445, + 0.009634708985686302, + 0.019956879317760468, + 0.01282982062548399, + 0.021885426715016365, + 0.011147662065923214, + 0.012621382251381874, + 0.012011837214231491, + 0.01447383128106594, + 0.006445754785090685, + 0.011664032004773617, + 0.012389834970235825, + 0.011165243573486805, + 0.023127490654587746, + 0.01301512960344553, + 0.011496491730213165, + 0.008444949984550476, + 0.0037802683655172586, + 0.005552899558097124, + 0.006971490569412708, + 0.0058847032487392426, + 0.004291631747037172, + 0.005661085247993469, + 0.00473411288112402, + 0.0045742434449493885, + 0.003433583304286003, + 0.0065055126324296, + 0.007552879396826029, + 0.008470972068607807, + 0.005382557399570942, + 0.004490520339459181, + 0.0050057596527040005, + 0.00402572937309742, + 0.007152268663048744, + 0.004730600863695145, + 0.0051676686853170395, + 0.005233987234532833, + 0.0058135986328125, + 0.007963922806084156, + 0.00507863936945796, + 0.005328493192791939, + 0.01462232880294323, + 0.018124325200915337, + 0.010407223366200924, + 0.028190432116389275, + 0.01569504663348198, + 0.013826475478708744, + 0.12793084979057312, + 0.11774245649576187, + 0.008631136268377304, + 0.019640130922198296, + 0.01725158654153347, + 0.017120348289608955, + 0.01193989347666502, + 0.009603203274309635, + 0.011069269850850105, + 0.007031145505607128, + 0.004186986479908228, + 0.004216931294649839, + 0.008768211118876934, + 0.007775002159178257, + 0.004728586878627539, + 0.003929030615836382, + 0.007393013220280409, + 0.004000483080744743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0015210260171443224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790915548801422 + ], + "opposite_logprob_tokens": [ + -3.5297908782958984 + ], + "logprob_reward": -0.029790915548801422, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790915548801422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18312692642211914, + 0.04359341412782669, + 0.11031138151884079, + 0.10412765294313431, + 0.06231074407696724, + 0.05198023095726967, + 0.0343317836523056, + 0.23340915143489838, + 0.09640824794769287, + 0.08040047436952591 + ], + "credit_map": [ + 0.011357644572854042, + 0.018781667575240135, + 0.020077893510460854, + 0.020980030298233032, + 0.02628203295171261, + 0.03565198555588722, + 0.018088318407535553, + 0.0319073460996151, + 0.0036710265558212996, + 0.0076756831258535385, + 0.00395595096051693, + 0.007210121490061283, + 0.005698484368622303, + 0.004441406577825546, + 0.006603720597922802, + 0.004337023012340069, + 0.006574454717338085, + 0.01700829342007637, + 0.013749952428042889, + 0.02174721471965313, + 0.008847949095070362, + 0.012292648665606976, + 0.016038063913583755, + 0.014052802696824074, + 0.0076450165361166, + 0.010377074591815472, + 0.011704490520060062, + 0.014450526796281338, + 0.013895843178033829, + 0.015109135769307613, + 0.015360441990196705, + 0.015585113316774368, + 0.005909076891839504, + 0.008596261031925678, + 0.008734602481126785, + 0.011561742052435875, + 0.006679215934127569, + 0.007996883243322372, + 0.006413877476006746, + 0.006419082172214985, + 0.00455486262217164, + 0.007967262528836727, + 0.008558697998523712, + 0.00725241145119071, + 0.006985363084822893, + 0.00480384286493063, + 0.006845423486083746, + 0.005012367852032185, + 0.005389966536313295, + 0.0045494199730455875, + 0.004260299727320671, + 0.004618831444531679, + 0.0037565664388239384, + 0.005230400711297989, + 0.0034223420079797506, + 0.003103958209976554, + 0.01550403144210577, + 0.042121678590774536, + 0.021875465288758278, + 0.03614582121372223, + 0.013951271772384644, + 0.014825746417045593, + 0.04632452130317688, + 0.04266060143709183, + 0.007314963731914759, + 0.02071245387196541, + 0.012836398556828499, + 0.01579410769045353, + 0.011698187328875065, + 0.008302867412567139, + 0.01092445943504572, + 0.00882481038570404, + 0.010019037872552872, + 0.008520565927028656, + 0.01470110658556223, + 0.010797432623803616, + 0.013154597021639347, + 0.005858385935425758, + 0.010465262457728386, + 0.006884084548801184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.029790915548801422 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430525541305542 + ], + "opposite_logprob_tokens": [ + -3.1680526733398438 + ], + "logprob_reward": -0.0430525541305542, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430525541305542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06955280154943466, + 0.07793617993593216, + 0.06310560554265976, + 0.023783521726727486, + 0.02944193407893181, + 0.02505270019173622, + 0.028599485754966736, + 0.6057335734367371, + 0.034352414309978485, + 0.04244180768728256 + ], + "credit_map": [ + 0.00400348287075758, + 0.009678334929049015, + 0.004673385992646217, + 0.00896866712719202, + 0.013612410984933376, + 0.011441596783697605, + 0.007334066554903984, + 0.009840843267738819, + 0.005327722989022732, + 0.016511334106326103, + 0.007558893878012896, + 0.012016917578876019, + 0.011067203246057034, + 0.00833870004862547, + 0.008774468675255775, + 0.008340932428836823, + 0.005163925234228373, + 0.013132570311427116, + 0.005434377118945122, + 0.00797642394900322, + 0.009636255912482738, + 0.00709500489756465, + 0.008162450045347214, + 0.006504592951387167, + 0.001971073215827346, + 0.0038209836930036545, + 0.0025870446115732193, + 0.003203378291800618, + 0.0033435262739658356, + 0.0030741484370082617, + 0.0031544924713671207, + 0.002628873335197568, + 0.0026447968557476997, + 0.004137604031711817, + 0.0032992998603731394, + 0.005565481726080179, + 0.003112702863290906, + 0.003450471442192793, + 0.003693472594022751, + 0.003538100514560938, + 0.0028325021266937256, + 0.0038710038643330336, + 0.003883764613419771, + 0.003054805798456073, + 0.00339764473028481, + 0.0020023430697619915, + 0.0028061154298484325, + 0.0032045175321400166, + 0.0030194763094186783, + 0.002784040290862322, + 0.003328489838168025, + 0.00305859069339931, + 0.0037712580524384975, + 0.005633159540593624, + 0.003331756452098489, + 0.0036727110855281353, + 0.024746015667915344, + 0.018740113824605942, + 0.009436523541808128, + 0.042303021997213364, + 0.02738822251558304, + 0.030683355405926704, + 0.23368245363235474, + 0.21875379979610443, + 0.0023528013844043016, + 0.004860570188611746, + 0.0031742823775857687, + 0.005523208994418383, + 0.004479025490581989, + 0.005241305101662874, + 0.005287913139909506, + 0.0034333027433604, + 0.0027043481823056936, + 0.0032650744542479515, + 0.006678200792521238, + 0.008484371937811375, + 0.005811044946312904, + 0.0034143810626119375, + 0.007201319094747305, + 0.004883060697466135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0430525541305542 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up tracking shot, panning steadily to the right, closely follows a mottled mural painted on a damp, water-stained, rough concrete wall deep within an abandoned subway tunnel. The mural features a simplified human figure, outlined in faded ochre and charcoal black. Its form is distorted and elongated, its surface riddled with fine cracks and patches of flaking. A harsh, cold, white searchlight, emanating from outside the frame, strikes the wall at a sharp angle, piercing through the thick, white simulated smoke that slowly swirls in the air. The light not only highlights the wall's every uneven texture, but also accentuates the mural's texture. Under the gaze of the camera, this human-shaped pattern, which should have been still, is strangely moving autonomously in an incoherent, twitching manner similar to stop-motion animation. It slides horizontally along the wall, and its paint edges seem to dissolve and re-solidify in front, driving tiny concrete fragments and dried paint powder to fall. The whole picture, with its cold tones, creates a strong sense of supernatural suspense and an unsettling claustrophobic atmosphere.", + "seed": 10835, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_4/index_6191_seed_10835_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": -0.02273067981004715, + "vlm_logprob": -0.02273067981004715, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02273067981004715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02654251642525196 + ], + "opposite_logprob_tokens": [ + -3.9015424251556396 + ], + "logprob_reward": -0.02654251642525196, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02654251642525196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13775870203971863, + 0.04597862437367439, + 0.06241167336702347, + 0.06950259208679199, + 0.10039370507001877, + 0.0586402527987957, + 0.15071529150009155, + 0.0880427211523056, + 0.1718108206987381, + 0.1147456169128418 + ], + "credit_map": [ + 0.015392141416668892, + 0.016833193600177765, + 0.01792122796177864, + 0.018318865448236465, + 0.01674826070666313, + 0.020102644339203835, + 0.018165551126003265, + 0.014276823960244656, + 0.004062442574650049, + 0.009296299889683723, + 0.005282872822135687, + 0.006044529844075441, + 0.0035117724910378456, + 0.005615745671093464, + 0.00582216028124094, + 0.006342803128063679, + 0.0066758617758750916, + 0.00822317972779274, + 0.009596230462193489, + 0.015571244060993195, + 0.00652061915025115, + 0.004537227563560009, + 0.007713864091783762, + 0.0035734393168240786, + 0.008735112845897675, + 0.007460202556103468, + 0.010314812883734703, + 0.007514019962400198, + 0.007871233858168125, + 0.008320375345647335, + 0.010759159922599792, + 0.00852767564356327, + 0.012368157505989075, + 0.006143426056951284, + 0.010618720203638077, + 0.007821973413228989, + 0.008644904009997845, + 0.012403764761984348, + 0.016166696324944496, + 0.026226064190268517, + 0.005573240574449301, + 0.008888796903192997, + 0.00977014284580946, + 0.008427081629633904, + 0.007848459295928478, + 0.006161379627883434, + 0.006196482572704554, + 0.00577466981485486, + 0.009181572124361992, + 0.026956794783473015, + 0.010599198751151562, + 0.02663702517747879, + 0.025181584060192108, + 0.017991187050938606, + 0.020481353625655174, + 0.0136865871027112, + 0.008156266994774342, + 0.017268208786845207, + 0.01013687439262867, + 0.011431668885052204, + 0.00902816466987133, + 0.010671189054846764, + 0.01333798747509718, + 0.008012363687157631, + 0.013643214479088783, + 0.026779737323522568, + 0.011395853012800217, + 0.042123567312955856, + 0.014172276481986046, + 0.017778130248188972, + 0.0165297519415617, + 0.029388301074504852, + 0.007305387407541275, + 0.014589603990316391, + 0.010456441901624203, + 0.01675681583583355, + 0.014119674451649189, + 0.016922978684306145, + 0.019543379545211792, + 0.015051338821649551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.02654251642525196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067700534127652645 + ], + "opposite_logprob_tokens": [ + -5.006770133972168 + ], + "logprob_reward": -0.0067700534127652645, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067700534127652645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08788402378559113, + 0.07876704633235931, + 0.09658274054527283, + 0.06829653680324554, + 0.07733190804719925, + 0.09823337942361832, + 0.30066853761672974, + 0.08492589741945267, + 0.07310616970062256, + 0.034203801304101944 + ], + "credit_map": [ + 0.009555553086102009, + 0.009492621757090092, + 0.009932922199368477, + 0.010702453553676605, + 0.012556170113384724, + 0.012514496222138405, + 0.01311307679861784, + 0.01001673098653555, + 0.008172391913831234, + 0.016616065055131912, + 0.007339305244386196, + 0.005333943758159876, + 0.005476945545524359, + 0.012849626131355762, + 0.006268596276640892, + 0.016710171476006508, + 0.010370750911533833, + 0.009309937246143818, + 0.012182800099253654, + 0.013897453434765339, + 0.012554657645523548, + 0.009880591183900833, + 0.016957968473434448, + 0.01142858900129795, + 0.00667173694819212, + 0.006097913254052401, + 0.008501912467181683, + 0.004896137397736311, + 0.01051850151270628, + 0.009755430743098259, + 0.011325106024742126, + 0.010529800318181515, + 0.008762389421463013, + 0.003922715783119202, + 0.00532435392960906, + 0.005550936795771122, + 0.008479354903101921, + 0.013328109867870808, + 0.011804413981735706, + 0.020159633830189705, + 0.007930897176265717, + 0.013095510192215443, + 0.006246168166399002, + 0.009659767150878906, + 0.013366379775106907, + 0.0179420355707407, + 0.01213705912232399, + 0.017855564132332802, + 0.017348067834973335, + 0.048936907202005386, + 0.02047000452876091, + 0.06473636627197266, + 0.052325084805488586, + 0.027428649365901947, + 0.039465099573135376, + 0.02995838038623333, + 0.007765629794448614, + 0.018743902444839478, + 0.009439792484045029, + 0.010381997562944889, + 0.008455593138933182, + 0.011614515446126461, + 0.010537386871874332, + 0.007987072691321373, + 0.005628284532576799, + 0.012562138959765434, + 0.0045449016615748405, + 0.009323623031377792, + 0.011239491403102875, + 0.00984511710703373, + 0.009837544523179531, + 0.010125065222382545, + 0.002275484148412943, + 0.00452021136879921, + 0.0027456271927803755, + 0.0040458980947732925, + 0.00665066996589303, + 0.003938955720514059, + 0.0041940780356526375, + 0.005832879338413477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.0067700534127652645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004109153989702463 + ], + "opposite_logprob_tokens": [ + -5.5041093826293945 + ], + "logprob_reward": -0.004109153989702463, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004109153989702463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09588176012039185, + 0.0744197815656662, + 0.043077193200588226, + 0.09797818213701248, + 0.06865166127681732, + 0.13617736101150513, + 0.2928667962551117, + 0.07359857857227325, + 0.08172182738780975, + 0.03562677651643753 + ], + "credit_map": [ + 0.009867711924016476, + 0.011246518231928349, + 0.010950327850878239, + 0.011051999405026436, + 0.011322597973048687, + 0.013672399334609509, + 0.011955385096371174, + 0.015814825892448425, + 0.006388846319168806, + 0.01081023644655943, + 0.006032952107489109, + 0.00823173951357603, + 0.006766651291400194, + 0.013082673773169518, + 0.007371070794761181, + 0.01573561690747738, + 0.004163236822932959, + 0.004760426003485918, + 0.005200937855988741, + 0.006980300415307283, + 0.005420522298663855, + 0.005188988987356424, + 0.006555955857038498, + 0.004806831944733858, + 0.00883682630956173, + 0.009637512266635895, + 0.008458035066723824, + 0.011142385192215443, + 0.013267012313008308, + 0.0204613134264946, + 0.014649336226284504, + 0.011525782756507397, + 0.008135603740811348, + 0.003869439009577036, + 0.004037842620164156, + 0.004615437239408493, + 0.007561386562883854, + 0.008779820054769516, + 0.011526537127792835, + 0.020125605165958405, + 0.010810193605720997, + 0.01631391979753971, + 0.008248594589531422, + 0.015505779534578323, + 0.017955070361495018, + 0.02547045424580574, + 0.015802530571818352, + 0.026070833206176758, + 0.016839206218719482, + 0.04129724204540253, + 0.016551505774259567, + 0.061943553388118744, + 0.05362462252378464, + 0.028589801862835884, + 0.04080362990498543, + 0.03321726992726326, + 0.004354133736342192, + 0.0063938298262655735, + 0.0056302896700799465, + 0.006228283513337374, + 0.008298526518046856, + 0.014347538352012634, + 0.011138894595205784, + 0.017207084223628044, + 0.005427923984825611, + 0.012403666041791439, + 0.005480690393596888, + 0.012674114666879177, + 0.01667511835694313, + 0.009372108615934849, + 0.010079952888190746, + 0.00960826687514782, + 0.002125110477209091, + 0.0036712237633764744, + 0.0026044417172670364, + 0.003212135983631015, + 0.007742690853774548, + 0.0047983466647565365, + 0.005489877890795469, + 0.005982955452054739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.004109153989702463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203708052635193 + ], + "opposite_logprob_tokens": [ + -2.8120369911193848 + ], + "logprob_reward": -0.06203708052635193, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203708052635193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12054883688688278, + 0.09377933293581009, + 0.11800562590360641, + 0.08878417313098907, + 0.11806085705757141, + 0.05052625760436058, + 0.15254651010036469, + 0.15743620693683624, + 0.060242291539907455, + 0.04006987437605858 + ], + "credit_map": [ + 0.013951880857348442, + 0.0124469343572855, + 0.010136115364730358, + 0.011957371607422829, + 0.01740809716284275, + 0.018994253128767014, + 0.01602700911462307, + 0.019627181813120842, + 0.007308844476938248, + 0.010850653983652592, + 0.007072990294545889, + 0.010790897533297539, + 0.009560546837747097, + 0.017376098781824112, + 0.011928611434996128, + 0.01889069378376007, + 0.010954633355140686, + 0.010911948047578335, + 0.016527609899640083, + 0.021429862827062607, + 0.011340905912220478, + 0.013031673617661, + 0.016540780663490295, + 0.01726822555065155, + 0.008037615567445755, + 0.006192397326231003, + 0.007450950797647238, + 0.009944011457264423, + 0.009930240921676159, + 0.01984991319477558, + 0.014303648844361305, + 0.013075404800474644, + 0.01032215915620327, + 0.006640459410846233, + 0.006470729596912861, + 0.009330586530268192, + 0.013943720608949661, + 0.022595973685383797, + 0.019174551591277122, + 0.029582686722278595, + 0.004003811627626419, + 0.005451403092592955, + 0.00863643828779459, + 0.005508412607014179, + 0.005707035306841135, + 0.008076196536421776, + 0.00623222254216671, + 0.006910736672580242, + 0.00859876163303852, + 0.014335472136735916, + 0.008279756642878056, + 0.031888220459222794, + 0.019535701721906662, + 0.02791564166545868, + 0.018223192542791367, + 0.02376978099346161, + 0.00823039747774601, + 0.01258962880820036, + 0.006858561187982559, + 0.009227401576936245, + 0.016845455393195152, + 0.03710470721125603, + 0.020007962360978127, + 0.04657210409641266, + 0.0036655559670180082, + 0.008414506912231445, + 0.004018341191112995, + 0.01494415383785963, + 0.008295844309031963, + 0.006306807976216078, + 0.00656878249719739, + 0.008028303273022175, + 0.0035714947152882814, + 0.004976092837750912, + 0.003772422671318054, + 0.004910503048449755, + 0.006466421764343977, + 0.004948822315782309, + 0.005685288459062576, + 0.0057388306595385075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.06203708052635193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.16114264726638794, + 0.04598645120859146, + 0.04966191202402115, + 0.12210919708013535, + 0.061649732291698456, + 0.07237095385789871, + 0.13103839755058289, + 0.16864511370658875, + 0.1310628056526184, + 0.0563327893614769 + ], + "credit_map": [ + 0.019440224394202232, + 0.01868634670972824, + 0.015974555164575577, + 0.020564023405313492, + 0.023288600146770477, + 0.018872501328587532, + 0.023594101890921593, + 0.020722297951579094, + 0.003408316522836685, + 0.007410009857267141, + 0.0038878154009580612, + 0.005538084078580141, + 0.004107389133423567, + 0.0071233403868973255, + 0.005860490258783102, + 0.00865100510418415, + 0.005480675492435694, + 0.005439312197268009, + 0.0055442629382014275, + 0.010541476309299469, + 0.006756564136594534, + 0.005674176383763552, + 0.0057404073886573315, + 0.004485037177801132, + 0.018469391390681267, + 0.011823241598904133, + 0.024369265884160995, + 0.013571729883551598, + 0.012754841707646847, + 0.016548387706279755, + 0.012566533870995045, + 0.012005808763206005, + 0.00796587485820055, + 0.0046054283156991005, + 0.006621725391596556, + 0.005567322950810194, + 0.005261973012238741, + 0.007936004549264908, + 0.00687143811956048, + 0.016819966956973076, + 0.0063050128519535065, + 0.010660913772881031, + 0.009673289954662323, + 0.008037171326577663, + 0.012848983518779278, + 0.009237782098352909, + 0.007151463069021702, + 0.008456345647573471, + 0.00811161007732153, + 0.017392845824360847, + 0.009216421283781528, + 0.02129816636443138, + 0.0279312152415514, + 0.0131105687469244, + 0.020030010491609573, + 0.013947549276053905, + 0.01324504055082798, + 0.027573226019740105, + 0.01426696963608265, + 0.015120698139071465, + 0.019544649869203568, + 0.03154825419187546, + 0.022137584164738655, + 0.025208676233887672, + 0.006181051954627037, + 0.01693674363195896, + 0.008203974924981594, + 0.03369594365358353, + 0.024052392691373825, + 0.013514524325728416, + 0.014294201508164406, + 0.01418397482484579, + 0.004016546998172998, + 0.009684506803750992, + 0.004893957171589136, + 0.00871188659220934, + 0.006939079612493515, + 0.00809541717171669, + 0.006691467948257923, + 0.0072999256663024426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.014194594696164131 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up tracking shot, panning steadily to the right, closely follows a mottled mural painted on a damp, water-stained, rough concrete wall deep within an abandoned subway tunnel. The mural features a simplified human figure, outlined in faded ochre and charcoal black. Its form is distorted and elongated, its surface riddled with fine cracks and patches of flaking. A harsh, cold, white searchlight, emanating from outside the frame, strikes the wall at a sharp angle, piercing through the thick, white simulated smoke that slowly swirls in the air. The light not only highlights the wall's every uneven texture, but also accentuates the mural's texture. Under the gaze of the camera, this human-shaped pattern, which should have been still, is strangely moving autonomously in an incoherent, twitching manner similar to stop-motion animation. It slides horizontally along the wall, and its paint edges seem to dissolve and re-solidify in front, driving tiny concrete fragments and dried paint powder to fall. The whole picture, with its cold tones, creates a strong sense of supernatural suspense and an unsettling claustrophobic atmosphere.", + "seed": 10836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_4/index_6191_seed_10836_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": -0.01971357485745102, + "vlm_logprob": -0.01971357485745102, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.01971357485745102, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021162305027246475 + ], + "opposite_logprob_tokens": [ + -4.396162509918213 + ], + "logprob_reward": -0.021162305027246475, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.021162305027246475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33534350991249084, + 0.0625414177775383, + 0.03129180148243904, + 0.04788300395011902, + 0.02058514952659607, + 0.26628661155700684, + 0.027510643005371094, + 0.06583830714225769, + 0.05666237697005272, + 0.08605718612670898 + ], + "credit_map": [ + 0.021982040256261826, + 0.030646860599517822, + 0.037701260298490524, + 0.04924783110618591, + 0.041287507861852646, + 0.04757564142346382, + 0.06333736330270767, + 0.04356498271226883, + 0.006400601472705603, + 0.0074230884201824665, + 0.009279437363147736, + 0.008533316664397717, + 0.005994289182126522, + 0.007198375184088945, + 0.008697514422237873, + 0.009014789015054703, + 0.0027408793102949858, + 0.003399504581466317, + 0.0038222952280193567, + 0.005690983962267637, + 0.00425713462755084, + 0.0030489997006952763, + 0.004814207553863525, + 0.00351779256016016, + 0.0038845406379550695, + 0.004175747279077768, + 0.009600802324712276, + 0.005373604129999876, + 0.005971910897642374, + 0.006043027155101299, + 0.007188689894974232, + 0.00564467441290617, + 0.0024474558886140585, + 0.0031051780097186565, + 0.003288192441686988, + 0.002325564157217741, + 0.0018579430179670453, + 0.0028813418466597795, + 0.002597287530079484, + 0.0020821846555918455, + 0.010189367458224297, + 0.00429856451228261, + 0.06456048786640167, + 0.008070133626461029, + 0.008958768099546432, + 0.1072503849864006, + 0.014544588513672352, + 0.04841430485248566, + 0.002479317830875516, + 0.002626132220029831, + 0.0034677060320973396, + 0.004686026368290186, + 0.003204525448381901, + 0.0034139009658247232, + 0.00427120178937912, + 0.0033618283923715353, + 0.006971124559640884, + 0.006208590231835842, + 0.0066435676999390125, + 0.010795919224619865, + 0.005002847407013178, + 0.010508211329579353, + 0.00990523025393486, + 0.009802810847759247, + 0.005132680293172598, + 0.004365785047411919, + 0.01061957236379385, + 0.005933865439146757, + 0.0033239759504795074, + 0.009275742806494236, + 0.005871823523193598, + 0.01213892363011837, + 0.012370366603136063, + 0.014684626832604408, + 0.012690508738160133, + 0.008770621381700039, + 0.012097285129129887, + 0.008670639246702194, + 0.009038357995450497, + 0.007734759245067835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.021162305027246475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00283946399576962 + ], + "opposite_logprob_tokens": [ + -5.8778395652771 + ], + "logprob_reward": -0.00283946399576962, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00283946399576962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28160637617111206, + 0.11863741278648376, + 0.06461287289857864, + 0.06231559440493584, + 0.06061141565442085, + 0.13425348699092865, + 0.056891366839408875, + 0.14290761947631836, + 0.03489062190055847, + 0.0432732030749321 + ], + "credit_map": [ + 0.017928985878825188, + 0.03203735500574112, + 0.028336673974990845, + 0.04437859728932381, + 0.03632018342614174, + 0.03872930258512497, + 0.04992852360010147, + 0.03394675999879837, + 0.012345049530267715, + 0.009820814244449139, + 0.01546675618737936, + 0.014812220819294453, + 0.011036031879484653, + 0.014836713671684265, + 0.023249439895153046, + 0.017070388421416283, + 0.0052571529522538185, + 0.006152677349746227, + 0.008650130592286587, + 0.011591902934014797, + 0.008754092268645763, + 0.006023000460118055, + 0.011809645220637321, + 0.0063742720521986485, + 0.004735615570098162, + 0.007309052161872387, + 0.006417182739824057, + 0.009070241823792458, + 0.008515588007867336, + 0.007582577411085367, + 0.011583738029003143, + 0.007101600058376789, + 0.005118494853377342, + 0.006249052006751299, + 0.007216470781713724, + 0.0068519823253154755, + 0.008671252988278866, + 0.012062294408679008, + 0.0071503655053675175, + 0.007291502319276333, + 0.005504667293280363, + 0.0026376438327133656, + 0.035240672528743744, + 0.005801721476018429, + 0.00500197010114789, + 0.05255017802119255, + 0.007010957691818476, + 0.020505687221884727, + 0.004736271221190691, + 0.004638861399143934, + 0.010400858707726002, + 0.007605405990034342, + 0.0050180028192698956, + 0.007619243580847979, + 0.007783995009958744, + 0.009088728576898575, + 0.013791438192129135, + 0.01564222201704979, + 0.016896134242415428, + 0.023841483518481255, + 0.012853346765041351, + 0.019664041697978973, + 0.02153189294040203, + 0.018687060102820396, + 0.004068769048899412, + 0.002973933005705476, + 0.007730978541076183, + 0.00402855034917593, + 0.0023030219599604607, + 0.004249900579452515, + 0.003551691072061658, + 0.005983775015920401, + 0.006099033635109663, + 0.006317662075161934, + 0.006251919083297253, + 0.004524942021816969, + 0.007041719276458025, + 0.004741385113447905, + 0.004259844310581684, + 0.0040366994217038155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.00283946399576962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017226393101736903 + ], + "opposite_logprob_tokens": [ + -6.376722812652588 + ], + "logprob_reward": -0.0017226393101736903, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017226393101736903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2603471875190735, + 0.10525409132242203, + 0.025883737951517105, + 0.05626485496759415, + 0.018946265801787376, + 0.24683816730976105, + 0.027237681671977043, + 0.153965026140213, + 0.04801046475768089, + 0.0572524331510067 + ], + "credit_map": [ + 0.018063433468341827, + 0.02423958294093609, + 0.02994137816131115, + 0.03679392486810684, + 0.03522379323840141, + 0.03681434318423271, + 0.04729965701699257, + 0.03197110444307327, + 0.009183412417769432, + 0.01543672289699316, + 0.012606276199221611, + 0.01631085015833378, + 0.010725808329880238, + 0.012669222429394722, + 0.016180604696273804, + 0.012141203507781029, + 0.0025381415616720915, + 0.002722527366131544, + 0.003998003900051117, + 0.003966533113270998, + 0.0028739511035382748, + 0.003447882831096649, + 0.0032272241078317165, + 0.003109476063400507, + 0.004398848861455917, + 0.006994658149778843, + 0.007662773132324219, + 0.008587323129177094, + 0.007521355524659157, + 0.007476212922483683, + 0.007880625315010548, + 0.005743062123656273, + 0.0014406011905521154, + 0.002331205876544118, + 0.003012461820617318, + 0.0029928276780992746, + 0.0017069715540856123, + 0.0032944295089691877, + 0.002307807793840766, + 0.001859961892478168, + 0.009805826470255852, + 0.005302192643284798, + 0.06502752006053925, + 0.008561541326344013, + 0.00913512334227562, + 0.09712691605091095, + 0.014770526438951492, + 0.037108536809682846, + 0.0029458291828632355, + 0.0021718244533985853, + 0.003790438175201416, + 0.005149850621819496, + 0.002274063415825367, + 0.002965709427371621, + 0.004033774137496948, + 0.003906194120645523, + 0.014292031526565552, + 0.016896849498152733, + 0.017437448725104332, + 0.026777053251862526, + 0.014188919216394424, + 0.02134786732494831, + 0.023845955729484558, + 0.019178912043571472, + 0.005791200790554285, + 0.0037569685373455286, + 0.008558793924748898, + 0.0073647527024149895, + 0.004282641224563122, + 0.005917643196880817, + 0.006291074678301811, + 0.006047390401363373, + 0.0074743409641087055, + 0.007962475530803204, + 0.007310174871236086, + 0.006010711658746004, + 0.00919840857386589, + 0.006865278352051973, + 0.007074471563100815, + 0.005356573965400457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0017226393101736903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790684580802917 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790684580802917, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790684580802917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21839575469493866, + 0.07014663517475128, + 0.035115018486976624, + 0.06771229952573776, + 0.0269040297716856, + 0.32878515124320984, + 0.03403248265385628, + 0.13815483450889587, + 0.03622317314147949, + 0.044530630111694336 + ], + "credit_map": [ + 0.014247254468500614, + 0.023373380303382874, + 0.024400506168603897, + 0.03346391022205353, + 0.03034033626317978, + 0.031178947538137436, + 0.03993813320994377, + 0.021453259512782097, + 0.008493123576045036, + 0.008554138243198395, + 0.00723616499453783, + 0.01146874949336052, + 0.007796569727361202, + 0.009010087698698044, + 0.010488225147128105, + 0.007099572103470564, + 0.002111648442223668, + 0.0038506858982145786, + 0.0035211576614528894, + 0.007764324080199003, + 0.004913729149848223, + 0.004003152716904879, + 0.006701463833451271, + 0.002248853910714388, + 0.005270803812891245, + 0.00827784650027752, + 0.010758903808891773, + 0.009335552342236042, + 0.009998461231589317, + 0.007919064722955227, + 0.009706782177090645, + 0.006444883998483419, + 0.002462154719978571, + 0.0034048433881253004, + 0.002913438016548753, + 0.0032898872159421444, + 0.0033906290773302317, + 0.0036078067496418953, + 0.003599771996960044, + 0.004235497210174799, + 0.012195807881653309, + 0.00624628271907568, + 0.06701308488845825, + 0.008518272079527378, + 0.01206928864121437, + 0.13303694128990173, + 0.018484149128198624, + 0.07122129201889038, + 0.003182773245498538, + 0.002610116498544812, + 0.0055361222475767136, + 0.005145612172782421, + 0.0022552453447133303, + 0.004361758939921856, + 0.005135448649525642, + 0.005805404856801033, + 0.011432292871177197, + 0.016552768647670746, + 0.016727183014154434, + 0.02603009343147278, + 0.013256995007395744, + 0.017445048317313194, + 0.02134637162089348, + 0.015364074148237705, + 0.0034389581996947527, + 0.0023193140514194965, + 0.0071684797294437885, + 0.00392572209239006, + 0.0029651587828993797, + 0.0030672599095851183, + 0.004946202039718628, + 0.008392076939344406, + 0.004424016922712326, + 0.00900277029722929, + 0.0042080567218363285, + 0.005564528051763773, + 0.0056062848307192326, + 0.004339899867773056, + 0.008206931874155998, + 0.0031781340949237347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.029790684580802917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043052781373262405 + ], + "opposite_logprob_tokens": [ + -3.1680526733398438 + ], + "logprob_reward": -0.043052781373262405, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043052781373262405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2451857626438141, + 0.05192273482680321, + 0.030269544571638107, + 0.055174149572849274, + 0.05284677445888519, + 0.43908601999282837, + 0.030324190855026245, + 0.046647701412439346, + 0.027120204642415047, + 0.02142288163304329 + ], + "credit_map": [ + 0.015049345791339874, + 0.026323067024350166, + 0.023698383942246437, + 0.039180926978588104, + 0.033512067049741745, + 0.03555819019675255, + 0.04781973734498024, + 0.024044081568717957, + 0.004459679592400789, + 0.006965125910937786, + 0.004598901607096195, + 0.00823872722685337, + 0.006427419371902943, + 0.0076416777446866035, + 0.008415818214416504, + 0.005175389349460602, + 0.002391115529462695, + 0.0033297683112323284, + 0.0035646334290504456, + 0.005804950837045908, + 0.004135500639677048, + 0.002961738733574748, + 0.004798228852450848, + 0.003283610101789236, + 0.004077402409166098, + 0.0053882659412920475, + 0.014378574676811695, + 0.006368815898895264, + 0.007216654252260923, + 0.007165353745222092, + 0.0058588446117937565, + 0.004720241762697697, + 0.0038971537724137306, + 0.004101657774299383, + 0.010470612905919552, + 0.004702964331954718, + 0.0052403612062335014, + 0.01603306084871292, + 0.0037337536923587322, + 0.0046672155149281025, + 0.02092757076025009, + 0.00989984069019556, + 0.05973593890666962, + 0.014337613247334957, + 0.016877667978405952, + 0.16921065747737885, + 0.026149675250053406, + 0.12194705009460449, + 0.0022512965369969606, + 0.0015962611651048064, + 0.006184826139360666, + 0.0049458895809948444, + 0.0031079454347491264, + 0.0037549252156168222, + 0.004679419565945864, + 0.0038036310579627752, + 0.004580477252602577, + 0.006334178149700165, + 0.0051494138315320015, + 0.0082774069160223, + 0.004228340927511454, + 0.005215360317379236, + 0.008052181452512741, + 0.004810350015759468, + 0.0023185221944004297, + 0.002995845163241029, + 0.0043695829808712006, + 0.003389677731320262, + 0.0027073246892541647, + 0.0028468037489801645, + 0.004039444494992495, + 0.004453005734831095, + 0.0021519018337130547, + 0.0035723743494600058, + 0.0022877834271639585, + 0.0024445566814392805, + 0.003441431326791644, + 0.002260912675410509, + 0.0024925393518060446, + 0.0027713824529200792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.043052781373262405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up tracking shot, panning steadily to the right, closely follows a mottled mural painted on a damp, water-stained, rough concrete wall deep within an abandoned subway tunnel. The mural features a simplified human figure, outlined in faded ochre and charcoal black. Its form is distorted and elongated, its surface riddled with fine cracks and patches of flaking. A harsh, cold, white searchlight, emanating from outside the frame, strikes the wall at a sharp angle, piercing through the thick, white simulated smoke that slowly swirls in the air. The light not only highlights the wall's every uneven texture, but also accentuates the mural's texture. Under the gaze of the camera, this human-shaped pattern, which should have been still, is strangely moving autonomously in an incoherent, twitching manner similar to stop-motion animation. It slides horizontally along the wall, and its paint edges seem to dissolve and re-solidify in front, driving tiny concrete fragments and dried paint powder to fall. The whole picture, with its cold tones, creates a strong sense of supernatural suspense and an unsettling claustrophobic atmosphere.", + "seed": 10837, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_4/index_6191_seed_10837_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": -0.02318174745887518, + "vlm_logprob": -0.02318174745887518, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02318174745887518, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029222847893834114 + ], + "opposite_logprob_tokens": [ + -3.7792229652404785 + ], + "logprob_reward": -0.029222847893834114, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.029222847893834114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13459895551204681, + 0.09695287048816681, + 0.13115611672401428, + 0.07739926129579544, + 0.07507248222827911, + 0.08077158778905869, + 0.09426054358482361, + 0.08545374125242233, + 0.06327319890260696, + 0.16106130182743073 + ], + "credit_map": [ + 0.008549468591809273, + 0.0146928196772933, + 0.013014545664191246, + 0.01813281513750553, + 0.02186228148639202, + 0.016232550144195557, + 0.027649197727441788, + 0.014465253800153732, + 0.004716163966804743, + 0.011333970353007317, + 0.010120619088411331, + 0.01299892459064722, + 0.014605444855988026, + 0.01383883599191904, + 0.01658007688820362, + 0.012758823111653328, + 0.008367117494344711, + 0.017086682841181755, + 0.01758875697851181, + 0.026456588879227638, + 0.01755216158926487, + 0.013196107931435108, + 0.015380455181002617, + 0.015528229996562004, + 0.005134148523211479, + 0.007759276777505875, + 0.008198259398341179, + 0.009979930706322193, + 0.006271461024880409, + 0.018381716683506966, + 0.0070611583068966866, + 0.014613297767937183, + 0.00663693668320775, + 0.005768708884716034, + 0.014214258641004562, + 0.014487755484879017, + 0.006726337131112814, + 0.012207480147480965, + 0.007063829340040684, + 0.007967166602611542, + 0.005059604533016682, + 0.0049491459503769875, + 0.008825070224702358, + 0.01222950592637062, + 0.0064865294843912125, + 0.017120299860835075, + 0.014574089087545872, + 0.011527342721819878, + 0.00900724995881319, + 0.007311583496630192, + 0.014553102664649487, + 0.012772409245371819, + 0.010927471332252026, + 0.009559370577335358, + 0.019287796691060066, + 0.010841548442840576, + 0.00509741622954607, + 0.0056517357006669044, + 0.012886498123407364, + 0.01648510806262493, + 0.010241622105240822, + 0.012965652160346508, + 0.011462554335594177, + 0.010663139633834362, + 0.0039809285663068295, + 0.004592309705913067, + 0.009843462146818638, + 0.007823072373867035, + 0.008623568341135979, + 0.010608135722577572, + 0.007881083525717258, + 0.00992063619196415, + 0.010010038502514362, + 0.012434255331754684, + 0.012776226736605167, + 0.013218064792454243, + 0.026837235316634178, + 0.024919454008340836, + 0.02766644023358822, + 0.03319956362247467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.029222847893834114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036376039497554302 + ], + "opposite_logprob_tokens": [ + -5.628637790679932 + ], + "logprob_reward": -0.0036376039497554302, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036376039497554302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10030277818441391, + 0.1458655148744583, + 0.09773267805576324, + 0.07482752948999405, + 0.09920616447925568, + 0.030847955495119095, + 0.06093795597553253, + 0.21394091844558716, + 0.09726081043481827, + 0.07907769083976746 + ], + "credit_map": [ + 0.007327351253479719, + 0.008337601087987423, + 0.010992933064699173, + 0.01667087711393833, + 0.011315048672258854, + 0.013314705342054367, + 0.01756768859922886, + 0.014776580035686493, + 0.006759656593203545, + 0.01328033022582531, + 0.010303863324224949, + 0.022632945328950882, + 0.020868565887212753, + 0.018829461187124252, + 0.033264659345149994, + 0.01992604322731495, + 0.005015525035560131, + 0.009094957262277603, + 0.009538120590150356, + 0.02013765275478363, + 0.017591865733265877, + 0.013545242138206959, + 0.014714882709085941, + 0.00809443648904562, + 0.007733550388365984, + 0.008204076439142227, + 0.007128656841814518, + 0.01030384749174118, + 0.005740619730204344, + 0.014636436477303505, + 0.008106592111289501, + 0.012973750941455364, + 0.008941388688981533, + 0.007156993728131056, + 0.020301148295402527, + 0.01902763918042183, + 0.00886510405689478, + 0.01696932315826416, + 0.008169238455593586, + 0.009775329381227493, + 0.002065656241029501, + 0.0026262968312948942, + 0.003651499282568693, + 0.0034120928030461073, + 0.004095206968486309, + 0.004869986325502396, + 0.006540586706250906, + 0.0035866310354322195, + 0.00465221144258976, + 0.0049297702498734, + 0.009163472801446915, + 0.011616012081503868, + 0.00651380093768239, + 0.008460472337901592, + 0.010006153024733067, + 0.005596064031124115, + 0.015318755991756916, + 0.013810435310006142, + 0.05530732497572899, + 0.02726559527218342, + 0.02954685129225254, + 0.02290254831314087, + 0.021643396466970444, + 0.028146017342805862, + 0.006835888605564833, + 0.006806692108511925, + 0.021690040826797485, + 0.010794006288051605, + 0.014487653970718384, + 0.011868617497384548, + 0.011639771051704884, + 0.013138145208358765, + 0.004089283291250467, + 0.00455847755074501, + 0.005949281621724367, + 0.0056450278498232365, + 0.008564344607293606, + 0.008596937172114849, + 0.018464377149939537, + 0.02320997044444084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0036376039497554302 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028275768272578716 + ], + "opposite_logprob_tokens": [ + -5.8778276443481445 + ], + "logprob_reward": -0.0028275768272578716, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028275768272578716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13467834889888763, + 0.07569807022809982, + 0.06971131265163422, + 0.06459823995828629, + 0.10215604305267334, + 0.04406784102320671, + 0.06530576944351196, + 0.23394396901130676, + 0.06962690502405167, + 0.14021345973014832 + ], + "credit_map": [ + 0.010119758546352386, + 0.011245775036513805, + 0.014138604514300823, + 0.020088281482458115, + 0.019825618714094162, + 0.019856754690408707, + 0.020945921540260315, + 0.018457645550370216, + 0.003914841450750828, + 0.0064654103480279446, + 0.006332005839794874, + 0.011814868077635765, + 0.01192233245819807, + 0.01055123656988144, + 0.013592764735221863, + 0.01110462099313736, + 0.004388698376715183, + 0.006473160348832607, + 0.005627375561743975, + 0.017684271559119225, + 0.006831703707575798, + 0.008326348848640919, + 0.012185229919850826, + 0.008194533176720142, + 0.006485366262495518, + 0.0062049985863268375, + 0.00667803268879652, + 0.011797010898590088, + 0.0045591010712087154, + 0.011216901242733002, + 0.008617740124464035, + 0.009039100259542465, + 0.008541153743863106, + 0.006689815316349268, + 0.020676296204328537, + 0.019276341423392296, + 0.009457720443606377, + 0.019408678635954857, + 0.007141221314668655, + 0.010964829474687576, + 0.0035866787657141685, + 0.002919226186349988, + 0.005475956480950117, + 0.006144134793430567, + 0.005167668219655752, + 0.007894843816757202, + 0.00729797687381506, + 0.005581360310316086, + 0.006730531807988882, + 0.0055476087145507336, + 0.01108470093458891, + 0.01158645935356617, + 0.0071472725830972195, + 0.007862092927098274, + 0.00910836923867464, + 0.006238739006221294, + 0.014567417092621326, + 0.012937597930431366, + 0.07632437348365784, + 0.023591192439198494, + 0.022670481353998184, + 0.022046290338039398, + 0.019770806655287743, + 0.042035818099975586, + 0.005169377196580172, + 0.005423700902611017, + 0.015854312106966972, + 0.008502659387886524, + 0.009979775175452232, + 0.007449619472026825, + 0.008712653070688248, + 0.008534814231097698, + 0.00688720028847456, + 0.00820154044777155, + 0.012074670754373074, + 0.010206087492406368, + 0.02036377601325512, + 0.013153832405805588, + 0.02790411189198494, + 0.04142225161194801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.0028275768272578716 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062035512179136276 + ], + "opposite_logprob_tokens": [ + -2.81203556060791 + ], + "logprob_reward": -0.062035512179136276, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062035512179136276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14979670941829681, + 0.17042416334152222, + 0.06935170292854309, + 0.17825038731098175, + 0.04374708980321884, + 0.0553547665476799, + 0.05043633282184601, + 0.1444365233182907, + 0.03948681801557541, + 0.09871544688940048 + ], + "credit_map": [ + 0.015670740976929665, + 0.01607002504169941, + 0.01657559536397457, + 0.02239314280450344, + 0.013540118932723999, + 0.023703424260020256, + 0.02335132472217083, + 0.01849234849214554, + 0.006963667925447226, + 0.01692516915500164, + 0.014566009864211082, + 0.023891042917966843, + 0.02422596886754036, + 0.023037610575556755, + 0.0338631197810173, + 0.026951588690280914, + 0.004265984985977411, + 0.00738281337544322, + 0.010102794505655766, + 0.013628796674311161, + 0.011832849122583866, + 0.006377608515322208, + 0.008813784457743168, + 0.006947077810764313, + 0.016811180859804153, + 0.015268257819116116, + 0.015427647158503532, + 0.03531121462583542, + 0.011775190010666847, + 0.0299138892441988, + 0.027388490736484528, + 0.02635454386472702, + 0.0019082543440163136, + 0.0028815018013119698, + 0.004292820114642382, + 0.005495346151292324, + 0.00406279182061553, + 0.009117810986936092, + 0.006251347716897726, + 0.009737220592796803, + 0.004939709324389696, + 0.006234698463231325, + 0.005978177767246962, + 0.011538471095263958, + 0.004430444445461035, + 0.006259363144636154, + 0.009328793734312057, + 0.006645109038800001, + 0.0033263356890529394, + 0.0029257165733724833, + 0.010720492340624332, + 0.008555815555155277, + 0.005370222497731447, + 0.005760638043284416, + 0.00983928982168436, + 0.003937823232263327, + 0.009455802850425243, + 0.008482208475470543, + 0.05062451586127281, + 0.01633579283952713, + 0.012834168039262295, + 0.012477799318730831, + 0.011934188194572926, + 0.022292057052254677, + 0.00392037071287632, + 0.0032407117541879416, + 0.006331346463412046, + 0.006803684402257204, + 0.004296610131859779, + 0.006082272157073021, + 0.004094439558684826, + 0.004717384465038776, + 0.004551047924906015, + 0.004553921055048704, + 0.011241234838962555, + 0.011847639456391335, + 0.011072119697928429, + 0.009155670180916786, + 0.021297689527273178, + 0.024996129795908928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.062035512179136276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 an abandoned subway tunnel featuring a damp, water-stained, rough concrete wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mottled mural visible on the wall featuring a simplified human figure outlined in faded ochre and charcoal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the human figure in the mural distorted, elongated, and riddled with fine cracks and patches of flaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the human figure move autonomously in an incoherent, twitching manner similar to stop-motion animation, sliding horizontally and causing paint edges to dissolve and re-solidify?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a tight, close-up tracking shot that pans steadily to the right while a harsh white searchlight strikes the wall at a sharp angle through swirling smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185196444392204 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185196444392204, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185196444392204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10849131643772125, + 0.06086360663175583, + 0.0990002304315567, + 0.06795211881399155, + 0.062423188239336014, + 0.05104168504476547, + 0.08683910965919495, + 0.17115655541419983, + 0.04532590135931969, + 0.24690629541873932 + ], + "credit_map": [ + 0.009550787508487701, + 0.012175055220723152, + 0.010140204802155495, + 0.01827307790517807, + 0.012369549833238125, + 0.018792733550071716, + 0.017474541440606117, + 0.009715372696518898, + 0.0032174030784517527, + 0.006499819457530975, + 0.0070951711386442184, + 0.010925955139100552, + 0.007157306186854839, + 0.008387843146920204, + 0.008885340765118599, + 0.008694764226675034, + 0.007114621810615063, + 0.0128119345754385, + 0.012388448230922222, + 0.019298812374472618, + 0.013055172748863697, + 0.010047229938209057, + 0.012233337387442589, + 0.012050668708980083, + 0.004547727294266224, + 0.0038989672902971506, + 0.00745442183688283, + 0.010671686381101608, + 0.005565493367612362, + 0.01346291322261095, + 0.008830546401441097, + 0.013520356267690659, + 0.005204465240240097, + 0.005243921652436256, + 0.012031028047204018, + 0.011064330115914345, + 0.006404713727533817, + 0.008732970803976059, + 0.004913080483675003, + 0.008828675374388695, + 0.003569032298400998, + 0.002634226344525814, + 0.007719087414443493, + 0.006948450114578009, + 0.0035128705203533173, + 0.008639994077384472, + 0.013103140518069267, + 0.004914884455502033, + 0.00701158307492733, + 0.007495447527617216, + 0.014407852664589882, + 0.013340387493371964, + 0.008738256990909576, + 0.00944615714251995, + 0.018961848691105843, + 0.007437573280185461, + 0.010516172274947166, + 0.010754711925983429, + 0.04523128643631935, + 0.01933222822844982, + 0.02375270426273346, + 0.018178941681981087, + 0.01676180772483349, + 0.02662871778011322, + 0.0030593611299991608, + 0.0039491597563028336, + 0.005125900264829397, + 0.00594053091481328, + 0.006505548022687435, + 0.006048679817467928, + 0.0061277844943106174, + 0.00856893789023161, + 0.01058169361203909, + 0.01987900584936142, + 0.01871352270245552, + 0.03439018875360489, + 0.052031517028808594, + 0.029353903606534004, + 0.0438496358692646, + 0.038106828927993774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.018185196444392204 + } + ], + "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/0000025/group_5/index_tensor([5682, 5682])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_5/index_tensor([5682, 5682])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..93903bd499530658881d60569f53207df7cf24fe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_5/index_tensor([5682, 5682])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic montage composed of a series of rapid, staggered extreme close-ups, the camera focused with a shallow depth of field on the fronds of a dense fern after a rainstorm. In the first shot, a delicate, pinnately divided fern frond dominates the frame, its surface coated in a thin film of water that makes the leaf's emerald green appear unusually saturated and deep. A crystal-clear, nearly perfectly spherical drop of water hangs delicately on the frond's tip, quivering slightly due to surface tension, its interior reflecting a miniature, distorted image of the veins beneath. A barely perceptible breeze gently tugs at the frond, swaying the droplet and causing light to play across its surface. The camera then cuts to a close-up of another fern frond, a string of even smaller droplets arranged along its central axis like a diamond necklace. Soft, diffused skylight evenly illuminates the scene, eliminating any distinct shadows and highlighting the clarity of each drop and the wet sheen of the frond. The entire sequence jumps rapidly between the microscopic details of different leaves, creating a rhythm that is both tranquil and full of vitality, conveying the fresh, clean and revitalized tranquil atmosphere of the forest floor after the rain.", + "seed": 20366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_5/index_5682_seed_20366_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": -0.04196493797935545, + "vlm_logprob": -0.04196493797935545, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.04196493797935545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.11615654826164246 + ], + "opposite_logprob_tokens": [ + -2.366156578063965 + ], + "logprob_reward": -0.11615654826164246, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11615654826164246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07330884039402008, + 0.09470462054014206, + 0.045137178152799606, + 0.0376819409430027, + 0.044976845383644104, + 0.03180782496929169, + 0.04859721288084984, + 0.22803232073783875, + 0.32775041460990906, + 0.06800273060798645 + ], + "credit_map": [ + 0.0072667477652430534, + 0.00670640729367733, + 0.01143238041549921, + 0.013743148185312748, + 0.006097183097153902, + 0.006556048523634672, + 0.007918400689959526, + 0.013588540256023407, + 0.010705247521400452, + 0.015910152345895767, + 0.014861474744975567, + 0.013745150528848171, + 0.008149438537657261, + 0.011715713888406754, + 0.011871355585753918, + 0.007746108807623386, + 0.004711775574833155, + 0.004193149972707033, + 0.006615025922656059, + 0.007409502286463976, + 0.008158047683537006, + 0.0032394607551395893, + 0.006242540664970875, + 0.004567683674395084, + 0.004138000775128603, + 0.007461564615368843, + 0.005797834135591984, + 0.005447488743811846, + 0.004526292439550161, + 0.004138529300689697, + 0.0026360047049820423, + 0.003536236472427845, + 0.006677302997559309, + 0.007780924439430237, + 0.0055412957444787025, + 0.005072210915386677, + 0.006284080911427736, + 0.005197230726480484, + 0.0043300604447722435, + 0.004093748517334461, + 0.005933962296694517, + 0.0036120875738561153, + 0.0024178477469831705, + 0.00396066065877676, + 0.002783979056403041, + 0.005780512932687998, + 0.004427651409059763, + 0.002891127485781908, + 0.005107237491756678, + 0.004934289492666721, + 0.0033534204121679068, + 0.005593996960669756, + 0.004401576239615679, + 0.0065789008513092995, + 0.010214555077254772, + 0.008413244970142841, + 0.011229321360588074, + 0.01161195058375597, + 0.02170741930603981, + 0.025628728792071342, + 0.03651602938771248, + 0.023465774953365326, + 0.06503438949584961, + 0.032838743180036545, + 0.022572360932826996, + 0.035052184015512466, + 0.01583319902420044, + 0.02847733534872532, + 0.1114555299282074, + 0.0690682902932167, + 0.025747312232851982, + 0.01954425498843193, + 0.005733071360737085, + 0.006586211733520031, + 0.009945648722350597, + 0.006577151361852884, + 0.005510563496500254, + 0.005268916487693787, + 0.00955433864146471, + 0.018826846033334732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.11615654826164246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.0028229409363120794 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.0028229409363120794, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028229409363120794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05487988889217377, + 0.18863599002361298, + 0.08720782399177551, + 0.042403653264045715, + 0.03299802541732788, + 0.026468079537153244, + 0.037586722522974014, + 0.07652223855257034, + 0.22424077987670898, + 0.22905679047107697 + ], + "credit_map": [ + 0.005031832028180361, + 0.005498681217432022, + 0.007519711274653673, + 0.008703788742423058, + 0.005052541848272085, + 0.005984074901789427, + 0.005384215619415045, + 0.011705044656991959, + 0.020484216511249542, + 0.029383117333054543, + 0.03151308372616768, + 0.03070797771215439, + 0.01856130361557007, + 0.020045017823576927, + 0.02198183536529541, + 0.015959447249770164, + 0.010922913439571857, + 0.007799258455634117, + 0.011645078659057617, + 0.01140796672552824, + 0.012765511870384216, + 0.008783555589616299, + 0.013882909901440144, + 0.010000641457736492, + 0.004595390986651182, + 0.008657711558043957, + 0.006083998363465071, + 0.0076421401463449, + 0.0045767975971102715, + 0.003305203514173627, + 0.0033793619368225336, + 0.004163054283708334, + 0.004098810255527496, + 0.004857806488871574, + 0.00402602506801486, + 0.00526792835444212, + 0.0034309898037463427, + 0.006758988369256258, + 0.0023757494054734707, + 0.0021817300003021955, + 0.003970080055296421, + 0.004035726189613342, + 0.0025620548985898495, + 0.004881602246314287, + 0.0025025950744748116, + 0.003014693036675453, + 0.003033761866390705, + 0.002467568963766098, + 0.004256771877408028, + 0.003469289280474186, + 0.0028530957642942667, + 0.004938535857945681, + 0.003775162622332573, + 0.004775430541485548, + 0.007813300006091595, + 0.005705135874450207, + 0.00603057025000453, + 0.0071063777431845665, + 0.010574045591056347, + 0.007963747717440128, + 0.010385849513113499, + 0.008741809986531734, + 0.01629951037466526, + 0.009420323185622692, + 0.01634601131081581, + 0.021241921931505203, + 0.01423593983054161, + 0.02413458749651909, + 0.07250897586345673, + 0.043445173650979996, + 0.01790701225399971, + 0.014421176165342331, + 0.018428556621074677, + 0.023129882290959358, + 0.03339368477463722, + 0.023449162021279335, + 0.019455570727586746, + 0.018595248460769653, + 0.032968442887067795, + 0.059636253863573074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.0028229409363120794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.0041007245890796185 + ], + "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": [ + 0.21073919534683228, + 0.10534107685089111, + 0.04549035057425499, + 0.03700144216418266, + 0.044104546308517456, + 0.040286410599946976, + 0.09982842952013016, + 0.024733781814575195, + 0.27191177010536194, + 0.12056294828653336 + ], + "credit_map": [ + 0.016134049743413925, + 0.015608851797878742, + 0.024355214089155197, + 0.04778415709733963, + 0.01512601412832737, + 0.016192017123103142, + 0.027294998988509178, + 0.04824388772249222, + 0.010264557786285877, + 0.015063784085214138, + 0.017834482714533806, + 0.013612904585897923, + 0.010003776289522648, + 0.014327061362564564, + 0.01233095396310091, + 0.011903563514351845, + 0.005096609238535166, + 0.0034218113869428635, + 0.006147890351712704, + 0.006343903951346874, + 0.006108605302870274, + 0.003670213744044304, + 0.00916000921279192, + 0.005541307386010885, + 0.003962766844779253, + 0.0051263184286653996, + 0.00474415672942996, + 0.005813605152070522, + 0.003839119803160429, + 0.0048681325279176235, + 0.0041444250382483006, + 0.004502920899540186, + 0.004062292166054249, + 0.006470119580626488, + 0.0035486549604684114, + 0.008065386675298214, + 0.005328180268406868, + 0.008178552612662315, + 0.0050937943160533905, + 0.0033575715497136116, + 0.006562904920428991, + 0.009375172667205334, + 0.002745922887697816, + 0.0037367157638072968, + 0.0030555459670722485, + 0.005667104851454496, + 0.004626541864126921, + 0.004516507498919964, + 0.012168025597929955, + 0.012131677009165287, + 0.007962396368384361, + 0.010300782509148121, + 0.00691763311624527, + 0.011956978589296341, + 0.019182784482836723, + 0.019208155572414398, + 0.0014685435453429818, + 0.0014700415777042508, + 0.0023148031905293465, + 0.0028283074498176575, + 0.0025773399975150824, + 0.0036379341036081314, + 0.0067983875051140785, + 0.0036384265404194593, + 0.02065541408956051, + 0.02424445003271103, + 0.015194658190011978, + 0.027784312143921852, + 0.09068967401981354, + 0.05453579127788544, + 0.020942246541380882, + 0.01786523498594761, + 0.009747451171278954, + 0.01171890739351511, + 0.01767187938094139, + 0.01439156848937273, + 0.011482525616884232, + 0.01070069707930088, + 0.015953712165355682, + 0.028896216303110123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0041007245890796185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.0486467182636261 + ], + "opposite_logprob_tokens": [ + -3.0486466884613037 + ], + "logprob_reward": -0.0486467182636261, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486467182636261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023190876469016075, + 0.034808896481990814, + 0.012472686357796192, + 0.025059813633561134, + 0.02311793528497219, + 0.018512912094593048, + 0.02329881303012371, + 0.0552198551595211, + 0.7593342065811157, + 0.024984097108244896 + ], + "credit_map": [ + 0.002304168650880456, + 0.002134599955752492, + 0.0037465172354131937, + 0.005466917064040899, + 0.0016083645168691874, + 0.0018737853970378637, + 0.002731033368036151, + 0.0033254874870181084, + 0.0042779818177223206, + 0.005633884109556675, + 0.004657769575715065, + 0.004442137666046619, + 0.0031609702855348587, + 0.004824821371585131, + 0.0043228548020124435, + 0.0034884733613580465, + 0.0012243574019521475, + 0.001193172880448401, + 0.0021178824827075005, + 0.0019730012863874435, + 0.002094281604513526, + 0.0010475094895809889, + 0.001538582961075008, + 0.001283897552639246, + 0.002334139309823513, + 0.005285654682666063, + 0.0030423260759562254, + 0.004357644356787205, + 0.0028030637186020613, + 0.002286314032971859, + 0.002339322119951248, + 0.0026113472413271666, + 0.002607643837109208, + 0.003929461818188429, + 0.0024638960603624582, + 0.002519200788810849, + 0.0029409904964268208, + 0.004234252963215113, + 0.002228120341897011, + 0.0021943668834865093, + 0.0033921480644494295, + 0.003965924959629774, + 0.0018336412031203508, + 0.0025509705301374197, + 0.0014192131347954273, + 0.0024421210400760174, + 0.0016020479379221797, + 0.0013068437110632658, + 0.0024308660067617893, + 0.0031268284656107426, + 0.0016603234689682722, + 0.0025887128431349993, + 0.002090177498757839, + 0.0025718724355101585, + 0.004914672579616308, + 0.00391535647213459, + 0.003204027656465769, + 0.0036655529402196407, + 0.004950990434736013, + 0.006197198759764433, + 0.00920324306935072, + 0.006690143141895533, + 0.012656226754188538, + 0.00865247007459402, + 0.05157214403152466, + 0.07639865577220917, + 0.04350239038467407, + 0.06972233951091766, + 0.2600092887878418, + 0.15878717601299286, + 0.05974937602877617, + 0.0395926758646965, + 0.0024843565188348293, + 0.0030374806374311447, + 0.003658008761703968, + 0.004474043846130371, + 0.002510884078219533, + 0.0019229911267757416, + 0.00277533452026546, + 0.00412099389359355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.0486467182636261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.03809775784611702 + ], + "opposite_logprob_tokens": [ + -3.288097858428955 + ], + "logprob_reward": -0.03809775784611702, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809775784611702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0931067243218422, + 0.042505260556936264, + 0.05038824677467346, + 0.06735308468341827, + 0.06578657776117325, + 0.026323925703763962, + 0.024952426552772522, + 0.07930399477481842, + 0.0471181757748127, + 0.5031615495681763 + ], + "credit_map": [ + 0.008889302611351013, + 0.008150502108037472, + 0.014747468754649162, + 0.017317604273557663, + 0.007595642004162073, + 0.008037136867642403, + 0.012988545931875706, + 0.015380521304905415, + 0.003920535556972027, + 0.007439020089805126, + 0.005129752214998007, + 0.005557991564273834, + 0.004057943820953369, + 0.00743845384567976, + 0.004703194834291935, + 0.004258365835994482, + 0.00547371618449688, + 0.004350218456238508, + 0.008941585198044777, + 0.01009519211947918, + 0.007647410035133362, + 0.002569237956777215, + 0.00480790575966239, + 0.006502984091639519, + 0.006284270901232958, + 0.013734331354498863, + 0.009064805693924427, + 0.010692157782614231, + 0.0076573570258915424, + 0.006597076542675495, + 0.005912798456847668, + 0.007410281337797642, + 0.00822236854583025, + 0.009169376455247402, + 0.005864969454705715, + 0.008779392577707767, + 0.006895304657518864, + 0.014515762217342854, + 0.005265654996037483, + 0.007073747459799051, + 0.005088783800601959, + 0.0031375668477267027, + 0.0021258711349219084, + 0.0040011717937886715, + 0.003093981184065342, + 0.003637097543105483, + 0.003111938014626503, + 0.002127514686435461, + 0.0025221863761544228, + 0.0021286916453391314, + 0.0019101537764072418, + 0.0029227687045931816, + 0.001929098623804748, + 0.002839909866452217, + 0.005801779683679342, + 0.004897839389741421, + 0.005738007836043835, + 0.006743232719600201, + 0.010992071591317654, + 0.007943693548440933, + 0.012576243840157986, + 0.009186913259327412, + 0.014767730608582497, + 0.011356092989444733, + 0.0033960342407226562, + 0.005585136823356152, + 0.0034613683819770813, + 0.005276930518448353, + 0.009189852513372898, + 0.006913356017321348, + 0.006492870859801769, + 0.006802629213780165, + 0.04045099765062332, + 0.04532497376203537, + 0.07538522034883499, + 0.04704037308692932, + 0.0385512039065361, + 0.03909965232014656, + 0.07684702426195145, + 0.14046213030815125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.03809775784611702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic montage composed of a series of rapid, staggered extreme close-ups, the camera focused with a shallow depth of field on the fronds of a dense fern after a rainstorm. In the first shot, a delicate, pinnately divided fern frond dominates the frame, its surface coated in a thin film of water that makes the leaf's emerald green appear unusually saturated and deep. A crystal-clear, nearly perfectly spherical drop of water hangs delicately on the frond's tip, quivering slightly due to surface tension, its interior reflecting a miniature, distorted image of the veins beneath. A barely perceptible breeze gently tugs at the frond, swaying the droplet and causing light to play across its surface. The camera then cuts to a close-up of another fern frond, a string of even smaller droplets arranged along its central axis like a diamond necklace. Soft, diffused skylight evenly illuminates the scene, eliminating any distinct shadows and highlighting the clarity of each drop and the wet sheen of the frond. The entire sequence jumps rapidly between the microscopic details of different leaves, creating a rhythm that is both tranquil and full of vitality, conveying the fresh, clean and revitalized tranquil atmosphere of the forest floor after the rain.", + "seed": 20367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_5/index_5682_seed_20367_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": -0.031310719437897205, + "vlm_logprob": -0.031310719437897205, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.031310719437897205, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.06896279752254486 + ], + "opposite_logprob_tokens": [ + -2.943962812423706 + ], + "logprob_reward": -0.06896279752254486, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06896279752254486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04291606694459915, + 0.02078435942530632, + 0.03884308040142059, + 0.16579367220401764, + 0.03318466991186142, + 0.5650056600570679, + 0.03595424070954323, + 0.029978634789586067, + 0.048723798245191574, + 0.01881588064134121 + ], + "credit_map": [ + 0.0038714329712092876, + 0.0049422746524214745, + 0.008631263859570026, + 0.0054575917311012745, + 0.006199250463396311, + 0.0050773099064826965, + 0.004969225265085697, + 0.0037677157670259476, + 0.0018004035810008645, + 0.001896703033708036, + 0.004132830537855625, + 0.0038205315358936787, + 0.002754745539277792, + 0.00211740517988801, + 0.0026991446502506733, + 0.0015625940868631005, + 0.003252436174079776, + 0.005421006586402655, + 0.0053502474911510944, + 0.007832658477127552, + 0.004396175034344196, + 0.0042212302796542645, + 0.0052513545379042625, + 0.003117968328297138, + 0.03884631022810936, + 0.01438815426081419, + 0.012352275662124157, + 0.019633345305919647, + 0.03673553094267845, + 0.015720048919320107, + 0.016979938372969627, + 0.011138048954308033, + 0.003772047581151128, + 0.005421438720077276, + 0.004174650181084871, + 0.004339267965406179, + 0.0040627713315188885, + 0.0031400974839925766, + 0.004193073604255915, + 0.0040813228115439415, + 0.05155766382813454, + 0.033129360526800156, + 0.027491513639688492, + 0.16388565301895142, + 0.02687145583331585, + 0.21138964593410492, + 0.022959288209676743, + 0.027721019461750984, + 0.0022302039433270693, + 0.0047700060531497, + 0.0035337801091372967, + 0.006341729778796434, + 0.009736756794154644, + 0.0025701236445456743, + 0.0038732814136892557, + 0.002898354781791568, + 0.0032685634214431047, + 0.002891610376536846, + 0.004623243119567633, + 0.0026592984795570374, + 0.004794072359800339, + 0.005280544515699148, + 0.0035188861656934023, + 0.0029424121603369713, + 0.006133907940238714, + 0.0050421105697751045, + 0.005871973931789398, + 0.005438037682324648, + 0.007742236368358135, + 0.005717301741242409, + 0.004156733863055706, + 0.008621488697826862, + 0.0031319372355937958, + 0.0021309126168489456, + 0.0026022023521363735, + 0.002280953573063016, + 0.0021381357219070196, + 0.003264258848503232, + 0.0016577988862991333, + 0.0016096786130219698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.06896279752254486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.0076345098204910755 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.0076345098204910755, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076345098204910755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017670536413788795, + 0.01495269127190113, + 0.05021664872765541, + 0.15447625517845154, + 0.04601801559329033, + 0.585080623626709, + 0.022870244458317757, + 0.047932542860507965, + 0.047011133283376694, + 0.013771231286227703 + ], + "credit_map": [ + 0.001341907656751573, + 0.0019907797686755657, + 0.0019602158572524786, + 0.002552727470174432, + 0.0024862869177013636, + 0.002949126996099949, + 0.0022006460931152105, + 0.002188845770433545, + 0.0015178888570517302, + 0.0015235445462167263, + 0.0022329166531562805, + 0.002887887414544821, + 0.0017545855371281505, + 0.0019732315558940172, + 0.0018860127311199903, + 0.0011766242096200585, + 0.004397326614707708, + 0.006618066690862179, + 0.007258130703121424, + 0.006703962571918964, + 0.006236241664737463, + 0.009320368990302086, + 0.006027159746736288, + 0.0036553952377289534, + 0.03862553834915161, + 0.012021920643746853, + 0.011121409013867378, + 0.013581646606326103, + 0.008412271738052368, + 0.008458729833364487, + 0.02049671858549118, + 0.04175802320241928, + 0.0048016211949288845, + 0.005867186468094587, + 0.004444710444658995, + 0.005288345273584127, + 0.006141130346804857, + 0.008084986358880997, + 0.006260058842599392, + 0.005129973869770765, + 0.04554566740989685, + 0.031820956617593765, + 0.028941381722688675, + 0.17879047989845276, + 0.02736533060669899, + 0.21937038004398346, + 0.024221522733569145, + 0.029024921357631683, + 0.0016383819747716188, + 0.0031042483169585466, + 0.003081021597608924, + 0.004412067122757435, + 0.003936196677386761, + 0.0016455232398584485, + 0.0027585618663579226, + 0.002294243313372135, + 0.006133574061095715, + 0.004573184531182051, + 0.0048930346965789795, + 0.0033494301605969667, + 0.0078011201694607735, + 0.012623125687241554, + 0.00375385326333344, + 0.004805220756679773, + 0.005857354495674372, + 0.0047695753164589405, + 0.005924449302256107, + 0.004611429758369923, + 0.00812500435858965, + 0.005903421435505152, + 0.00341287674382329, + 0.008407022804021835, + 0.002494256943464279, + 0.0015589752001687884, + 0.0016457276651635766, + 0.0015334895579144359, + 0.0019664007704705, + 0.0015884296735748649, + 0.0014192614471539855, + 0.0015646903775632381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.0076345098204910755 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.003200054634362459 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.003200054634362459, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003200054634362459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04304041340947151, + 0.04592743143439293, + 0.0667814388871193, + 0.22691121697425842, + 0.0353834368288517, + 0.4187701642513275, + 0.06573668122291565, + 0.04417891427874565, + 0.03828507661819458, + 0.014985217712819576 + ], + "credit_map": [ + 0.004846648778766394, + 0.004754369612783194, + 0.010067370720207691, + 0.0080387182533741, + 0.0044587356969714165, + 0.00394843053072691, + 0.003317804541438818, + 0.0036083392333239317, + 0.005699698347598314, + 0.008159484714269638, + 0.005778333637863398, + 0.005719909444451332, + 0.004381169565021992, + 0.0053860186599195, + 0.00759492814540863, + 0.003207889385521412, + 0.0073395841754972935, + 0.011336966417729855, + 0.008869725279510021, + 0.012868420220911503, + 0.006352463737130165, + 0.006060175597667694, + 0.008601145818829536, + 0.0053529576398432255, + 0.06257276237010956, + 0.018395062536001205, + 0.009804734960198402, + 0.01237532775849104, + 0.011202487163245678, + 0.012658022344112396, + 0.0319458432495594, + 0.06795697659254074, + 0.003425943199545145, + 0.005667082499712706, + 0.004869014024734497, + 0.005607662256807089, + 0.004263534210622311, + 0.005303296260535717, + 0.0033169095404446125, + 0.002929993439465761, + 0.03382081165909767, + 0.01995404064655304, + 0.022038737311959267, + 0.12911392748355865, + 0.01957853138446808, + 0.15499825775623322, + 0.017192739993333817, + 0.02207314595580101, + 0.0036277954932302237, + 0.010233326815068722, + 0.005618898663669825, + 0.010659629479050636, + 0.01944335363805294, + 0.004944708663970232, + 0.006573759019374847, + 0.00463521433994174, + 0.004938230849802494, + 0.004937843419611454, + 0.004338644444942474, + 0.0047963154502213, + 0.005818514619022608, + 0.005869114305824041, + 0.006028444040566683, + 0.007451808080077171, + 0.004545678850263357, + 0.00401205662637949, + 0.005844620987772942, + 0.004081455059349537, + 0.00704146409407258, + 0.004699208773672581, + 0.003338867798447609, + 0.004721721168607473, + 0.0015533027471974492, + 0.0017404177924618125, + 0.0018616748275235295, + 0.002232693135738373, + 0.001759804436005652, + 0.002043802523985505, + 0.0024747513234615326, + 0.0013187716249376535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.003200054634362459 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.043059516698122025 + ], + "opposite_logprob_tokens": [ + -3.1680595874786377 + ], + "logprob_reward": -0.043059516698122025, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043059516698122025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025348542258143425, + 0.029891567304730415, + 0.044849127531051636, + 0.09459604322910309, + 0.03801765665411949, + 0.5524550080299377, + 0.03700706735253334, + 0.05800497531890869, + 0.11233744025230408, + 0.007492606993764639 + ], + "credit_map": [ + 0.002132440684363246, + 0.0022987499833106995, + 0.004114748910069466, + 0.004045870155096054, + 0.0025129811838269234, + 0.0027236698660999537, + 0.005352654494345188, + 0.0021674255840480328, + 0.0034342475701123476, + 0.005012426525354385, + 0.0047659119591116905, + 0.004403489641845226, + 0.0032197057735174894, + 0.00333048845641315, + 0.003475046483799815, + 0.002250249730423093, + 0.0032677981071174145, + 0.005572916939854622, + 0.0065001314505934715, + 0.005328466184437275, + 0.007180305663496256, + 0.0071548400446772575, + 0.007131799124181271, + 0.0027128702495247126, + 0.020838236436247826, + 0.006536736618727446, + 0.00916951522231102, + 0.012393188662827015, + 0.010127966292202473, + 0.005655658431351185, + 0.011807290837168694, + 0.018067458644509315, + 0.003916935063898563, + 0.0056847985833883286, + 0.004287580493837595, + 0.004951682407408953, + 0.0050000399351119995, + 0.005685275420546532, + 0.004608505871146917, + 0.0038828374817967415, + 0.05627197027206421, + 0.026664869859814644, + 0.026288120076060295, + 0.15597355365753174, + 0.026076991111040115, + 0.21316054463386536, + 0.02150668576359749, + 0.02651224099099636, + 0.00279350602068007, + 0.0057731200940907, + 0.006135362666100264, + 0.0039612483233213425, + 0.003952228929847479, + 0.003923287149518728, + 0.005274063441902399, + 0.005194253288209438, + 0.007766273804008961, + 0.006225356366485357, + 0.008204986341297626, + 0.006056249141693115, + 0.007120187394320965, + 0.008603657595813274, + 0.00846863817423582, + 0.0055596292950212955, + 0.013410999439656734, + 0.011978983879089355, + 0.013736912980675697, + 0.012191438116133213, + 0.020309003069996834, + 0.013913199305534363, + 0.009431170299649239, + 0.017365742474794388, + 0.0008966257446445525, + 0.0010827239602804184, + 0.0008992028306238353, + 0.0010064832167699933, + 0.0009119801106862724, + 0.0011646498460322618, + 0.0008752864087000489, + 0.0006556548760272563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.043059516698122025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.03369671851396561 + ], + "opposite_logprob_tokens": [ + -3.4086966514587402 + ], + "logprob_reward": -0.03369671851396561, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369671851396561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01845656707882881, + 0.009739969857037067, + 0.02423226647078991, + 0.11547248810529709, + 0.016399867832660675, + 0.7175421714782715, + 0.019887400791049004, + 0.02496541477739811, + 0.04587153345346451, + 0.007432347629219294 + ], + "credit_map": [ + 0.002185977529734373, + 0.002237931592389941, + 0.003114457707852125, + 0.002059112535789609, + 0.002188224345445633, + 0.003363339463248849, + 0.0014752446440979838, + 0.0018322780961170793, + 0.0011845006374642253, + 0.0010304811876267195, + 0.0013432276900857687, + 0.0014204318867996335, + 0.0013324072351679206, + 0.0011989633785560727, + 0.0014261911856010556, + 0.0008037660154514015, + 0.0022364319302141666, + 0.00225252378731966, + 0.002918725600466132, + 0.003469937713816762, + 0.003002151846885681, + 0.004869426134973764, + 0.0032853351440280676, + 0.002197731751948595, + 0.023844163864850998, + 0.01096536498516798, + 0.0036968891508877277, + 0.0048696547746658325, + 0.00827043317258358, + 0.007471011020243168, + 0.019750015810132027, + 0.03660494461655617, + 0.001967592863366008, + 0.001981765730306506, + 0.001726703834719956, + 0.0020391664002090693, + 0.0017779911868274212, + 0.002582569606602192, + 0.0018174175638705492, + 0.00250665913335979, + 0.0675237849354744, + 0.03517993539571762, + 0.03555699810385704, + 0.20889702439308167, + 0.033504512161016464, + 0.27405449748039246, + 0.028347060084342957, + 0.0344783291220665, + 0.0014222452882677317, + 0.003216607728973031, + 0.0025680463295429945, + 0.002244367962703109, + 0.0028680702671408653, + 0.0023635709658265114, + 0.002384851686656475, + 0.0028196389321237803, + 0.0028076237067580223, + 0.0033060384448617697, + 0.0033213035203516483, + 0.002982436679303646, + 0.0032101834658533335, + 0.003938273526728153, + 0.002668873406946659, + 0.0027306813281029463, + 0.005046457517892122, + 0.004778665956109762, + 0.005697928834706545, + 0.004790520761162043, + 0.008373013697564602, + 0.005851146299391985, + 0.0036788685247302055, + 0.007654930464923382, + 0.0010156533680856228, + 0.000867788156028837, + 0.0008493285276927054, + 0.0006465761689469218, + 0.0010303426533937454, + 0.0011027322616428137, + 0.0009670931613072753, + 0.0009528325754217803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.03369671851396561 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic montage composed of a series of rapid, staggered extreme close-ups, the camera focused with a shallow depth of field on the fronds of a dense fern after a rainstorm. In the first shot, a delicate, pinnately divided fern frond dominates the frame, its surface coated in a thin film of water that makes the leaf's emerald green appear unusually saturated and deep. A crystal-clear, nearly perfectly spherical drop of water hangs delicately on the frond's tip, quivering slightly due to surface tension, its interior reflecting a miniature, distorted image of the veins beneath. A barely perceptible breeze gently tugs at the frond, swaying the droplet and causing light to play across its surface. The camera then cuts to a close-up of another fern frond, a string of even smaller droplets arranged along its central axis like a diamond necklace. Soft, diffused skylight evenly illuminates the scene, eliminating any distinct shadows and highlighting the clarity of each drop and the wet sheen of the frond. The entire sequence jumps rapidly between the microscopic details of different leaves, creating a rhythm that is both tranquil and full of vitality, conveying the fresh, clean and revitalized tranquil atmosphere of the forest floor after the rain.", + "seed": 20368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_5/index_5682_seed_20368_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": -0.01916649176273495, + "vlm_logprob": -0.01916649176273495, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.01916649176273495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.03096075728535652 + ], + "opposite_logprob_tokens": [ + -3.78096079826355 + ], + "logprob_reward": -0.03096075728535652, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03096075728535652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036117538809776306, + 0.15280860662460327, + 0.15777912735939026, + 0.0388229563832283, + 0.10721734911203384, + 0.08185359835624695, + 0.0737578421831131, + 0.18241213262081146, + 0.0708981305360794, + 0.09833276271820068 + ], + "credit_map": [ + 0.0037426133640110493, + 0.0030483584851026535, + 0.004060457926243544, + 0.004006906412541866, + 0.008542290888726711, + 0.003662953618913889, + 0.006177432835102081, + 0.002876524580642581, + 0.012923545204102993, + 0.020326612517237663, + 0.011076072230935097, + 0.01594359800219536, + 0.020188195630908012, + 0.020742686465382576, + 0.02457503229379654, + 0.027032865211367607, + 0.03783616051077843, + 0.009427336044609547, + 0.014273238368332386, + 0.01750064268708229, + 0.015534627251327038, + 0.014172465540468693, + 0.008894957602024078, + 0.04013969749212265, + 0.008306608535349369, + 0.00427453126758337, + 0.004813039675354958, + 0.0038672303780913353, + 0.004054514691233635, + 0.004891033284366131, + 0.004277706611901522, + 0.004338291473686695, + 0.009586316533386707, + 0.013634983450174332, + 0.01941819116473198, + 0.014149009250104427, + 0.014450845308601856, + 0.011669693514704704, + 0.007796254474669695, + 0.016512064263224602, + 0.008365283720195293, + 0.009542251005768776, + 0.01934022456407547, + 0.010342779569327831, + 0.010172986425459385, + 0.0077398899011313915, + 0.009929954074323177, + 0.0064202263019979, + 0.008597061038017273, + 0.009007376618683338, + 0.011801483109593391, + 0.011297649703919888, + 0.007198311388492584, + 0.009553857147693634, + 0.00963678490370512, + 0.006665320135653019, + 0.01401564758270979, + 0.008570601232349873, + 0.0543789267539978, + 0.020129920914769173, + 0.010114414617419243, + 0.05065375566482544, + 0.013884557411074638, + 0.010664289817214012, + 0.007426383439451456, + 0.01243016216903925, + 0.0071546523831784725, + 0.0073615615256130695, + 0.008408554829657078, + 0.01014059130102396, + 0.011222069151699543, + 0.006754154339432716, + 0.013156019151210785, + 0.014757908880710602, + 0.013596714474260807, + 0.011456235311925411, + 0.008529028855264187, + 0.010024036280810833, + 0.013294568285346031, + 0.013518244959414005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.03096075728535652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.0019452712731435895 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019452712731435895, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019452712731435895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1369272619485855, + 0.18050315976142883, + 0.16739995777606964, + 0.09578311443328857, + 0.05433996021747589, + 0.05278686806559563, + 0.05590010806918144, + 0.11048569530248642, + 0.10472951829433441, + 0.04114440083503723 + ], + "credit_map": [ + 0.01290920004248619, + 0.01256259623914957, + 0.015214294195175171, + 0.01773405820131302, + 0.033192142844200134, + 0.014580383896827698, + 0.01888415962457657, + 0.011850416660308838, + 0.02001490630209446, + 0.03417695686221123, + 0.015552581287920475, + 0.024557003751397133, + 0.018508870154619217, + 0.019039124250411987, + 0.021753234788775444, + 0.026900477707386017, + 0.0254659540951252, + 0.0136679382994771, + 0.02177267149090767, + 0.025688275694847107, + 0.02471040189266205, + 0.017235280945897102, + 0.01743021048605442, + 0.02142920158803463, + 0.006831256672739983, + 0.009493019431829453, + 0.012857306748628616, + 0.01182079128921032, + 0.011002413928508759, + 0.018864737823605537, + 0.01177254505455494, + 0.01314103789627552, + 0.004345308989286423, + 0.00546348886564374, + 0.010549616068601608, + 0.005887852981686592, + 0.008583483286201954, + 0.00679050013422966, + 0.004716804251074791, + 0.008002901449799538, + 0.0064496612176299095, + 0.005992353893816471, + 0.0082714157178998, + 0.00643574446439743, + 0.005262553226202726, + 0.00994005799293518, + 0.00583303440362215, + 0.004602042492479086, + 0.006015211343765259, + 0.0066238283179700375, + 0.006697851698845625, + 0.006969941779971123, + 0.005792081821709871, + 0.0069266981445252895, + 0.007033874746412039, + 0.009840614162385464, + 0.010779681615531445, + 0.00958345364779234, + 0.019520850852131844, + 0.018144987523555756, + 0.011241728439927101, + 0.012010585516691208, + 0.017862897366285324, + 0.01134149543941021, + 0.010725601576268673, + 0.01553299929946661, + 0.011034727096557617, + 0.012698918581008911, + 0.012293905951082706, + 0.012401745654642582, + 0.01686382107436657, + 0.013177785091102123, + 0.00508140167221427, + 0.003463290398940444, + 0.004794011358171701, + 0.004300953354686499, + 0.0047761197201907635, + 0.007114197593182325, + 0.007312810514122248, + 0.004301611799746752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0019452712731435895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.02059059962630272 + ], + "opposite_logprob_tokens": [ + -3.8955905437469482 + ], + "logprob_reward": -0.02059059962630272, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059059962630272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2678582966327667, + 0.05815453082323074, + 0.2676154673099518, + 0.05522725358605385, + 0.04425111413002014, + 0.03956246003508568, + 0.06801405549049377, + 0.06310258060693741, + 0.06164575740695, + 0.0745684802532196 + ], + "credit_map": [ + 0.017806459218263626, + 0.02107815258204937, + 0.02722986228764057, + 0.02732856199145317, + 0.0890272855758667, + 0.030355921015143394, + 0.04149323329329491, + 0.013538854196667671, + 0.006933261174708605, + 0.003828988876193762, + 0.006155018229037523, + 0.00911551434546709, + 0.00822125282138586, + 0.008598286658525467, + 0.007834495045244694, + 0.007467714603990316, + 0.06785424053668976, + 0.016337832435965538, + 0.025785742327570915, + 0.020554527640342712, + 0.020541725680232048, + 0.025358816608786583, + 0.018730424344539642, + 0.07245215773582458, + 0.0048026107251644135, + 0.004360218066722155, + 0.009847280569374561, + 0.008022296242415905, + 0.006891581695526838, + 0.007196225691586733, + 0.005408134777098894, + 0.008698910474777222, + 0.003551255678758025, + 0.005201724823564291, + 0.010235474444925785, + 0.006554737687110901, + 0.004363874904811382, + 0.0037917534355074167, + 0.004317309241741896, + 0.006234985310584307, + 0.004259283654391766, + 0.004448507912456989, + 0.007202433422207832, + 0.004330669064074755, + 0.004308745730668306, + 0.008097682148218155, + 0.0033583780750632286, + 0.0035567618906497955, + 0.008658668957650661, + 0.006727864965796471, + 0.010542902164161205, + 0.010178197175264359, + 0.006710399407893419, + 0.009515110403299332, + 0.00968234334141016, + 0.005998569540679455, + 0.0062979175709187984, + 0.005154123529791832, + 0.013513613492250443, + 0.009938781149685383, + 0.004516264423727989, + 0.011053917929530144, + 0.0073724049143493176, + 0.005255556665360928, + 0.005629431456327438, + 0.009690139442682266, + 0.006123917177319527, + 0.006755013018846512, + 0.007172753103077412, + 0.008959008380770683, + 0.011575797572731972, + 0.0057396967895329, + 0.0077235340140759945, + 0.006857077591121197, + 0.008932780474424362, + 0.008189857006072998, + 0.006987270433455706, + 0.015239625237882137, + 0.013140791095793247, + 0.007497547660022974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02059059962630272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.008639458566904068 + ], + "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": [ + 0.3651077151298523, + 0.22151722013950348, + 0.11956413835287094, + 0.04879573732614517, + 0.028982700780034065, + 0.027926422655582428, + 0.02210722677409649, + 0.05697735399007797, + 0.04580843448638916, + 0.063213050365448 + ], + "credit_map": [ + 0.028221722692251205, + 0.027648190036416054, + 0.047691430896520615, + 0.0367337241768837, + 0.10848264396190643, + 0.04127422347664833, + 0.055125340819358826, + 0.019930439069867134, + 0.026827454566955566, + 0.052572231739759445, + 0.016200285404920578, + 0.021215541288256645, + 0.021032564342021942, + 0.02439192309975624, + 0.014773625880479813, + 0.0445035845041275, + 0.03039868362247944, + 0.007277682889252901, + 0.011864031665027142, + 0.00859750621020794, + 0.009221824817359447, + 0.011677899397909641, + 0.007990718819200993, + 0.03253578022122383, + 0.005959469359368086, + 0.005767907947301865, + 0.0058023580349981785, + 0.0065923333168029785, + 0.004193620756268501, + 0.007941198535263538, + 0.007443771231919527, + 0.005095074884593487, + 0.0027844400610774755, + 0.0035913463216274977, + 0.003845711937174201, + 0.00473389495164156, + 0.00449996767565608, + 0.003179432824254036, + 0.003409924218431115, + 0.002937980927526951, + 0.0028277833480387926, + 0.003300501499325037, + 0.0044550783932209015, + 0.0028825909830629826, + 0.00367399537935853, + 0.003181067295372486, + 0.004008903633803129, + 0.0035964983981102705, + 0.002321753418073058, + 0.004888602998107672, + 0.002625736640766263, + 0.002508277539163828, + 0.0031063135247677565, + 0.0019964221864938736, + 0.0023009369615465403, + 0.0023591823410242796, + 0.004044139292091131, + 0.0034655481576919556, + 0.015452086925506592, + 0.006089923903346062, + 0.004337047692388296, + 0.014259670861065388, + 0.0053593371994793415, + 0.003969598561525345, + 0.004306836519390345, + 0.00834961049258709, + 0.0047022136859595776, + 0.006318700034171343, + 0.005495859310030937, + 0.004814081825315952, + 0.007247450761497021, + 0.00457367766648531, + 0.007353377994149923, + 0.009100096300244331, + 0.008223040960729122, + 0.007680484559386969, + 0.008323636837303638, + 0.008459754288196564, + 0.009613037109375, + 0.004459619056433439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008639458566904068 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.03369637206196785 + ], + "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": [ + 0.142700657248497, + 0.07207176834344864, + 0.2548767924308777, + 0.05346919223666191, + 0.06358989328145981, + 0.15592528879642487, + 0.05539960041642189, + 0.0446968711912632, + 0.049583327025175095, + 0.10768668353557587 + ], + "credit_map": [ + 0.01632942631840706, + 0.01090871263295412, + 0.019548408687114716, + 0.02039892040193081, + 0.030653830617666245, + 0.01540812011808157, + 0.019091159105300903, + 0.010362060740590096, + 0.00708054518327117, + 0.006256517954170704, + 0.0063999127596616745, + 0.009547040797770023, + 0.012608266435563564, + 0.008253654465079308, + 0.015073828399181366, + 0.006851998623460531, + 0.0629539042711258, + 0.016303468495607376, + 0.02728554792702198, + 0.018122704699635506, + 0.022593660280108452, + 0.021595977246761322, + 0.016491813585162163, + 0.06952967494726181, + 0.00856983195990324, + 0.0058504738844931126, + 0.005019145552068949, + 0.004398208111524582, + 0.005491656716912985, + 0.009451346471905708, + 0.005779418163001537, + 0.008909105323255062, + 0.007313092239201069, + 0.007141693029552698, + 0.008426299318671227, + 0.0121906828135252, + 0.008956831879913807, + 0.005016499198973179, + 0.007926619611680508, + 0.006618169602006674, + 0.015248515643179417, + 0.021140798926353455, + 0.02172226272523403, + 0.019959082826972008, + 0.019976142793893814, + 0.023757148534059525, + 0.017610086128115654, + 0.016511231660842896, + 0.006130265537649393, + 0.0067911469377577305, + 0.005985155235975981, + 0.007154188118875027, + 0.0060503920540213585, + 0.008820172399282455, + 0.009062196128070354, + 0.005406079348176718, + 0.004974983166903257, + 0.0049767266027629375, + 0.006955835968255997, + 0.007357423659414053, + 0.004154532216489315, + 0.005541535094380379, + 0.00566548528149724, + 0.005070345476269722, + 0.005637929309159517, + 0.005747475195676088, + 0.005267335567623377, + 0.0069477190263569355, + 0.006097708363085985, + 0.008879991248250008, + 0.005132841411978006, + 0.005872318055480719, + 0.009504009038209915, + 0.013793903402984142, + 0.015048005618155003, + 0.01419801078736782, + 0.012130165472626686, + 0.0165821872651577, + 0.01766744814813137, + 0.008762949146330357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.03369637206196785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic montage composed of a series of rapid, staggered extreme close-ups, the camera focused with a shallow depth of field on the fronds of a dense fern after a rainstorm. In the first shot, a delicate, pinnately divided fern frond dominates the frame, its surface coated in a thin film of water that makes the leaf's emerald green appear unusually saturated and deep. A crystal-clear, nearly perfectly spherical drop of water hangs delicately on the frond's tip, quivering slightly due to surface tension, its interior reflecting a miniature, distorted image of the veins beneath. A barely perceptible breeze gently tugs at the frond, swaying the droplet and causing light to play across its surface. The camera then cuts to a close-up of another fern frond, a string of even smaller droplets arranged along its central axis like a diamond necklace. Soft, diffused skylight evenly illuminates the scene, eliminating any distinct shadows and highlighting the clarity of each drop and the wet sheen of the frond. The entire sequence jumps rapidly between the microscopic details of different leaves, creating a rhythm that is both tranquil and full of vitality, conveying the fresh, clean and revitalized tranquil atmosphere of the forest floor after the rain.", + "seed": 20369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_5/index_5682_seed_20369_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": -0.021298753400333225, + "vlm_logprob": -0.021298753400333225, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.021298753400333225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.040822528302669525 + ], + "opposite_logprob_tokens": [ + -3.5408225059509277 + ], + "logprob_reward": -0.040822528302669525, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.040822528302669525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10625606030225754, + 0.07235405594110489, + 0.16596613824367523, + 0.2844020426273346, + 0.05926186218857765, + 0.1244376078248024, + 0.05526309832930565, + 0.02164429798722267, + 0.04004860669374466, + 0.070366270840168 + ], + "credit_map": [ + 0.005365409888327122, + 0.007166588678956032, + 0.008699865080416203, + 0.010871674865484238, + 0.00962692592293024, + 0.016204984858632088, + 0.026198139414191246, + 0.022122470661997795, + 0.006031806115061045, + 0.005276479292660952, + 0.00625448115170002, + 0.010300879366695881, + 0.005071846768260002, + 0.008068474009633064, + 0.020105762407183647, + 0.01124432310461998, + 0.02728286013007164, + 0.01276999432593584, + 0.016007909551262856, + 0.020856129005551338, + 0.012650742195546627, + 0.03838124871253967, + 0.018239261582493782, + 0.019778000190854073, + 0.019970053806900978, + 0.027756493538618088, + 0.019900912418961525, + 0.05159720405936241, + 0.042353011667728424, + 0.04222485423088074, + 0.0539100207388401, + 0.02668951265513897, + 0.005993138533085585, + 0.006705089006572962, + 0.00937468558549881, + 0.007545868866145611, + 0.009684007614850998, + 0.006385189946740866, + 0.006343659479171038, + 0.007230221759527922, + 0.009278729557991028, + 0.01632261835038662, + 0.014245974831283092, + 0.011365935206413269, + 0.019668249413371086, + 0.014468341134488583, + 0.013808402232825756, + 0.02527935430407524, + 0.005986636504530907, + 0.006150966510176659, + 0.007915783673524857, + 0.006661661434918642, + 0.007745633367449045, + 0.005353001412004232, + 0.006848693825304508, + 0.008600725792348385, + 0.0029201055876910686, + 0.002557630417868495, + 0.002637193538248539, + 0.0015500751323997974, + 0.0022182678803801537, + 0.0023592656943947077, + 0.0030808367300778627, + 0.004320924170315266, + 0.005101691000163555, + 0.0087053207680583, + 0.003854227950796485, + 0.0038362927734851837, + 0.005069698207080364, + 0.00418899254873395, + 0.004882016684859991, + 0.004410366527736187, + 0.008091701194643974, + 0.005719117354601622, + 0.006686789449304342, + 0.0054330588318407536, + 0.005906311329454184, + 0.013652496971189976, + 0.017853805795311928, + 0.007022992707788944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.040822528302669525 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.0019445574143901467 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019445574143901467, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019445574143901467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8021109104156494, + 0.02383267693221569, + 0.027884379029273987, + 0.018853498622775078, + 0.016788911074399948, + 0.03240922465920448, + 0.015265967696905136, + 0.016024956479668617, + 0.03669692575931549, + 0.010132551193237305 + ], + "credit_map": [ + 0.02112349309027195, + 0.02056395821273327, + 0.019158022478222847, + 0.0927010178565979, + 0.02839680202305317, + 0.03368145599961281, + 0.3308180868625641, + 0.25566810369491577, + 0.0013304554158821702, + 0.0013620367972180247, + 0.001391334692016244, + 0.0030110913794487715, + 0.0011670527746900916, + 0.0019716620445251465, + 0.008669261820614338, + 0.004929786082357168, + 0.0035737527068704367, + 0.002019121777266264, + 0.0022155591286718845, + 0.005822444800287485, + 0.0028492400888353586, + 0.00408181082457304, + 0.003489434951916337, + 0.003833015915006399, + 0.00142251409124583, + 0.0018330637831240892, + 0.0015318410005420446, + 0.0030872533097863197, + 0.0024190694093704224, + 0.0030601692851632833, + 0.003358800197020173, + 0.0021407892927527428, + 0.0025571403093636036, + 0.0018325757700949907, + 0.002707219449803233, + 0.0018507923232391477, + 0.00245228735730052, + 0.0023471293970942497, + 0.0014287105295807123, + 0.001613057916983962, + 0.0026814851444214582, + 0.004734761547297239, + 0.0033172154799103737, + 0.002865659072995186, + 0.005159479100257158, + 0.003823184175416827, + 0.0037605019751936197, + 0.006066938862204552, + 0.0012612283462658525, + 0.0039914436638355255, + 0.0011442259419709444, + 0.0008509612525813282, + 0.00425702566280961, + 0.0010325188050046563, + 0.0011609389912337065, + 0.0015676250914111733, + 0.0017987781902775168, + 0.0013449372490867972, + 0.0019579101353883743, + 0.0013739026617258787, + 0.0015457578701898456, + 0.0017179576680064201, + 0.0032458556815981865, + 0.0030398585367947817, + 0.004376461263746023, + 0.009028671309351921, + 0.002835275139659643, + 0.002856170991435647, + 0.004880454856902361, + 0.0056623537093400955, + 0.003713911399245262, + 0.0033436305820941925, + 0.0010816038120537996, + 0.0008695794967934489, + 0.0009134573629125953, + 0.0011355993337929249, + 0.0008490158361382782, + 0.0024199984036386013, + 0.0018876898102462292, + 0.0009756085928529501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0019445574143901467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.006742938421666622 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742938421666622, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742938421666622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6807299852371216, + 0.18689897656440735, + 0.010269001126289368, + 0.013373661786317825, + 0.012985862791538239, + 0.025424502789974213, + 0.0188030693680048, + 0.016086183488368988, + 0.028446190059185028, + 0.006982504855841398 + ], + "credit_map": [ + 0.017558565363287926, + 0.017238425090909004, + 0.013834110461175442, + 0.06898221373558044, + 0.023921234533190727, + 0.027957186102867126, + 0.28830569982528687, + 0.2229325920343399, + 0.006571522913873196, + 0.004726339131593704, + 0.003697877051308751, + 0.023214465007185936, + 0.007093878462910652, + 0.009627217426896095, + 0.08389193564653397, + 0.048075757920742035, + 0.0016214605420827866, + 0.0007255241507664323, + 0.0014195512048900127, + 0.0015101720346137881, + 0.00132092391140759, + 0.0012416341342031956, + 0.0017265758942812681, + 0.0007031594286672771, + 0.0009781268890947104, + 0.0012715639313682914, + 0.0010541855590417981, + 0.0024563204497098923, + 0.0019706562161445618, + 0.0018854467198252678, + 0.0021648250985890627, + 0.001592538203112781, + 0.0018325087148696184, + 0.0013729323400184512, + 0.0018219236517325044, + 0.0011177612468600273, + 0.0015522571047767997, + 0.0024318911600857973, + 0.0012040063738822937, + 0.0016525835962966084, + 0.0020314897410571575, + 0.0034429728984832764, + 0.002438083291053772, + 0.002111011417582631, + 0.003905594814568758, + 0.002743394346907735, + 0.003078973153606057, + 0.0056729866191744804, + 0.0012633462902158499, + 0.006122308783233166, + 0.0011334032751619816, + 0.0006395107484422624, + 0.006352498661726713, + 0.0013306491309776902, + 0.00101543462369591, + 0.0009459196589887142, + 0.0023050240706652403, + 0.0016774036921560764, + 0.0019592747557908297, + 0.0013268321054056287, + 0.0014264152850955725, + 0.0018998769810423255, + 0.0031177597120404243, + 0.0023735989816486835, + 0.0038300296291708946, + 0.007690959144383669, + 0.0024035004898905754, + 0.0017076365184038877, + 0.0041716983541846275, + 0.00307407695800066, + 0.002951244357973337, + 0.0026170467026531696, + 0.0009740851819515228, + 0.00044911057921126485, + 0.0006032905075699091, + 0.0005039432435296476, + 0.00048289416008628905, + 0.0009768168674781919, + 0.0020258163567632437, + 0.0009665488614700735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.03369683399796486 + ], + "opposite_logprob_tokens": [ + -3.4086968898773193 + ], + "logprob_reward": -0.03369683399796486, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369683399796486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8804359436035156, + 0.005910966079682112, + 0.0867481380701065, + 0.004877233877778053, + 0.004290865734219551, + 0.004202635958790779, + 0.0039042753633111715, + 0.004143937956541777, + 0.002419649390503764, + 0.0030663423240184784 + ], + "credit_map": [ + 0.023077163845300674, + 0.023018300533294678, + 0.021099969744682312, + 0.10161597281694412, + 0.03146086633205414, + 0.03720723092556, + 0.3642466366291046, + 0.2787098288536072, + 0.0006658426718786359, + 0.0005600759177468717, + 0.00043620436917990446, + 0.001343102427199483, + 0.0004594538186211139, + 0.0006565987132489681, + 0.0011877960059791803, + 0.000601892068516463, + 0.009128200821578503, + 0.004863721318542957, + 0.006339657120406628, + 0.010900773108005524, + 0.008278848603367805, + 0.029423672705888748, + 0.00832523312419653, + 0.009488037787377834, + 0.00028925223159603775, + 0.000385297869797796, + 0.00057465914869681, + 0.0009119676542468369, + 0.000557170482352376, + 0.0006259938236325979, + 0.0009577646269463003, + 0.0005751284188590944, + 0.00030157185392454267, + 0.0003600399650167674, + 0.0004599417152348906, + 0.0008174621616490185, + 0.00037352635990828276, + 0.0010758597636595368, + 0.00040429478394798934, + 0.0004981692181900144, + 0.0003570925327949226, + 0.0005104253650642931, + 0.000556030950974673, + 0.0005692692357115448, + 0.0006020586588419974, + 0.00044335174607113004, + 0.0006814851658418775, + 0.00048292253632098436, + 0.00026325174258090556, + 0.0010011913254857063, + 0.0003164942900184542, + 0.0002466757141519338, + 0.0010034716688096523, + 0.00023089129535946995, + 0.00037281165714375675, + 0.0004694878589361906, + 0.0005193655961193144, + 0.0004472786677069962, + 0.0005282795173116028, + 0.00033181984326802194, + 0.0004153327608946711, + 0.0004463593941181898, + 0.0008421916281804442, + 0.0006133103161118925, + 0.00038097487413324416, + 0.0006222493248060346, + 0.00019249039178248495, + 0.0001472957810619846, + 0.00027022839640267193, + 0.0002150173531845212, + 0.0002621236490085721, + 0.0003292697365395725, + 0.00028717576060444117, + 0.0003172026190441102, + 0.00036184219061397016, + 0.0003325947909615934, + 0.000327820802340284, + 0.0005208498914726079, + 0.0005272767157293856, + 0.00039157981518656015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03369683399796486 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the shots presented as a rapid montage of extreme close-ups?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject matter focused on the fronds of a dense fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern surface coated in water with saturated emerald green coloration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a spherical water drop visible hanging from a frond tip, quivering slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffused, eliminating distinct 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": [ + -0.023286908864974976 + ], + "opposite_logprob_tokens": [ + -3.773286819458008 + ], + "logprob_reward": -0.023286908864974976, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286908864974976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6479970216751099, + 0.1490178257226944, + 0.05993881821632385, + 0.017538299784064293, + 0.016718504950404167, + 0.01896671950817108, + 0.04590969532728195, + 0.018109234049916267, + 0.012395639903843403, + 0.013408329337835312 + ], + "credit_map": [ + 0.017639562487602234, + 0.017420342192053795, + 0.017529215663671494, + 0.08057292550802231, + 0.022881118580698967, + 0.02770427241921425, + 0.26368263363838196, + 0.2005668729543686, + 0.005627323407679796, + 0.004011611919850111, + 0.0036414412315934896, + 0.01813453808426857, + 0.00580385560169816, + 0.006934576667845249, + 0.06617601960897446, + 0.038688451051712036, + 0.009935293346643448, + 0.004144866950809956, + 0.005556977353990078, + 0.005883780308067799, + 0.005343134049326181, + 0.017572201788425446, + 0.005679470486938953, + 0.005823086481541395, + 0.001030373270623386, + 0.0012353913625702262, + 0.001867697574198246, + 0.0032391026616096497, + 0.001770333736203611, + 0.0031489990651607513, + 0.0030917800031602383, + 0.002154621295630932, + 0.0022243442945182323, + 0.0022374303080141544, + 0.002080319682136178, + 0.0017416727496311069, + 0.0028110442217439413, + 0.002418365329504013, + 0.001399042084813118, + 0.0018062854651361704, + 0.0011938209645450115, + 0.002341943560168147, + 0.001959388144314289, + 0.002199725480750203, + 0.0028722635470330715, + 0.002430009189993143, + 0.0032912089955061674, + 0.0026783582288771868, + 0.0032535779755562544, + 0.012422863394021988, + 0.0030750667210668325, + 0.0025784149765968323, + 0.014834643341600895, + 0.002977965632453561, + 0.0031699221581220627, + 0.0035972404293715954, + 0.0031114111188799143, + 0.0018074638210237026, + 0.0019113554153591394, + 0.0013662519631907344, + 0.0021793947089463472, + 0.0017114408547058702, + 0.002805009251460433, + 0.003216906450688839, + 0.0017227573553100228, + 0.0019094031304121017, + 0.0011667913058772683, + 0.0008281809277832508, + 0.002274206606671214, + 0.0014081767294555902, + 0.0016066166572272778, + 0.001479505910538137, + 0.0019475208828225732, + 0.001078694243915379, + 0.001493764342740178, + 0.001019662362523377, + 0.001108198193833232, + 0.0014322735369205475, + 0.00302052590996027, + 0.002307688584551215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023286908864974976 + } + ], + "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/0000025/group_6/index_tensor([7585, 7585])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_6/index_tensor([7585, 7585])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e0267b974882549aabd27b24a8d790eb70c1e07 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_6/index_tensor([7585, 7585])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing medium close-up focuses on a young soldier, leaning wearily against the damp earthen wall of a muddy trench, against a gloomy sky shrouded in leaden clouds. He appears to be in his early twenties, his face smeared with dirt and sweat, dark circles under his eyes betraying a look of profound weariness. His short, rain-soaked, dark hair is matted to his forehead. He wears a worn, dark green military jacket, the collar turned up to protect against the chill. With slightly trembling but determined hands, he pulls a damp matchbox from his jacket pocket, carefully extracts a match, and vigorously strikes it against the rough side. A small, bright orange flame bursts forth, briefly illuminating his tense jaw and sunken eye sockets, and also revealing the clammy texture of the surrounding rough earth and rotting wooden support structures. He quickly brought the precious flame close to the wrinkled, seemingly last cigarette in his mouth. The cigarette butt hissed faintly at the touch of the flame, igniting into a single crimson point that stood out against the dim light. He took a deep breath, then slowly exhaled a cloud of white smoke that condensed in the cold air. His furrowed brow seemed to relax for a moment. The entire scene was enveloped in a repressive, cold hue. Only that speck of crimson and the momentary emptiness in the soldier's eyes created a complex atmosphere of both exhaustion and tranquility in the intervals between battles.", + "seed": 7156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_6/index_7585_seed_7156_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": -0.0176500310190022, + "vlm_logprob": -0.0176500310190022, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.0176500310190022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019582398235797882 + ], + "opposite_logprob_tokens": [ + -4.644582271575928 + ], + "logprob_reward": -0.019582398235797882, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.019582398235797882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16266340017318726, + 0.14114068448543549, + 0.07423575222492218, + 0.0853869840502739, + 0.07570837438106537, + 0.05196571722626686, + 0.08712533116340637, + 0.13031083345413208, + 0.13230940699577332, + 0.059153489768505096 + ], + "credit_map": [ + 0.01858333684504032, + 0.024640299379825592, + 0.021563058719038963, + 0.020660951733589172, + 0.011680103838443756, + 0.008381826803088188, + 0.03286898508667946, + 0.02428482472896576, + 0.00992563460022211, + 0.011116951704025269, + 0.02346639335155487, + 0.02710118517279625, + 0.012661186046898365, + 0.01031283289194107, + 0.02118057571351528, + 0.025375939905643463, + 0.00383470649830997, + 0.005026828497648239, + 0.018493736162781715, + 0.013234935700893402, + 0.0059747640043497086, + 0.005362323019653559, + 0.010730111971497536, + 0.011578351259231567, + 0.006741143763065338, + 0.005136288236826658, + 0.01580454595386982, + 0.0208448376506567, + 0.006364507135003805, + 0.0051619610749185085, + 0.011977588757872581, + 0.013356116600334644, + 0.005759579129517078, + 0.005071235354989767, + 0.01110806129872799, + 0.01483509223908186, + 0.00890408642590046, + 0.004418604541569948, + 0.014456174336373806, + 0.011155540123581886, + 0.005047392100095749, + 0.004618857987225056, + 0.00843601394444704, + 0.007409689016640186, + 0.004080560058355331, + 0.0037576600443571806, + 0.008947644382715225, + 0.009667899459600449, + 0.004352466203272343, + 0.004866085015237331, + 0.01956598088145256, + 0.013213693164288998, + 0.008994261734187603, + 0.006312813144177198, + 0.016072802245616913, + 0.013747229240834713, + 0.00793851725757122, + 0.009192107245326042, + 0.018908057361841202, + 0.026877041906118393, + 0.015214220620691776, + 0.006257561966776848, + 0.024989182129502296, + 0.020934151485562325, + 0.010052159428596497, + 0.013044623658061028, + 0.021243751049041748, + 0.025237681344151497, + 0.013908660039305687, + 0.01373683288693428, + 0.015526965260505676, + 0.019558727741241455, + 0.004942058585584164, + 0.0049678850919008255, + 0.0078070275485515594, + 0.012589250691235065, + 0.006572545040398836, + 0.00440005399286747, + 0.009655906818807125, + 0.008218765258789062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.019582398235797882 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025132279843091965 + ], + "opposite_logprob_tokens": [ + -6.002513408660889 + ], + "logprob_reward": -0.0025132279843091965, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025132279843091965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1912447065114975, + 0.046711403876543045, + 0.051688581705093384, + 0.07709656655788422, + 0.10789567232131958, + 0.08522068709135056, + 0.09089913219213486, + 0.06526795029640198, + 0.12467062473297119, + 0.159304678440094 + ], + "credit_map": [ + 0.01281263679265976, + 0.014462778344750404, + 0.01923331804573536, + 0.028119802474975586, + 0.027307109907269478, + 0.014677982777357101, + 0.042448434978723526, + 0.03218265622854233, + 0.0028317009564489126, + 0.0035649538040161133, + 0.008236292749643326, + 0.011534028686583042, + 0.0030443668365478516, + 0.0018123554764315486, + 0.007677927613258362, + 0.008009776473045349, + 0.0023661425802856684, + 0.0024837420787662268, + 0.011350134387612343, + 0.010128282010555267, + 0.004792856518179178, + 0.005251860711723566, + 0.007461852859705687, + 0.007853711023926735, + 0.004129624925553799, + 0.003804814303293824, + 0.016079803928732872, + 0.021376533433794975, + 0.005357230082154274, + 0.004067941103130579, + 0.00943392887711525, + 0.012846690602600574, + 0.005429020617157221, + 0.006415426731109619, + 0.02307649329304695, + 0.027147306129336357, + 0.007871970534324646, + 0.005226852837949991, + 0.016214771196246147, + 0.016513828188180923, + 0.006572650279849768, + 0.00845965463668108, + 0.014766883105039597, + 0.01449031662195921, + 0.007045180536806583, + 0.00506878225132823, + 0.011924797669053078, + 0.01689242385327816, + 0.006775974296033382, + 0.009082361124455929, + 0.02069246955215931, + 0.010930498130619526, + 0.007970674894750118, + 0.005413154140114784, + 0.017768751829862595, + 0.012265246361494064, + 0.004581010434776545, + 0.004654861055314541, + 0.009172276593744755, + 0.012531235814094543, + 0.008819987066090107, + 0.003901975229382515, + 0.011803668923676014, + 0.009802935644984245, + 0.009232212789356709, + 0.011608662083745003, + 0.016588641330599785, + 0.024249812588095665, + 0.0132400868460536, + 0.009240812622010708, + 0.021086761727929115, + 0.019423633813858032, + 0.0143797118216753, + 0.019499583169817924, + 0.020900219678878784, + 0.03145436942577362, + 0.017212357372045517, + 0.01142826210707426, + 0.021399859338998795, + 0.023030312731862068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0025132279843091965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963034927845001 + ], + "opposite_logprob_tokens": [ + -5.13096284866333 + ], + "logprob_reward": -0.005963034927845001, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963034927845001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11079134792089462, + 0.09653611481189728, + 0.08489767462015152, + 0.07614816725254059, + 0.20750126242637634, + 0.06840912252664566, + 0.11339415609836578, + 0.08357477188110352, + 0.14192695915699005, + 0.016820505261421204 + ], + "credit_map": [ + 0.007762860041111708, + 0.008785362355411053, + 0.014910269528627396, + 0.02436203509569168, + 0.008086786605417728, + 0.008607692085206509, + 0.016288015991449356, + 0.021988311782479286, + 0.005928871687501669, + 0.00567084364593029, + 0.012483829632401466, + 0.022109152749180794, + 0.009364972822368145, + 0.007317238952964544, + 0.016145119443535805, + 0.017516080290079117, + 0.004611469805240631, + 0.003986538853496313, + 0.014737198129296303, + 0.023053541779518127, + 0.006454217713326216, + 0.005193090066313744, + 0.01301378384232521, + 0.013847826980054379, + 0.0071972431614995, + 0.004557022359222174, + 0.01201956532895565, + 0.012506182305514812, + 0.010085070505738258, + 0.005370734259486198, + 0.012596816755831242, + 0.011815527454018593, + 0.010022839531302452, + 0.012738353572785854, + 0.04321649298071861, + 0.049516819417476654, + 0.012198356911540031, + 0.010012059472501278, + 0.03858339414000511, + 0.031212903559207916, + 0.0044165197759866714, + 0.005140520632266998, + 0.011249067261815071, + 0.013758189976215363, + 0.005921538919210434, + 0.003837020369246602, + 0.013989630155265331, + 0.01009663101285696, + 0.008030141703784466, + 0.01091811154037714, + 0.021107980981469154, + 0.013761361129581928, + 0.00944527331739664, + 0.004835956264287233, + 0.028701765462756157, + 0.016593554988503456, + 0.005216992925852537, + 0.005678758956491947, + 0.014011671766638756, + 0.011405511759221554, + 0.008391686715185642, + 0.004707165062427521, + 0.020241854712367058, + 0.01392111461609602, + 0.010872839950025082, + 0.013107364065945148, + 0.02219543047249317, + 0.028455544263124466, + 0.013670148327946663, + 0.013077735900878906, + 0.018666628748178482, + 0.02188127115368843, + 0.0017626321641728282, + 0.002472413470968604, + 0.002758729737251997, + 0.0027817452792078257, + 0.0014279534807428718, + 0.001215089694596827, + 0.002138025127351284, + 0.0022639150265604258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.005963034927845001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052535575814545155 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.0052535575814545155, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052535575814545155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2372855544090271, + 0.08851943910121918, + 0.056337542831897736, + 0.1514538675546646, + 0.1486033946275711, + 0.08423195034265518, + 0.0699789896607399, + 0.04162595793604851, + 0.06961096078157425, + 0.052352264523506165 + ], + "credit_map": [ + 0.020666565746068954, + 0.02747872844338417, + 0.03403260186314583, + 0.04443323239684105, + 0.018628031015396118, + 0.010379758663475513, + 0.037278447300195694, + 0.04438820853829384, + 0.009521200321614742, + 0.010250437073409557, + 0.015271113254129887, + 0.014085730537772179, + 0.006956827826797962, + 0.006990250200033188, + 0.011127660982310772, + 0.014316222630441189, + 0.004772865679115057, + 0.004441542085260153, + 0.011237656697630882, + 0.011336635798215866, + 0.004541717004030943, + 0.0036624325439333916, + 0.007301448378711939, + 0.00904324185103178, + 0.01214019488543272, + 0.01028391718864441, + 0.029939979314804077, + 0.039480339735746384, + 0.009846506640315056, + 0.008363238535821438, + 0.01901058852672577, + 0.02238912135362625, + 0.010731585323810577, + 0.012503157369792461, + 0.024220360442996025, + 0.030651627108454704, + 0.01617305539548397, + 0.00839900877326727, + 0.02380497381091118, + 0.022119617089629173, + 0.005870360415428877, + 0.007548211142420769, + 0.01234227605164051, + 0.014861172065138817, + 0.007354908622801304, + 0.005077387671917677, + 0.014578952454030514, + 0.016598686575889587, + 0.007979544810950756, + 0.005764963570982218, + 0.010964705608785152, + 0.00892614759504795, + 0.006042766850441694, + 0.0048794313333928585, + 0.01490097213536501, + 0.01052046101540327, + 0.0047013163566589355, + 0.006827026139944792, + 0.007382451556622982, + 0.005084814503788948, + 0.003715441096574068, + 0.0027546698693186045, + 0.005916242953389883, + 0.0052439970895648, + 0.005244077183306217, + 0.006682653911411762, + 0.012876728549599648, + 0.012576249428093433, + 0.007474740967154503, + 0.007577360142022371, + 0.006899953354150057, + 0.010279196314513683, + 0.005533337127417326, + 0.009078266099095345, + 0.007358055096119642, + 0.009167569689452648, + 0.0039491127245128155, + 0.003752517281100154, + 0.0065119946375489235, + 0.0070014153607189655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0052535575814545155 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549379363656044 + ], + "opposite_logprob_tokens": [ + -2.9299378395080566 + ], + "logprob_reward": -0.0549379363656044, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549379363656044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07384487986564636, + 0.05432792752981186, + 0.03932885825634003, + 0.20782016217708588, + 0.13723666965961456, + 0.12914013862609863, + 0.11270776391029358, + 0.10320806503295898, + 0.10724437236785889, + 0.035141121596097946 + ], + "credit_map": [ + 0.005644082557410002, + 0.006726742256432772, + 0.0076138172298669815, + 0.01134577114135027, + 0.01097065769135952, + 0.006418447010219097, + 0.01418768148869276, + 0.010937687940895557, + 0.0037267212755978107, + 0.005326193757355213, + 0.007321061100810766, + 0.008692892268300056, + 0.004050201736390591, + 0.003919203765690327, + 0.012029178440570831, + 0.009262478910386562, + 0.0028311240021139383, + 0.004044354893267155, + 0.0070985304191708565, + 0.008260155096650124, + 0.0033910456113517284, + 0.002581377048045397, + 0.005609966814517975, + 0.005512305069714785, + 0.013222499750554562, + 0.010404611937701702, + 0.036605626344680786, + 0.0640902891755104, + 0.015206847339868546, + 0.009344841353595257, + 0.02629067562520504, + 0.03265475854277611, + 0.00609360309317708, + 0.00840196292847395, + 0.02814589999616146, + 0.03295901045203209, + 0.011304584331810474, + 0.006109607871621847, + 0.025649357587099075, + 0.018572630360722542, + 0.008741837926208973, + 0.010411208495497704, + 0.019261619076132774, + 0.02480434998869896, + 0.011758483946323395, + 0.007426130585372448, + 0.02003457210958004, + 0.02670193836092949, + 0.008721498772501945, + 0.008106905966997147, + 0.013021457940340042, + 0.016733987256884575, + 0.015830589458346367, + 0.007311870343983173, + 0.025457164272665977, + 0.017524292692542076, + 0.005933642853051424, + 0.008055846206843853, + 0.019857456907629967, + 0.0169106125831604, + 0.009032279253005981, + 0.00651986338198185, + 0.021592849865555763, + 0.015305510722100735, + 0.008900770917534828, + 0.008229520171880722, + 0.014069256372749805, + 0.02206604741513729, + 0.010192809626460075, + 0.008729522116482258, + 0.017404915764927864, + 0.017651528120040894, + 0.0031468451488763094, + 0.0042638108134269714, + 0.005494553595781326, + 0.006944615859538317, + 0.003132858779281378, + 0.0021701576188206673, + 0.0052144755609333515, + 0.004773803986608982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.0549379363656044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium close-up focuses on a young soldier, leaning wearily against the damp earthen wall of a muddy trench, against a gloomy sky shrouded in leaden clouds. He appears to be in his early twenties, his face smeared with dirt and sweat, dark circles under his eyes betraying a look of profound weariness. His short, rain-soaked, dark hair is matted to his forehead. He wears a worn, dark green military jacket, the collar turned up to protect against the chill. With slightly trembling but determined hands, he pulls a damp matchbox from his jacket pocket, carefully extracts a match, and vigorously strikes it against the rough side. A small, bright orange flame bursts forth, briefly illuminating his tense jaw and sunken eye sockets, and also revealing the clammy texture of the surrounding rough earth and rotting wooden support structures. He quickly brought the precious flame close to the wrinkled, seemingly last cigarette in his mouth. The cigarette butt hissed faintly at the touch of the flame, igniting into a single crimson point that stood out against the dim light. He took a deep breath, then slowly exhaled a cloud of white smoke that condensed in the cold air. His furrowed brow seemed to relax for a moment. The entire scene was enveloped in a repressive, cold hue. Only that speck of crimson and the momentary emptiness in the soldier's eyes created a complex atmosphere of both exhaustion and tranquility in the intervals between battles.", + "seed": 7157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_6/index_7585_seed_7157_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": -0.059540279861539605, + "vlm_logprob": -0.059540279861539605, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.059540279861539605, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025841688737273216 + ], + "opposite_logprob_tokens": [ + -4.15084171295166 + ], + "logprob_reward": -0.025841688737273216, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.025841688737273216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061226025223731995, + 0.07544906437397003, + 0.033285558223724365, + 0.06060970574617386, + 0.024193961173295975, + 0.07233616709709167, + 0.29323476552963257, + 0.05032157897949219, + 0.22807657718658447, + 0.10126660019159317 + ], + "credit_map": [ + 0.00397905707359314, + 0.004984616301953793, + 0.009694356471300125, + 0.010179898701608181, + 0.008144732564687729, + 0.003787468886002898, + 0.011087102815508842, + 0.009368793107569218, + 0.00391333969309926, + 0.0044776033610105515, + 0.012565533630549908, + 0.012106701731681824, + 0.013827245682477951, + 0.004123271442949772, + 0.014737102203071117, + 0.009698263369500637, + 0.002031740266829729, + 0.0030730499420315027, + 0.004778092727065086, + 0.0044736010022461414, + 0.0045046117156744, + 0.002854573540389538, + 0.005845976062119007, + 0.005723914131522179, + 0.004139095079153776, + 0.004165665712207556, + 0.011251341551542282, + 0.005852066446095705, + 0.00489668408408761, + 0.0031399012077599764, + 0.012058238498866558, + 0.015106713399291039, + 0.00176151841878891, + 0.0019461599877104163, + 0.0049272701144218445, + 0.0036180736497044563, + 0.0018992897821590304, + 0.0023909315932542086, + 0.003191579831764102, + 0.0044591366313397884, + 0.004435218404978514, + 0.002641400322318077, + 0.006768664810806513, + 0.0034227133728563786, + 0.005211416631937027, + 0.003325361991301179, + 0.019730057567358017, + 0.026801330968737602, + 0.03814856335520744, + 0.014795196242630482, + 0.02658180147409439, + 0.013455270789563656, + 0.009733160026371479, + 0.010633876547217369, + 0.06368240714073181, + 0.11620450019836426, + 0.007952374406158924, + 0.0037240770179778337, + 0.00502267898991704, + 0.002344956388697028, + 0.0026669695507735014, + 0.0030317201744765043, + 0.00822488684207201, + 0.01735391654074192, + 0.03085656836628914, + 0.04493100196123123, + 0.019481467083096504, + 0.009106799028813839, + 0.014658614993095398, + 0.007562344428151846, + 0.05820683389902115, + 0.043272946029901505, + 0.022491101175546646, + 0.022674856707453728, + 0.009396311827003956, + 0.006602643057703972, + 0.007002570666372776, + 0.003311637556180358, + 0.01717928610742092, + 0.012608194723725319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.025841688737273216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041167521849274635 + ], + "opposite_logprob_tokens": [ + -5.504116535186768 + ], + "logprob_reward": -0.0041167521849274635, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041167521849274635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07128407061100006, + 0.048419926315546036, + 0.035292334854602814, + 0.056249529123306274, + 0.04739876091480255, + 0.08626504242420197, + 0.17363840341567993, + 0.2400972545146942, + 0.06962692737579346, + 0.17172771692276 + ], + "credit_map": [ + 0.006237114313989878, + 0.007585208397358656, + 0.008156560361385345, + 0.00813966803252697, + 0.012549979612231255, + 0.006063432898372412, + 0.012771053239703178, + 0.009781058877706528, + 0.002706803847104311, + 0.0030794055201113224, + 0.007749546784907579, + 0.008246316574513912, + 0.007323273923248053, + 0.0031466330401599407, + 0.008948097936809063, + 0.007219846360385418, + 0.0027329775039106607, + 0.003358460497111082, + 0.004955918062478304, + 0.005079125985503197, + 0.0035444763489067554, + 0.004484421573579311, + 0.006082974374294281, + 0.005053980275988579, + 0.0034135840833187103, + 0.004346675705164671, + 0.008904937654733658, + 0.006317234132438898, + 0.005943867843598127, + 0.0028348020277917385, + 0.015903864055871964, + 0.008584566414356232, + 0.002466023899614811, + 0.002607316942885518, + 0.006727631203830242, + 0.007524081040173769, + 0.005442104768007994, + 0.004562458023428917, + 0.00964694656431675, + 0.008422196842730045, + 0.0073902481235563755, + 0.002626314526423812, + 0.007179132662713528, + 0.0027781266253441572, + 0.004609361756592989, + 0.01933448016643524, + 0.013105770573019981, + 0.029241610318422318, + 0.02297503501176834, + 0.008765781298279762, + 0.016150834038853645, + 0.0082342978566885, + 0.005537325982004404, + 0.006489331368356943, + 0.03730111941695213, + 0.06818467378616333, + 0.05109989270567894, + 0.015787366777658463, + 0.01883745938539505, + 0.007286801002919674, + 0.00782595295459032, + 0.009497429244220257, + 0.040092840790748596, + 0.08966951072216034, + 0.0072417291812598705, + 0.010861949995160103, + 0.010052108205854893, + 0.0036985657643526793, + 0.0038071535527706146, + 0.002313770353794098, + 0.01765293814241886, + 0.013998712413012981, + 0.029403086751699448, + 0.030674561858177185, + 0.012255340814590454, + 0.011425190605223179, + 0.014551323838531971, + 0.008903843350708485, + 0.03688327595591545, + 0.027631089091300964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.0041167521849274635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016158401966095 + ], + "opposite_logprob_tokens": [ + -1.701615810394287 + ], + "logprob_reward": -0.2016158401966095, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016158401966095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08643589168787003, + 0.10628587007522583, + 0.03808106109499931, + 0.060382213443517685, + 0.03635084629058838, + 0.04640745371580124, + 0.052143990993499756, + 0.4437136650085449, + 0.0451127365231514, + 0.08508630096912384 + ], + "credit_map": [ + 0.008216969668865204, + 0.007240878418087959, + 0.012814231216907501, + 0.00903104618191719, + 0.011294088326394558, + 0.006582087371498346, + 0.016808994114398956, + 0.01444759126752615, + 0.0057121929712593555, + 0.007038572337478399, + 0.014790866523981094, + 0.016425518319010735, + 0.019041217863559723, + 0.008060554042458534, + 0.02040635235607624, + 0.014810589142143726, + 0.0034678264055401087, + 0.003436504164710641, + 0.007320767734199762, + 0.004345436580479145, + 0.0028439979068934917, + 0.004822420887649059, + 0.00507889362052083, + 0.006765215191990137, + 0.0037365106400102377, + 0.004387378692626953, + 0.006272617727518082, + 0.008905251510441303, + 0.006749384570866823, + 0.0038318142760545015, + 0.017573604360222816, + 0.008925656788051128, + 0.0019300024723634124, + 0.0024827392771840096, + 0.0043991548009216785, + 0.004995505791157484, + 0.004211319610476494, + 0.00230460986495018, + 0.009346653707325459, + 0.006680862978100777, + 0.0025768980849534273, + 0.0015824161237105727, + 0.00434903334826231, + 0.004829843528568745, + 0.004822103772312403, + 0.005710580851882696, + 0.00953014474362135, + 0.01300643291324377, + 0.005871347617357969, + 0.002439880045130849, + 0.004844224080443382, + 0.004140679258853197, + 0.0027011078782379627, + 0.0023390771821141243, + 0.010348853655159473, + 0.019458819180727005, + 0.09561339765787125, + 0.02965112403035164, + 0.035817187279462814, + 0.013181323185563087, + 0.012978597544133663, + 0.01698286645114422, + 0.07341008633375168, + 0.1660790741443634, + 0.007971936836838722, + 0.009591500274837017, + 0.007256155833601952, + 0.003751174546778202, + 0.0036790810991078615, + 0.002719320124015212, + 0.004373644944280386, + 0.005769921466708183, + 0.020027298480272293, + 0.02078436128795147, + 0.006538005545735359, + 0.004671004600822926, + 0.006034856662154198, + 0.0026102515403181314, + 0.012738265097141266, + 0.011682254262268543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.2016158401966095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095975775271654 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095975775271654, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095975775271654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11328799277544022, + 0.05303684249520302, + 0.03554996848106384, + 0.0320722833275795, + 0.026286205276846886, + 0.05097489804029465, + 0.30554935336112976, + 0.2596372365951538, + 0.03572755306959152, + 0.08787763118743896 + ], + "credit_map": [ + 0.007354689296334982, + 0.007662761956453323, + 0.022882793098688126, + 0.024074897170066833, + 0.017600901424884796, + 0.005407551769167185, + 0.017696822062134743, + 0.0106075843796134, + 0.0030321995727717876, + 0.0031892703846096992, + 0.009944233112037182, + 0.009362098760902882, + 0.008774020709097385, + 0.003085443517193198, + 0.009802193380892277, + 0.005847386550158262, + 0.0030383397825062275, + 0.0040203966200351715, + 0.005962880328297615, + 0.004705299157649279, + 0.0030298035126179457, + 0.0037493344862014055, + 0.005684677977114916, + 0.005359241273254156, + 0.002156394999474287, + 0.0026010298170149326, + 0.006837344728410244, + 0.0032561824191361666, + 0.0031213376205414534, + 0.001735537895001471, + 0.005761559586971998, + 0.006602900102734566, + 0.0019415206043049693, + 0.002527369884774089, + 0.004235217347741127, + 0.0026989136822521687, + 0.002259021159261465, + 0.002709928434342146, + 0.005712015088647604, + 0.004202221520245075, + 0.004670586436986923, + 0.0018483585445210338, + 0.004340184386819601, + 0.001872081309556961, + 0.0027701479848474264, + 0.011476046405732632, + 0.007319619879126549, + 0.01667787693440914, + 0.039594363421201706, + 0.015313073992729187, + 0.027458440512418747, + 0.013733511790633202, + 0.008767541497945786, + 0.009581049904227257, + 0.06701274961233139, + 0.12408864498138428, + 0.05572159215807915, + 0.017344633117318153, + 0.02032109722495079, + 0.0066413539461791515, + 0.007530517876148224, + 0.010157725773751736, + 0.04382765665650368, + 0.09809268265962601, + 0.005949399434030056, + 0.007278698496520519, + 0.002452343236654997, + 0.0015500482404604554, + 0.0017565510934218764, + 0.0016880846815183759, + 0.008005062118172646, + 0.007047370076179504, + 0.02200595661997795, + 0.022536322474479675, + 0.004665045998990536, + 0.005645120516419411, + 0.005132145714014769, + 0.002092213835567236, + 0.014667747542262077, + 0.011133088730275631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.004095975775271654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203114241361618 + ], + "opposite_logprob_tokens": [ + -2.8120310306549072 + ], + "logprob_reward": -0.06203114241361618, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203114241361618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06425448507070541, + 0.0406576506793499, + 0.020455583930015564, + 0.042131442576646805, + 0.10401053726673126, + 0.33454838395118713, + 0.07574775815010071, + 0.07744067162275314, + 0.06981291621923447, + 0.170940563082695 + ], + "credit_map": [ + 0.005374416243284941, + 0.006008499767631292, + 0.008769289590418339, + 0.010757414624094963, + 0.009792231023311615, + 0.003358484013006091, + 0.011066332459449768, + 0.009127817116677761, + 0.0027449182234704494, + 0.0029051180463284254, + 0.005537056364119053, + 0.007774496916681528, + 0.0064560528844594955, + 0.002378602046519518, + 0.007670002523809671, + 0.005191401112824678, + 0.0020102427806705236, + 0.0019112017471343279, + 0.003387247212231159, + 0.002000776818022132, + 0.002030224073678255, + 0.002649628324434161, + 0.0029581417329609394, + 0.003508122405037284, + 0.002514166757464409, + 0.002473348518833518, + 0.007952200248837471, + 0.004720055032521486, + 0.004559569526463747, + 0.0024204510264098644, + 0.008584946393966675, + 0.008906705304980278, + 0.006647533271461725, + 0.006814771331846714, + 0.01760503649711609, + 0.010688243433833122, + 0.010707689449191093, + 0.0058256108313798904, + 0.029266297817230225, + 0.016455357894301414, + 0.049183133989572525, + 0.012284059077501297, + 0.02421627938747406, + 0.009570011869072914, + 0.018428677693009377, + 0.10006020218133926, + 0.03155047819018364, + 0.08925556391477585, + 0.00983868632465601, + 0.003774140262976289, + 0.007484977133572102, + 0.004037376958876848, + 0.0032318695448338985, + 0.0028094390872865915, + 0.016017040237784386, + 0.028554227203130722, + 0.01508594024926424, + 0.0052771130576729774, + 0.0072923507541418076, + 0.003493943717330694, + 0.0038159508258104324, + 0.004688093904405832, + 0.011484280228614807, + 0.02630300633609295, + 0.015995033085346222, + 0.017325714230537415, + 0.006796479690819979, + 0.0044693415984511375, + 0.007468838710337877, + 0.004017682280391455, + 0.0060377586632966995, + 0.007702072151005268, + 0.04414689913392067, + 0.044207148253917694, + 0.009126248769462109, + 0.008868901990354061, + 0.010599999688565731, + 0.005376233719289303, + 0.026415232568979263, + 0.02219988964498043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.06203114241361618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium close-up focuses on a young soldier, leaning wearily against the damp earthen wall of a muddy trench, against a gloomy sky shrouded in leaden clouds. He appears to be in his early twenties, his face smeared with dirt and sweat, dark circles under his eyes betraying a look of profound weariness. His short, rain-soaked, dark hair is matted to his forehead. He wears a worn, dark green military jacket, the collar turned up to protect against the chill. With slightly trembling but determined hands, he pulls a damp matchbox from his jacket pocket, carefully extracts a match, and vigorously strikes it against the rough side. A small, bright orange flame bursts forth, briefly illuminating his tense jaw and sunken eye sockets, and also revealing the clammy texture of the surrounding rough earth and rotting wooden support structures. He quickly brought the precious flame close to the wrinkled, seemingly last cigarette in his mouth. The cigarette butt hissed faintly at the touch of the flame, igniting into a single crimson point that stood out against the dim light. He took a deep breath, then slowly exhaled a cloud of white smoke that condensed in the cold air. His furrowed brow seemed to relax for a moment. The entire scene was enveloped in a repressive, cold hue. Only that speck of crimson and the momentary emptiness in the soldier's eyes created a complex atmosphere of both exhaustion and tranquility in the intervals between battles.", + "seed": 7158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_6/index_7585_seed_7158_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": -0.025705670844763516, + "vlm_logprob": -0.025705670844763516, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.025705670844763516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02361365780234337 + ], + "opposite_logprob_tokens": [ + -4.523613452911377 + ], + "logprob_reward": -0.02361365780234337, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02361365780234337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17809249460697174, + 0.1138458400964737, + 0.05829405039548874, + 0.06798943877220154, + 0.06760752946138382, + 0.10293590277433395, + 0.11360440403223038, + 0.11025428026914597, + 0.10941703617572784, + 0.07795899361371994 + ], + "credit_map": [ + 0.013917919248342514, + 0.013142910785973072, + 0.024530550464987755, + 0.033126071095466614, + 0.016174957156181335, + 0.018003080040216446, + 0.03441670536994934, + 0.024780310690402985, + 0.007772727869451046, + 0.007871146313846111, + 0.01610313169658184, + 0.014634443446993828, + 0.010851874947547913, + 0.007740678731352091, + 0.028886986896395683, + 0.01998485065996647, + 0.004988424945622683, + 0.007674078457057476, + 0.006182678043842316, + 0.005792285781353712, + 0.005111930891871452, + 0.0039030087646096945, + 0.016041170805692673, + 0.008600474335253239, + 0.004597304854542017, + 0.004924986977130175, + 0.012078401632606983, + 0.007090784143656492, + 0.008403032086789608, + 0.012266210280358791, + 0.00811699591577053, + 0.010511724278330803, + 0.004681081976741552, + 0.005472073797136545, + 0.007135395426303148, + 0.009017007425427437, + 0.00607705395668745, + 0.004639665130525827, + 0.019317904487252235, + 0.011267349123954773, + 0.0069196876138448715, + 0.0064175608567893505, + 0.01810934953391552, + 0.014963548630475998, + 0.00983519572764635, + 0.012737425044178963, + 0.01903260126709938, + 0.014920531772077084, + 0.009406891651451588, + 0.006870506796985865, + 0.02582652121782303, + 0.0210874006152153, + 0.007737844716757536, + 0.009199813939630985, + 0.01724199578166008, + 0.016233431175351143, + 0.008275254629552364, + 0.009517569094896317, + 0.013257242739200592, + 0.027327150106430054, + 0.009667757898569107, + 0.012178744189441204, + 0.013556348159909248, + 0.016474222764372826, + 0.0066787428222596645, + 0.00806702021509409, + 0.01697710156440735, + 0.019041361287236214, + 0.008605221286416054, + 0.012678682804107666, + 0.022207964211702347, + 0.015160942450165749, + 0.006537665147334337, + 0.008612031117081642, + 0.00908246822655201, + 0.013687199912965298, + 0.006116663571447134, + 0.01146641094237566, + 0.00955389253795147, + 0.012902665883302689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02361365780234337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001963712740689516 + ], + "opposite_logprob_tokens": [ + -6.2519636154174805 + ], + "logprob_reward": -0.001963712740689516, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001963712740689516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28505802154541016, + 0.05796467140316963, + 0.04318581894040108, + 0.04478560760617256, + 0.06605471670627594, + 0.03486481308937073, + 0.07467544078826904, + 0.09705553203821182, + 0.23006896674633026, + 0.0662863552570343 + ], + "credit_map": [ + 0.019640639424324036, + 0.019075509160757065, + 0.05694756656885147, + 0.06116606667637825, + 0.021934321150183678, + 0.022173184901475906, + 0.04218266159296036, + 0.04193807765841484, + 0.004581846762448549, + 0.004660376813262701, + 0.008317249827086926, + 0.008430249989032745, + 0.006967392284423113, + 0.005909095983952284, + 0.011116047389805317, + 0.007982416078448296, + 0.003769148839637637, + 0.005742087494581938, + 0.006429816130548716, + 0.005522837862372398, + 0.0032223931048065424, + 0.0041502430103719234, + 0.0074761309660971165, + 0.006873162463307381, + 0.00251912884414196, + 0.0026062075048685074, + 0.004798355046659708, + 0.005690260790288448, + 0.004871041979640722, + 0.0056452881544828415, + 0.011414320208132267, + 0.007241006474941969, + 0.005122978240251541, + 0.0062957704067230225, + 0.005397318862378597, + 0.009690748527646065, + 0.005311054643243551, + 0.005641918629407883, + 0.018322087824344635, + 0.010272840969264507, + 0.0029236730188131332, + 0.002866133349016309, + 0.00655432278290391, + 0.004682846367359161, + 0.0038681314326822758, + 0.003782967571169138, + 0.0059807319194078445, + 0.004206008277833462, + 0.005166715011000633, + 0.005401956848800182, + 0.014530056156218052, + 0.017761709168553352, + 0.004893516190350056, + 0.006803715601563454, + 0.009560049511492252, + 0.010557715781033039, + 0.007153728511184454, + 0.009425840340554714, + 0.01480189710855484, + 0.01928659901022911, + 0.011452030390501022, + 0.008447849191725254, + 0.013504032045602798, + 0.012983549386262894, + 0.0144492844119668, + 0.018335336819291115, + 0.03951934725046158, + 0.05092541128396988, + 0.01645357720553875, + 0.023854278028011322, + 0.031013410538434982, + 0.0355183370411396, + 0.005226570181548595, + 0.006965175736695528, + 0.007626798003911972, + 0.014978514984250069, + 0.005304350052028894, + 0.006924165412783623, + 0.00973519217222929, + 0.009525586850941181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.001963712740689516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796753525733948 + ], + "opposite_logprob_tokens": [ + -4.634796619415283 + ], + "logprob_reward": -0.009796753525733948, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796753525733948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17103759944438934, + 0.08039532601833344, + 0.056698571890592575, + 0.12475498765707016, + 0.30924004316329956, + 0.056658580899238586, + 0.061720870435237885, + 0.032527800649404526, + 0.07707327604293823, + 0.029892917722463608 + ], + "credit_map": [ + 0.011372760869562626, + 0.010254433378577232, + 0.03100665658712387, + 0.03231038898229599, + 0.013409980572760105, + 0.011260299012064934, + 0.036747682839632034, + 0.024675391614437103, + 0.006416273768991232, + 0.0076149399392306805, + 0.00789272878319025, + 0.009669821709394455, + 0.007314074318856001, + 0.004569960292428732, + 0.023122703656554222, + 0.013794831931591034, + 0.0031110465060919523, + 0.003897510003298521, + 0.005887989420443773, + 0.005134615581482649, + 0.005067456513643265, + 0.003692615544423461, + 0.016748638823628426, + 0.013158698566257954, + 0.007702404633164406, + 0.006003872491419315, + 0.010614985600113869, + 0.019849281758069992, + 0.010992813855409622, + 0.006642869208008051, + 0.04005787521600723, + 0.022890890017151833, + 0.021481920033693314, + 0.02003299631178379, + 0.026217643171548843, + 0.04560573771595955, + 0.020866625010967255, + 0.020968394353985786, + 0.09684666246175766, + 0.05722007527947426, + 0.00430472893640399, + 0.0038226949982345104, + 0.010706331580877304, + 0.008997133933007717, + 0.005222210194915533, + 0.009546879678964615, + 0.007087484933435917, + 0.006971118971705437, + 0.0036698244512081146, + 0.00367800728417933, + 0.012071472592651844, + 0.014949913136661053, + 0.004532572813332081, + 0.0077935815788805485, + 0.0076653300784528255, + 0.007360164076089859, + 0.002257692627608776, + 0.00279635121114552, + 0.0052360473200678825, + 0.005945793818682432, + 0.0036627978552132845, + 0.004734320566058159, + 0.003878202987834811, + 0.004016593564301729, + 0.0049412865191698074, + 0.006131881382316351, + 0.00717952661216259, + 0.01767122559249401, + 0.005868830718100071, + 0.009013820439577103, + 0.014946188777685165, + 0.011320510879158974, + 0.002100075129419565, + 0.002634975593537092, + 0.006063454784452915, + 0.0064841886050999165, + 0.0020446458365768194, + 0.003034202614799142, + 0.003609033301472664, + 0.003922342788428068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.009796753525733948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419882569462061 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.01419882569462061, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419882569462061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12846721708774567, + 0.36892732977867126, + 0.06703055649995804, + 0.05126229301095009, + 0.07134708017110825, + 0.14113110303878784, + 0.034027695655822754, + 0.04705150052905083, + 0.056071240454912186, + 0.0346839539706707 + ], + "credit_map": [ + 0.00808970257639885, + 0.008132897317409515, + 0.01928580179810524, + 0.024143299087882042, + 0.018180081620812416, + 0.010754602029919624, + 0.022407852113246918, + 0.017472991719841957, + 0.03395717591047287, + 0.0447610579431057, + 0.07214804738759995, + 0.05450594052672386, + 0.030822182074189186, + 0.03133772686123848, + 0.050582390278577805, + 0.050812818109989166, + 0.005694420076906681, + 0.008433768525719643, + 0.015208769589662552, + 0.007780773099511862, + 0.006788214668631554, + 0.008339579217135906, + 0.007276748772710562, + 0.007508279290050268, + 0.0037173298187553883, + 0.003867111401632428, + 0.009902416728436947, + 0.005867029074579477, + 0.005006053019315004, + 0.008281721733510494, + 0.007161651272326708, + 0.00745898624882102, + 0.004488206002861261, + 0.0045064110308885574, + 0.005992718506604433, + 0.010507242754101753, + 0.005422639194875956, + 0.00442567327991128, + 0.02271903306245804, + 0.013285158202052116, + 0.01000440213829279, + 0.00940953753888607, + 0.0253272857517004, + 0.02482983097434044, + 0.011406008154153824, + 0.024128835648298264, + 0.015487626194953918, + 0.020537571981549263, + 0.002794562140479684, + 0.0016704576555639505, + 0.006840802729129791, + 0.006276126950979233, + 0.002166505204513669, + 0.0027289248537272215, + 0.005116864573210478, + 0.006433448288589716, + 0.0033454031217843294, + 0.0036625093780457973, + 0.010381614789366722, + 0.008529281243681908, + 0.0037158888299018145, + 0.0037996864411979914, + 0.0067977155558764935, + 0.006819403264671564, + 0.003398977452889085, + 0.00399411516264081, + 0.009209650568664074, + 0.00970239844173193, + 0.005005104932934046, + 0.005632230546325445, + 0.011091174557805061, + 0.008037589490413666, + 0.002513859886676073, + 0.0030843312852084637, + 0.006672333460301161, + 0.007462038658559322, + 0.002287545008584857, + 0.0029471188317984343, + 0.004910844843834639, + 0.004805884324014187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.01419882569462061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895540446043015 + ], + "opposite_logprob_tokens": [ + -2.5789554119110107 + ], + "logprob_reward": -0.07895540446043015, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895540446043015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18226175010204315, + 0.10313193500041962, + 0.0428759790956974, + 0.06388770788908005, + 0.06636057049036026, + 0.12214905023574829, + 0.09921830147504807, + 0.13992048799991608, + 0.1392856240272522, + 0.04090850055217743 + ], + "credit_map": [ + 0.011139310896396637, + 0.011701256968080997, + 0.025349028408527374, + 0.0424325056374073, + 0.01580014079809189, + 0.01821533963084221, + 0.03050452098250389, + 0.027119677513837814, + 0.005759659688919783, + 0.004663753788918257, + 0.00712047703564167, + 0.012764187529683113, + 0.012976933270692825, + 0.005289493128657341, + 0.035126056522130966, + 0.01943139359354973, + 0.0030944852624088526, + 0.003797410288825631, + 0.003813068848103285, + 0.006423672661185265, + 0.0060089267790317535, + 0.004121896345168352, + 0.009458224289119244, + 0.006158302538096905, + 0.0044815419241786, + 0.004261985886842012, + 0.012236528098583221, + 0.0067028687335550785, + 0.005207211710512638, + 0.004206950310617685, + 0.015557111240923405, + 0.011233517900109291, + 0.00513519486412406, + 0.005433321930468082, + 0.007586479652673006, + 0.00917741097509861, + 0.0064710527658462524, + 0.00421252241358161, + 0.017512094229459763, + 0.010832513682544231, + 0.007588510867208242, + 0.006797032430768013, + 0.013883200474083424, + 0.027909042313694954, + 0.012040683999657631, + 0.014303673058748245, + 0.016212835907936096, + 0.023414092138409615, + 0.005834026727825403, + 0.008440906181931496, + 0.01110607199370861, + 0.02872513420879841, + 0.00808017048984766, + 0.0071783107705414295, + 0.0166460108011961, + 0.013207684271037579, + 0.009828872047364712, + 0.01134043000638485, + 0.029239192605018616, + 0.024298636242747307, + 0.012722489424049854, + 0.01272188127040863, + 0.021096928045153618, + 0.018672095611691475, + 0.007899375632405281, + 0.008173955604434013, + 0.026373717933893204, + 0.02705199271440506, + 0.011607908643782139, + 0.01564759947359562, + 0.02626560442149639, + 0.016265496611595154, + 0.0038702820893377066, + 0.006185201928019524, + 0.005923624150454998, + 0.006791032385081053, + 0.004756495356559753, + 0.004149151500314474, + 0.004993857350200415, + 0.004238861612975597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.07895540446043015 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium close-up focuses on a young soldier, leaning wearily against the damp earthen wall of a muddy trench, against a gloomy sky shrouded in leaden clouds. He appears to be in his early twenties, his face smeared with dirt and sweat, dark circles under his eyes betraying a look of profound weariness. His short, rain-soaked, dark hair is matted to his forehead. He wears a worn, dark green military jacket, the collar turned up to protect against the chill. With slightly trembling but determined hands, he pulls a damp matchbox from his jacket pocket, carefully extracts a match, and vigorously strikes it against the rough side. A small, bright orange flame bursts forth, briefly illuminating his tense jaw and sunken eye sockets, and also revealing the clammy texture of the surrounding rough earth and rotting wooden support structures. He quickly brought the precious flame close to the wrinkled, seemingly last cigarette in his mouth. The cigarette butt hissed faintly at the touch of the flame, igniting into a single crimson point that stood out against the dim light. He took a deep breath, then slowly exhaled a cloud of white smoke that condensed in the cold air. His furrowed brow seemed to relax for a moment. The entire scene was enveloped in a repressive, cold hue. Only that speck of crimson and the momentary emptiness in the soldier's eyes created a complex atmosphere of both exhaustion and tranquility in the intervals between battles.", + "seed": 7159, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_6/index_7585_seed_7159_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": -0.01953574698418379, + "vlm_logprob": -0.01953574698418379, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.01953574698418379, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020945429801940918 + ], + "opposite_logprob_tokens": [ + -4.2709455490112305 + ], + "logprob_reward": -0.020945429801940918, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020945429801940918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15689365565776825, + 0.06253636628389359, + 0.07304401695728302, + 0.1568480134010315, + 0.16542303562164307, + 0.17503781616687775, + 0.05438293144106865, + 0.05923198163509369, + 0.029188916087150574, + 0.06741325557231903 + ], + "credit_map": [ + 0.008309855125844479, + 0.009023407474160194, + 0.02564574033021927, + 0.02980615571141243, + 0.015166113153100014, + 0.00927293673157692, + 0.03701939433813095, + 0.022650061175227165, + 0.004470093175768852, + 0.007489131763577461, + 0.009820857085287571, + 0.008993301540613174, + 0.008502132259309292, + 0.005215000826865435, + 0.008836854249238968, + 0.009209003299474716, + 0.005380756687372923, + 0.005154469516128302, + 0.011411125771701336, + 0.011548218317329884, + 0.007656664587557316, + 0.004534033592790365, + 0.015104396268725395, + 0.01225435920059681, + 0.018461743369698524, + 0.02004225365817547, + 0.028083639219403267, + 0.02320164628326893, + 0.016148900613188744, + 0.009710059501230717, + 0.0152649050578475, + 0.025934863835573196, + 0.0183261726051569, + 0.01666584238409996, + 0.02699454128742218, + 0.0317070335149765, + 0.013226804323494434, + 0.015213501639664173, + 0.019710201770067215, + 0.023578954860568047, + 0.017737021669745445, + 0.017317064106464386, + 0.02314942330121994, + 0.025453440845012665, + 0.0178244486451149, + 0.01614823006093502, + 0.03300045058131218, + 0.024407727643847466, + 0.0036529034841805696, + 0.004756920970976353, + 0.008261127397418022, + 0.008426264859735966, + 0.005582105368375778, + 0.006897374521940947, + 0.008738995529711246, + 0.008067239075899124, + 0.003503573825582862, + 0.00556356692686677, + 0.012090190313756466, + 0.011221625842154026, + 0.005684277508407831, + 0.003919237293303013, + 0.00885710958391428, + 0.00839240476489067, + 0.002462745178490877, + 0.002774439984932542, + 0.0038282698951661587, + 0.0044412147253751755, + 0.003505052998661995, + 0.003183815162628889, + 0.004844955168664455, + 0.004148424603044987, + 0.00448028277605772, + 0.006505947560071945, + 0.01045959535986185, + 0.011948589235544205, + 0.007649392820894718, + 0.007812258321791887, + 0.008919741958379745, + 0.00963745266199112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.020945429801940918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025064502842724323 + ], + "opposite_logprob_tokens": [ + -6.002506256103516 + ], + "logprob_reward": -0.0025064502842724323, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025064502842724323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11976899206638336, + 0.1752428561449051, + 0.1007501408457756, + 0.2633027136325836, + 0.07164017856121063, + 0.08635146170854568, + 0.033858831971883774, + 0.06223132088780403, + 0.03200433775782585, + 0.054849181324243546 + ], + "credit_map": [ + 0.00855928286910057, + 0.01101017277687788, + 0.018868405371904373, + 0.02961287647485733, + 0.007837049663066864, + 0.0072090052999556065, + 0.018718523904681206, + 0.017953669652342796, + 0.008217411115765572, + 0.014565114863216877, + 0.039457421749830246, + 0.045597899705171585, + 0.01692979224026203, + 0.007109508849680424, + 0.02044043317437172, + 0.022925257682800293, + 0.00690816342830658, + 0.005771513096988201, + 0.012243879027664661, + 0.02160361036658287, + 0.007695494219660759, + 0.004485298879444599, + 0.02489301562309265, + 0.01714915782213211, + 0.015636980533599854, + 0.020090579986572266, + 0.041187189519405365, + 0.040912870317697525, + 0.023199355229735374, + 0.014486645348370075, + 0.056294035166502, + 0.05149505287408829, + 0.00777601171284914, + 0.007363055367022753, + 0.008729690685868263, + 0.01169960293918848, + 0.006166006438434124, + 0.005570529028773308, + 0.014195007272064686, + 0.010140271857380867, + 0.0063984086737036705, + 0.005725896917283535, + 0.008004961535334587, + 0.011681764386594296, + 0.009038249030709267, + 0.006045536603778601, + 0.02474009245634079, + 0.01471654698252678, + 0.0021365887951105833, + 0.0016306506004184484, + 0.004144078120589256, + 0.005025219637900591, + 0.0036969527136534452, + 0.003286990337073803, + 0.008893298916518688, + 0.005045051220804453, + 0.0038154497742652893, + 0.006857112981379032, + 0.009413724765181541, + 0.014659728854894638, + 0.005597380921244621, + 0.004130709916353226, + 0.008959081955254078, + 0.008798129856586456, + 0.002380561316385865, + 0.00231256359256804, + 0.005961222108453512, + 0.007495306897908449, + 0.002444265177473426, + 0.0027864379808306694, + 0.004370577167719603, + 0.004253402817994356, + 0.003891044296324253, + 0.005214683711528778, + 0.011860406957566738, + 0.010265065357089043, + 0.0048926277086138725, + 0.004088427405804396, + 0.007046151906251907, + 0.007590772118419409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0025064502842724323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552776373922825 + ], + "opposite_logprob_tokens": [ + -4.387552738189697 + ], + "logprob_reward": -0.012552776373922825, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552776373922825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2383188009262085, + 0.06747470051050186, + 0.28300970792770386, + 0.15992046892642975, + 0.0863540768623352, + 0.029712896794080734, + 0.048315420746803284, + 0.03606585040688515, + 0.037754639983177185, + 0.013073432259261608 + ], + "credit_map": [ + 0.011789878830313683, + 0.014028680510818958, + 0.02193463034927845, + 0.04828809201717377, + 0.019360419362783432, + 0.018333371728658676, + 0.06071366369724274, + 0.04387005418539047, + 0.003472303505986929, + 0.004863866604864597, + 0.012548042461276054, + 0.015538172796368599, + 0.004670858383178711, + 0.002684668404981494, + 0.012630614452064037, + 0.01106617134064436, + 0.025405865162611008, + 0.01785385236144066, + 0.04714538902044296, + 0.06663984060287476, + 0.01960545778274536, + 0.022627444937825203, + 0.03792792558670044, + 0.04580392315983772, + 0.009472248144447803, + 0.011483501642942429, + 0.02562209777534008, + 0.028381740674376488, + 0.013170505873858929, + 0.009907162748277187, + 0.03538781404495239, + 0.02649538777768612, + 0.005372219253331423, + 0.006125921383500099, + 0.014254461042582989, + 0.014633839949965477, + 0.005281879100948572, + 0.0032414626330137253, + 0.022830095142126083, + 0.014614195562899113, + 0.0038536677602678537, + 0.0037380214780569077, + 0.00582377053797245, + 0.003433659439906478, + 0.0027054979000240564, + 0.001970478566363454, + 0.004013958387076855, + 0.004173842258751392, + 0.004054916091263294, + 0.003578516189008951, + 0.00821587536484003, + 0.006757794879376888, + 0.004264997784048319, + 0.005118444561958313, + 0.009535145945847034, + 0.006789728067815304, + 0.002353133400902152, + 0.004650282673537731, + 0.004918502178043127, + 0.008700587786734104, + 0.00321995303966105, + 0.0022029541432857513, + 0.004938106052577496, + 0.005082330666482449, + 0.0019225692376494408, + 0.0027111226227134466, + 0.007707423530519009, + 0.011332102119922638, + 0.002452873857691884, + 0.002318534068763256, + 0.004352635238319635, + 0.0049573760479688644, + 0.0017694670241326094, + 0.0012854981468990445, + 0.0017494115745648742, + 0.002028110669925809, + 0.001084120012819767, + 0.0013076186878606677, + 0.0017332960851490498, + 0.0021159108728170395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.012552776373922825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673760985955596 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.00673760985955596, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673760985955596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1461084634065628, + 0.057185105979442596, + 0.12201744318008423, + 0.11321786046028137, + 0.14290404319763184, + 0.15881268680095673, + 0.0351090207695961, + 0.11318442225456238, + 0.04364773631095886, + 0.06781328469514847 + ], + "credit_map": [ + 0.01096526999026537, + 0.010668731294572353, + 0.02307140827178955, + 0.03384830057621002, + 0.008616314269602299, + 0.00783457886427641, + 0.026889188215136528, + 0.02421467751264572, + 0.003901370568200946, + 0.0067791626788675785, + 0.008177999407052994, + 0.010384180583059788, + 0.0067211356945335865, + 0.00423813983798027, + 0.009127555415034294, + 0.007855557836592197, + 0.010662558488547802, + 0.00894634984433651, + 0.018567422404885292, + 0.031346358358860016, + 0.009275741875171661, + 0.006986097898334265, + 0.014984567649662495, + 0.021248338744044304, + 0.011973345652222633, + 0.01223113015294075, + 0.01794281043112278, + 0.018645478412508965, + 0.00908572506159544, + 0.007703311275690794, + 0.01936628855764866, + 0.016269763931632042, + 0.019868405535817146, + 0.018531370908021927, + 0.019682878628373146, + 0.022942787036299706, + 0.013894813135266304, + 0.014910100027918816, + 0.013771764002740383, + 0.01930190622806549, + 0.016840634867548943, + 0.018384018912911415, + 0.01781674660742283, + 0.02272266149520874, + 0.01608138158917427, + 0.014747616834938526, + 0.03023049607872963, + 0.021989116445183754, + 0.0023839918430894613, + 0.002435685833916068, + 0.0051059178076684475, + 0.004410446621477604, + 0.004219658672809601, + 0.0037554975133389235, + 0.0076985787600278854, + 0.0050992402248084545, + 0.007926776073873043, + 0.010542352683842182, + 0.01701435074210167, + 0.024020133540034294, + 0.010031303390860558, + 0.012719765305519104, + 0.01624998450279236, + 0.014679751358926296, + 0.004063975531607866, + 0.0026996745727956295, + 0.005002189427614212, + 0.00982380099594593, + 0.002853330224752426, + 0.008042886853218079, + 0.006170226261019707, + 0.004991648718714714, + 0.004789219703525305, + 0.006032106466591358, + 0.011750033125281334, + 0.009881064295768738, + 0.005987526848912239, + 0.00932358205318451, + 0.011891811154782772, + 0.008157937787473202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.00673760985955596 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young soldier in his early twenties leaning against a damp earthen wall of a muddy trench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a worn, dark green military jacket with the collar turned up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soldier strike a match to light a cigarette, resulting in a visible flame and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a gloomy sky with leaden clouds and rotting wooden support structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slowly advancing medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493646860122681 + ], + "opposite_logprob_tokens": [ + -2.929936408996582 + ], + "logprob_reward": -0.05493646860122681, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493646860122681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15248101949691772, + 0.08351510763168335, + 0.09471455961465836, + 0.20095977187156677, + 0.09916035830974579, + 0.09890706837177277, + 0.04490618035197258, + 0.04572479799389839, + 0.13994571566581726, + 0.03968539461493492 + ], + "credit_map": [ + 0.006020741071552038, + 0.006040922366082668, + 0.023694608360528946, + 0.03718738630414009, + 0.01283292192965746, + 0.005184252746403217, + 0.038367729634046555, + 0.023152461275458336, + 0.006313465069979429, + 0.011470518074929714, + 0.011468715034425259, + 0.01457990426570177, + 0.008949090726673603, + 0.007095338311046362, + 0.01288533303886652, + 0.010752742178738117, + 0.00598343089222908, + 0.005960209295153618, + 0.01245211437344551, + 0.011688739992678165, + 0.009442602284252644, + 0.005459620151668787, + 0.025818729773163795, + 0.01790911890566349, + 0.011593110859394073, + 0.014601781032979488, + 0.03263157606124878, + 0.034436095505952835, + 0.015871889889240265, + 0.013300986029207706, + 0.04027724638581276, + 0.03824708238244057, + 0.005401152186095715, + 0.004976230673491955, + 0.012165147811174393, + 0.014103959314525127, + 0.010411885567009449, + 0.005391176789999008, + 0.029281798750162125, + 0.01742900162935257, + 0.008025002665817738, + 0.006780319847166538, + 0.012797173112630844, + 0.01177801564335823, + 0.01055529247969389, + 0.005640068557113409, + 0.027467796579003334, + 0.01586340367794037, + 0.0025327804032713175, + 0.004666510503739119, + 0.008710545487701893, + 0.005989858414977789, + 0.004177832044661045, + 0.0030971872620284557, + 0.008755378425121307, + 0.00697608757764101, + 0.003792162286117673, + 0.00653774943202734, + 0.005823688115924597, + 0.00821689423173666, + 0.004802108742296696, + 0.0052114212885499, + 0.00480312667787075, + 0.006537647917866707, + 0.0069182561710476875, + 0.010650025680661201, + 0.026853369548916817, + 0.04102043807506561, + 0.010317569598555565, + 0.008912485092878342, + 0.016639966517686844, + 0.018633613362908363, + 0.0030563203617930412, + 0.0041899667121469975, + 0.006196909584105015, + 0.008005155250430107, + 0.0037895278073847294, + 0.0035668127238750458, + 0.005730277858674526, + 0.00515042245388031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.05493646860122681 + } + ], + "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/0000025/group_7/index_tensor([6896, 6896])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_tensor([6896, 6896])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eecbc1c8d6900f6e2300b87510dddc14c1758a73 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_tensor([6896, 6896])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a precarious, handheld, first-person perspective, the viewer is positioned within an ancient, ornate palace courtyard, where the ground suddenly trembles and heaves with violent force. The scene is bathed in midday sunlight, which casts sharp, deep shadows across the surrounding marble columns and arched colonnades. What begins as a low, guttural rumble gradually escalates into a piercing vibration, causing the camera to tilt downward. The view focuses on the floor, where intricate mosaic tiles, featuring geometric patterns of deep lapis lazuli, terracotta, and polished ivory, begin to fracture. A network of fine hairline cracks rapidly spiderweb across the polished surface, before the tiles buckle upward and part with a series of sharp, thrilling resonations. Immediately afterwards, a deep surface fissure opens, decisively cutting through the stone floor, revealing a dark, jagged fissure in the raw earth beneath. The entire catastrophic event is captured in a surreal and clean style, rendering every shattered fragment and rising dust mote with razor-sharp photographic detail, emphasizing the severe, irreversible damage inflicted on the beautiful historic structure and creating an atmosphere of sudden, terrifying destruction.", + "seed": 8720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_6896_seed_8720_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": -0.2264680251479149, + "vlm_logprob": -0.2264680251479149, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.2264680251479149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37716445326805115 + ], + "opposite_logprob_tokens": [ + -1.2521644830703735 + ], + "logprob_reward": -0.37716445326805115, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.37716445326805115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0564599372446537, + 0.019483355805277824, + 0.08208031207323074, + 0.03983788937330246, + 0.045855872333049774, + 0.09773102402687073, + 0.3027461767196655, + 0.011680686846375465, + 0.3198062479496002, + 0.024318519979715347 + ], + "credit_map": [ + 0.006087115965783596, + 0.010262569412589073, + 0.007353549357503653, + 0.00803465861827135, + 0.004928965587168932, + 0.0060196490958333015, + 0.006160785909742117, + 0.007612643297761679, + 0.0030435933731496334, + 0.001682303729467094, + 0.0026995621155947447, + 0.0032005049288272858, + 0.001629250473342836, + 0.002388333436101675, + 0.00272069638594985, + 0.0021191113628447056, + 0.003741078544408083, + 0.00702429935336113, + 0.012859757989645004, + 0.028104489669203758, + 0.005063954275101423, + 0.0034668876323848963, + 0.016569573432207108, + 0.005250272341072559, + 0.0031490835826843977, + 0.0029332702979445457, + 0.01183687336742878, + 0.005718025844544172, + 0.0022738270927220583, + 0.0023060638923197985, + 0.006026616785675287, + 0.005594129674136639, + 0.0032764202915132046, + 0.004886654671281576, + 0.0071341716684401035, + 0.004756629932671785, + 0.003708942560479045, + 0.0029239016585052013, + 0.0077791581861674786, + 0.011389994993805885, + 0.013222324661910534, + 0.01509313564747572, + 0.009038225747644901, + 0.009640227071940899, + 0.010843489319086075, + 0.014171205461025238, + 0.012703881599009037, + 0.01301853358745575, + 0.01429666392505169, + 0.024398675188422203, + 0.04320010542869568, + 0.017633652314543724, + 0.11793296039104462, + 0.06197122111916542, + 0.0120459645986557, + 0.011266936548054218, + 0.0011811184231191874, + 0.0011690614046528935, + 0.0018487968482077122, + 0.001576353213749826, + 0.0009959344752132893, + 0.0017457848880439997, + 0.0015255670296028256, + 0.0016380702145397663, + 0.02296505868434906, + 0.016845526173710823, + 0.047355324029922485, + 0.022242410108447075, + 0.015750402584671974, + 0.072423055768013, + 0.10035695880651474, + 0.021867502480745316, + 0.0025241586845368147, + 0.0023745691869407892, + 0.0028748998884111643, + 0.0067945425398647785, + 0.0016371875535696745, + 0.002902187639847398, + 0.0030306053813546896, + 0.002180368173867464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.37716445326805115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215279370546341 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.014215279370546341, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215279370546341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1313825249671936, + 0.029270874336361885, + 0.03688475117087364, + 0.048838451504707336, + 0.14188875257968903, + 0.0428980253636837, + 0.08189614117145538, + 0.03127091005444527, + 0.3441876769065857, + 0.11148186028003693 + ], + "credit_map": [ + 0.013007262721657753, + 0.02522934600710869, + 0.019448574632406235, + 0.01673288084566593, + 0.009426278993487358, + 0.012893128208816051, + 0.015337657183408737, + 0.019307399168610573, + 0.0033222383353859186, + 0.0030644226353615522, + 0.005118510220199823, + 0.005589033477008343, + 0.0019511793507263064, + 0.003035672940313816, + 0.00400963332504034, + 0.0031801825389266014, + 0.0031171771697700024, + 0.003617497393861413, + 0.005986921954900026, + 0.0050732740201056, + 0.004259247798472643, + 0.0026131232734769583, + 0.005964300595223904, + 0.006253210827708244, + 0.0041836947202682495, + 0.007001012563705444, + 0.005612543318420649, + 0.00640845624729991, + 0.004603498615324497, + 0.004501736257225275, + 0.00808145385235548, + 0.008446058258414268, + 0.009300941601395607, + 0.010777609422802925, + 0.027287418022751808, + 0.02224491909146309, + 0.012312165461480618, + 0.006500640418380499, + 0.02560994029045105, + 0.027855122461915016, + 0.0048994882963597775, + 0.005145242437720299, + 0.004246317315846682, + 0.0049661388620734215, + 0.004274075850844383, + 0.007939739152789116, + 0.005683847703039646, + 0.005743173882365227, + 0.004737774841487408, + 0.003236772259697318, + 0.03170191869139671, + 0.011019921861588955, + 0.008759543299674988, + 0.01125484798103571, + 0.005365425255149603, + 0.00581993767991662, + 0.003106525167822838, + 0.0029159076511859894, + 0.00388296227902174, + 0.0031035481952130795, + 0.0031526973471045494, + 0.004605327267199755, + 0.005002379883080721, + 0.005501564592123032, + 0.01484235841780901, + 0.013773951679468155, + 0.03289761394262314, + 0.020077243447303772, + 0.016207972541451454, + 0.10606301575899124, + 0.12271804362535477, + 0.01760747656226158, + 0.016556838527321815, + 0.006892780773341656, + 0.007235265802592039, + 0.040982089936733246, + 0.006629199720919132, + 0.012496665120124817, + 0.014776984229683876, + 0.005912046413868666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.014215279370546341 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038111645728349686 + ], + "opposite_logprob_tokens": [ + -3.288111686706543 + ], + "logprob_reward": -0.038111645728349686, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038111645728349686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030329136177897453, + 0.019970491528511047, + 0.054903484880924225, + 0.0707533210515976, + 0.11842861026525497, + 0.055416978895664215, + 0.07609331607818604, + 0.047981586307287216, + 0.45585715770721436, + 0.07026593387126923 + ], + "credit_map": [ + 0.003960480447858572, + 0.004365843720734119, + 0.0044636898674070835, + 0.0038347213994711637, + 0.002739159855991602, + 0.0032497630454599857, + 0.004055771976709366, + 0.003659707959741354, + 0.0018504170002415776, + 0.0016902030911296606, + 0.0030441710259765387, + 0.002970555331557989, + 0.0020511881448328495, + 0.002605943940579891, + 0.0023240363225340843, + 0.00343397818505764, + 0.002497360110282898, + 0.004354362841695547, + 0.007936076261103153, + 0.015625422820448875, + 0.0048070028424263, + 0.002553586382418871, + 0.010576237924396992, + 0.006553441286087036, + 0.005346337333321571, + 0.006170707289129496, + 0.013775825500488281, + 0.01603817194700241, + 0.005579729098826647, + 0.004761592485010624, + 0.010213261470198631, + 0.00886769499629736, + 0.007276157382875681, + 0.008536233566701412, + 0.023864202201366425, + 0.015333687886595726, + 0.009322582744061947, + 0.005954016000032425, + 0.02610156126320362, + 0.022040171548724174, + 0.0061712199822068214, + 0.005916565656661987, + 0.006105736829340458, + 0.007375698070973158, + 0.005463021341711283, + 0.00881461426615715, + 0.007503055036067963, + 0.00806706678122282, + 0.0032323559280484915, + 0.010194490663707256, + 0.0072810593992471695, + 0.0029174298979341984, + 0.02632961980998516, + 0.017340686172246933, + 0.0044472538866102695, + 0.0043504247441887856, + 0.004946243949234486, + 0.00384370144456625, + 0.0057359728962183, + 0.004318172577768564, + 0.005851002875715494, + 0.008022816851735115, + 0.006835664622485638, + 0.008428010158240795, + 0.01934417523443699, + 0.020289018750190735, + 0.045770544558763504, + 0.027649929746985435, + 0.022187821567058563, + 0.13961197435855865, + 0.15920250117778778, + 0.021801209077239037, + 0.00493472209200263, + 0.006430536042898893, + 0.005790686700493097, + 0.019205857068300247, + 0.006247419863939285, + 0.012010999023914337, + 0.008876988664269447, + 0.006768719293177128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.038111645728349686 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002581655979156 + ], + "opposite_logprob_tokens": [ + -2.69502592086792 + ], + "logprob_reward": -0.07002581655979156, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002581655979156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02597671188414097, + 0.02198176272213459, + 0.05925607308745384, + 0.04772060737013817, + 0.044776882976293564, + 0.05304117500782013, + 0.36364585161209106, + 0.07641550153493881, + 0.28374937176704407, + 0.023436032235622406 + ], + "credit_map": [ + 0.0047105285339057446, + 0.003911636769771576, + 0.0029524783603847027, + 0.0036555666010826826, + 0.002292195102199912, + 0.0030185766518115997, + 0.0026098673697561026, + 0.0028258629608899355, + 0.002620317041873932, + 0.0021672227885574102, + 0.0022139069624245167, + 0.0026919173542410135, + 0.0025564166717231274, + 0.0031761343125253916, + 0.00235162815079093, + 0.004204222001135349, + 0.0028660984244197607, + 0.004565345123410225, + 0.0068216524086892605, + 0.016802599653601646, + 0.004646309185773134, + 0.0024490824434906244, + 0.012903223745524883, + 0.00820176675915718, + 0.003152240999042988, + 0.006300415378063917, + 0.006819401402026415, + 0.01371519174426794, + 0.0026912055909633636, + 0.0030984922777861357, + 0.005395286716520786, + 0.0065483772195875645, + 0.0036881333217024803, + 0.00479196198284626, + 0.007345161400735378, + 0.0065170153975486755, + 0.0049961325712502, + 0.00315732485614717, + 0.008045719005167484, + 0.006235437002032995, + 0.004797728732228279, + 0.0058961412869393826, + 0.006136229261755943, + 0.009919014759361744, + 0.004592397715896368, + 0.007304258644580841, + 0.0075477235950529575, + 0.006847684271633625, + 0.015052205882966518, + 0.04831420257687569, + 0.12430711835622787, + 0.0420595183968544, + 0.039315856993198395, + 0.04677986726164818, + 0.026600543409585953, + 0.021216558292508125, + 0.007032810710370541, + 0.006413029972463846, + 0.010853121988475323, + 0.00934805441647768, + 0.006101451814174652, + 0.010644730180501938, + 0.01179638784378767, + 0.014225911349058151, + 0.01909075863659382, + 0.009854269213974476, + 0.027052391320466995, + 0.10987480729818344, + 0.009268909692764282, + 0.07589495182037354, + 0.018693041056394577, + 0.014020266011357307, + 0.005100930575281382, + 0.0014207609929144382, + 0.0021204722579568624, + 0.006759749259799719, + 0.001735682599246502, + 0.0020512870978564024, + 0.0025355324614793062, + 0.0017116173403337598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.07002581655979156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328229308128357 + ], + "opposite_logprob_tokens": [ + -0.7578229308128357 + ], + "logprob_reward": -0.6328229308128357, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328229308128357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015721123665571213, + 0.009445786476135254, + 0.009046264924108982, + 0.011451225727796555, + 0.015384715050458908, + 0.011443914845585823, + 0.07257624715566635, + 0.0113017987459898, + 0.7871125936508179, + 0.056516293436288834 + ], + "credit_map": [ + 0.002238987013697624, + 0.003383956616744399, + 0.0021453979425132275, + 0.0022229624446481466, + 0.0012861228315159678, + 0.0017229507211595774, + 0.001464399858377874, + 0.001256345771253109, + 0.0008866420248523355, + 0.0008017625659704208, + 0.0011269134702160954, + 0.0015412989305332303, + 0.0009153978317044675, + 0.0014241908211261034, + 0.0015930348308756948, + 0.001156546175479889, + 0.0006920606247149408, + 0.0010368707589805126, + 0.0010857597226276994, + 0.0012165711959823966, + 0.0008973357616923749, + 0.0007110643782652915, + 0.0013764742761850357, + 0.0020301288459450006, + 0.0006572384154424071, + 0.0013113380409777164, + 0.0013931100256741047, + 0.0030345830600708723, + 0.0005622899625450373, + 0.0006958761368878186, + 0.0015830215997993946, + 0.002213768195360899, + 0.0015862231375649571, + 0.0015091703971847892, + 0.0019791957456618547, + 0.0018453346565365791, + 0.0016598744550719857, + 0.0011196413543075323, + 0.0025531521532684565, + 0.0031321225687861443, + 0.0015326716238632798, + 0.001131141441874206, + 0.001359540387056768, + 0.001196664641611278, + 0.001085619442164898, + 0.0020443168468773365, + 0.0015099203446879983, + 0.0015840400010347366, + 0.003881010925397277, + 0.013477454893290997, + 0.021045712754130363, + 0.006764414720237255, + 0.01152781117707491, + 0.006433442234992981, + 0.005685703828930855, + 0.003760700114071369, + 0.0018297543283551931, + 0.0011892179027199745, + 0.0017831996083259583, + 0.0019568903371691704, + 0.0008539955015294254, + 0.0013004436623305082, + 0.0012744477717205882, + 0.0011138494592159986, + 0.03593364730477333, + 0.04053424671292305, + 0.07657862454652786, + 0.21480251848697662, + 0.0333237387239933, + 0.26172521710395813, + 0.09068524837493896, + 0.03352935612201691, + 0.00899071991443634, + 0.00364087987691164, + 0.0035288981162011623, + 0.020603742450475693, + 0.0035026653204113245, + 0.005912154447287321, + 0.00721456715837121, + 0.003122663591057062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.6328229308128357 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a precarious, handheld, first-person perspective, the viewer is positioned within an ancient, ornate palace courtyard, where the ground suddenly trembles and heaves with violent force. The scene is bathed in midday sunlight, which casts sharp, deep shadows across the surrounding marble columns and arched colonnades. What begins as a low, guttural rumble gradually escalates into a piercing vibration, causing the camera to tilt downward. The view focuses on the floor, where intricate mosaic tiles, featuring geometric patterns of deep lapis lazuli, terracotta, and polished ivory, begin to fracture. A network of fine hairline cracks rapidly spiderweb across the polished surface, before the tiles buckle upward and part with a series of sharp, thrilling resonations. Immediately afterwards, a deep surface fissure opens, decisively cutting through the stone floor, revealing a dark, jagged fissure in the raw earth beneath. The entire catastrophic event is captured in a surreal and clean style, rendering every shattered fragment and rising dust mote with razor-sharp photographic detail, emphasizing the severe, irreversible damage inflicted on the beautiful historic structure and creating an atmosphere of sudden, terrifying destruction.", + "seed": 8721, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_6896_seed_8721_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": -0.23087576553225517, + "vlm_logprob": -0.23087576553225517, + "vlm_yes_no_margin": 1.749999988079071, + "vlm_token_credit": -0.23087576553225517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31463608145713806 + ], + "opposite_logprob_tokens": [ + -1.4396361112594604 + ], + "logprob_reward": -0.31463608145713806, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.31463608145713806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1316061019897461, + 0.4001839756965637, + 0.0380125567317009, + 0.11407950520515442, + 0.06273160874843597, + 0.0589110292494297, + 0.049149367958307266, + 0.0492381788790226, + 0.025403590872883797, + 0.07068412005901337 + ], + "credit_map": [ + 0.015000936575233936, + 0.0198239516466856, + 0.020794695243239403, + 0.0152674475684762, + 0.013104312121868134, + 0.011454043909907341, + 0.012192470021545887, + 0.023968245834112167, + 0.027353206649422646, + 0.022472387179732323, + 0.014969629235565662, + 0.017005255445837975, + 0.03882870450615883, + 0.12659896910190582, + 0.12166833877563477, + 0.031287457793951035, + 0.002788878744468093, + 0.0024030476342886686, + 0.0032777932938188314, + 0.003489239141345024, + 0.0032220776192843914, + 0.007694162894040346, + 0.01055561751127243, + 0.004581738263368607, + 0.007415438070893288, + 0.005884923506528139, + 0.011083426885306835, + 0.005848458968102932, + 0.008204946294426918, + 0.0101919025182724, + 0.03334541991353035, + 0.0321049839258194, + 0.0051747895777225494, + 0.005256175063550472, + 0.0036230830010026693, + 0.007679431699216366, + 0.010121054947376251, + 0.012686967849731445, + 0.006725136190652847, + 0.011464963667094707, + 0.007430612575262785, + 0.008403230458498001, + 0.005533348303288221, + 0.011577947065234184, + 0.004195189103484154, + 0.004431989975273609, + 0.007538062520325184, + 0.00980064831674099, + 0.005383297335356474, + 0.00587508687749505, + 0.0074301064014434814, + 0.00995799619704485, + 0.0038730893284082413, + 0.0037209223955869675, + 0.0075953188352286816, + 0.005313549190759659, + 0.0049865213222801685, + 0.005877884104847908, + 0.009363677352666855, + 0.005320356227457523, + 0.005411969963461161, + 0.004914885852485895, + 0.008371591567993164, + 0.0049912878312170506, + 0.0037323979195207357, + 0.002772173611447215, + 0.003299685660749674, + 0.0023525040596723557, + 0.0031409256625920534, + 0.003683890914544463, + 0.002727022161707282, + 0.003694989951327443, + 0.008001388050615788, + 0.019737806171178818, + 0.007121211849153042, + 0.004180599469691515, + 0.015508354641497135, + 0.004321637563407421, + 0.003884833538904786, + 0.007928281091153622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.31463608145713806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134314715862274 + ], + "opposite_logprob_tokens": [ + -1.3134315013885498 + ], + "logprob_reward": -0.3134314715862274, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134314715862274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11143174767494202, + 0.40013498067855835, + 0.03541138023138046, + 0.11642488092184067, + 0.08536890149116516, + 0.08527307212352753, + 0.01449153944849968, + 0.03782406821846962, + 0.03213483467698097, + 0.08150459825992584 + ], + "credit_map": [ + 0.013033458963036537, + 0.017919598147273064, + 0.016744114458560944, + 0.010051612742245197, + 0.010547171346843243, + 0.009306848049163818, + 0.009973080828785896, + 0.02385585941374302, + 0.02091262675821781, + 0.018443571403622627, + 0.015897156670689583, + 0.018741333857178688, + 0.04666745290160179, + 0.11026289314031601, + 0.14210860431194305, + 0.027101313695311546, + 0.0032989359460771084, + 0.0028818659484386444, + 0.0037166269030421972, + 0.00446186400949955, + 0.00592588959261775, + 0.007048942614346743, + 0.003948817495256662, + 0.004128437023609877, + 0.005657591857016087, + 0.003422848414629698, + 0.010048412717878819, + 0.004271919839084148, + 0.0092468885704875, + 0.04301443696022034, + 0.02090481109917164, + 0.019857970997691154, + 0.006796031259000301, + 0.006874450482428074, + 0.0068342718295753, + 0.00928752776235342, + 0.010507659055292606, + 0.016305724158883095, + 0.009534258395433426, + 0.019228985533118248, + 0.018182925879955292, + 0.011731909587979317, + 0.004843829199671745, + 0.006324262358248234, + 0.005187307018786669, + 0.006891735829412937, + 0.01975206658244133, + 0.012359040789306164, + 0.0018756872741505504, + 0.002346678404137492, + 0.0022082773502916098, + 0.002249439712613821, + 0.0015679937787353992, + 0.0018209714908152819, + 0.0011757205938920379, + 0.001246770960278809, + 0.005479717161506414, + 0.004433863330632448, + 0.00780831091105938, + 0.004531777463853359, + 0.004513994324952364, + 0.0034601292572915554, + 0.0033396917860955, + 0.004256587941199541, + 0.0038708485662937164, + 0.0028182934038341045, + 0.005292115267366171, + 0.003492070594802499, + 0.004096408374607563, + 0.003529443871229887, + 0.0036627021618187428, + 0.005372950807213783, + 0.009636521339416504, + 0.021730905398726463, + 0.008696788921952248, + 0.005597811192274094, + 0.016080057248473167, + 0.005901549477130175, + 0.005753105506300926, + 0.008107863366603851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.3134314715862274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048679646104574203 + ], + "opposite_logprob_tokens": [ + -3.0486795902252197 + ], + "logprob_reward": -0.048679646104574203, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048679646104574203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05188703164458275, + 0.2313721477985382, + 0.08830048888921738, + 0.03502793610095978, + 0.0723886787891388, + 0.04980667680501938, + 0.03915930166840553, + 0.11499477177858353, + 0.19677351415157318, + 0.12028944492340088 + ], + "credit_map": [ + 0.007091511506587267, + 0.006814301013946533, + 0.007510970812290907, + 0.005378204397857189, + 0.006051885429769754, + 0.004919683560729027, + 0.0050783706828951836, + 0.00904210563749075, + 0.01319016795605421, + 0.012745351530611515, + 0.008330333977937698, + 0.009525352157652378, + 0.024511678144335747, + 0.07177074998617172, + 0.07765170931816101, + 0.013646810315549374, + 0.007337101735174656, + 0.005652590189129114, + 0.007482254412025213, + 0.016919773072004318, + 0.01635129004716873, + 0.009841511026024818, + 0.008323104120790958, + 0.016392862424254417, + 0.003464481560513377, + 0.0024979510344564915, + 0.005423101596534252, + 0.0028959885239601135, + 0.003605116857215762, + 0.00703879538923502, + 0.007252969779074192, + 0.0028495334554463625, + 0.007761071901768446, + 0.0076537965796887875, + 0.010035401210188866, + 0.009626802057027817, + 0.011374960653483868, + 0.0091762226074934, + 0.007555051706731319, + 0.009205379523336887, + 0.008013126440346241, + 0.005885464604943991, + 0.005225198343396187, + 0.006529009900987148, + 0.0029661746229976416, + 0.004556605126708746, + 0.009564680978655815, + 0.0070664179511368275, + 0.005663527641445398, + 0.005230015143752098, + 0.007427203468978405, + 0.0052725705318152905, + 0.002334680873900652, + 0.004126162268221378, + 0.005773105192929506, + 0.003332037478685379, + 0.01624903827905655, + 0.013143553398549557, + 0.021190466359257698, + 0.013705774210393429, + 0.011458970606327057, + 0.011511795222759247, + 0.015610848553478718, + 0.012124334461987019, + 0.024235324934124947, + 0.025931432843208313, + 0.03060799650847912, + 0.015369625762104988, + 0.03209557384252548, + 0.02659401297569275, + 0.015949102118611336, + 0.025990452617406845, + 0.00635676272213459, + 0.036444488912820816, + 0.008425629697740078, + 0.006265365052968264, + 0.03855542093515396, + 0.006642263848334551, + 0.006025238428264856, + 0.01157428603619337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.048679646104574203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868316650390625 + ], + "opposite_logprob_tokens": [ + -3.0486831665039062 + ], + "logprob_reward": -0.04868316650390625, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868316650390625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08272019773721695, + 0.06533920764923096, + 0.04714132472872734, + 0.05543356388807297, + 0.11054053157567978, + 0.04744430258870125, + 0.11780694872140884, + 0.06172460690140724, + 0.24473203718662262, + 0.16711728274822235 + ], + "credit_map": [ + 0.009464401751756668, + 0.012549186125397682, + 0.012876401655375957, + 0.009976543486118317, + 0.008432885631918907, + 0.006352452095597982, + 0.007348633371293545, + 0.015719689428806305, + 0.010070755146443844, + 0.005709468852728605, + 0.004883348476141691, + 0.003742019645869732, + 0.007537076715379953, + 0.00904142390936613, + 0.011358347721397877, + 0.012996764853596687, + 0.0026308130472898483, + 0.0026024924591183662, + 0.0035769320093095303, + 0.0071928794495761395, + 0.005885718390345573, + 0.011986354365944862, + 0.005834971554577351, + 0.00743116345256567, + 0.004560793749988079, + 0.004360384773463011, + 0.009387986734509468, + 0.0041432506404817104, + 0.006603800226002932, + 0.008618926629424095, + 0.008200286887586117, + 0.009558135643601418, + 0.009245889261364937, + 0.007903304882347584, + 0.007299151737242937, + 0.016096968203783035, + 0.011441485956311226, + 0.01796163059771061, + 0.014771874062716961, + 0.02582022361457348, + 0.010948650538921356, + 0.003864604514092207, + 0.002653507050126791, + 0.005950463470071554, + 0.00234132818877697, + 0.0025071250274777412, + 0.013360266573727131, + 0.005818354897201061, + 0.014695998281240463, + 0.02052844502031803, + 0.016952989622950554, + 0.02425430528819561, + 0.00919093657284975, + 0.01371837966144085, + 0.010534808039665222, + 0.00793108157813549, + 0.0074106426909565926, + 0.0064538330771028996, + 0.011671431362628937, + 0.0073903207667171955, + 0.0077324528247118, + 0.004786758683621883, + 0.0083591528236866, + 0.00792001560330391, + 0.034330155700445175, + 0.02445785328745842, + 0.03325279429554939, + 0.024238580837845802, + 0.03324892371892929, + 0.028726577758789062, + 0.03084774687886238, + 0.03562939912080765, + 0.017832346260547638, + 0.024345139041543007, + 0.017937088385224342, + 0.016870955005288124, + 0.039631184190511703, + 0.018117710947990417, + 0.012252302840352058, + 0.02013055607676506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.04868316650390625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42894846200942993 + ], + "opposite_logprob_tokens": [ + -1.0539484024047852 + ], + "logprob_reward": -0.42894846200942993, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42894846200942993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09468869864940643, + 0.1432909518480301, + 0.041354164481163025, + 0.08010721951723099, + 0.03747771307826042, + 0.06087981164455414, + 0.1528252363204956, + 0.06869655102491379, + 0.08839797228574753, + 0.23228169977664948 + ], + "credit_map": [ + 0.01185890194028616, + 0.01589111052453518, + 0.013508906587958336, + 0.010464032180607319, + 0.0077340067364275455, + 0.006227326579391956, + 0.011132811196148396, + 0.01787160150706768, + 0.015769679099321365, + 0.009818359278142452, + 0.008134185336530209, + 0.00648391293361783, + 0.011964108794927597, + 0.03460114821791649, + 0.03295257315039635, + 0.023566996678709984, + 0.002960139885544777, + 0.0023210057988762856, + 0.003029919695109129, + 0.00424167001619935, + 0.003762018634006381, + 0.010679329745471478, + 0.008957523852586746, + 0.0054025547578930855, + 0.0056564765982329845, + 0.004467736929655075, + 0.00765089550986886, + 0.00483738724142313, + 0.00583534175530076, + 0.00738320080563426, + 0.02238902449607849, + 0.02188715524971485, + 0.00301811913959682, + 0.0026113102212548256, + 0.00222314172424376, + 0.005425121169537306, + 0.006055379752069712, + 0.007540805730968714, + 0.0038393645081669092, + 0.00676446920260787, + 0.009856480173766613, + 0.010190918110311031, + 0.004222184885293245, + 0.008089400827884674, + 0.004523714073002338, + 0.005850909743458033, + 0.0072318920865654945, + 0.010914315469563007, + 0.018195606768131256, + 0.021841708570718765, + 0.02249716781079769, + 0.03691043332219124, + 0.01133769005537033, + 0.014615124091506004, + 0.014749553054571152, + 0.012677943333983421, + 0.009236849844455719, + 0.00846161786466837, + 0.013439432717859745, + 0.00693912198767066, + 0.007203029468655586, + 0.006511296145617962, + 0.005917233880609274, + 0.010987965390086174, + 0.010681171901524067, + 0.010658334009349346, + 0.012809001840651035, + 0.007293141447007656, + 0.015051305294036865, + 0.010727718472480774, + 0.008477929048240185, + 0.012699373997747898, + 0.013499048538506031, + 0.08238614350557327, + 0.011686681769788265, + 0.0106661943718791, + 0.07856928557157516, + 0.011695458553731441, + 0.009217163547873497, + 0.01456171553581953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.42894846200942993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a precarious, handheld, first-person perspective, the viewer is positioned within an ancient, ornate palace courtyard, where the ground suddenly trembles and heaves with violent force. The scene is bathed in midday sunlight, which casts sharp, deep shadows across the surrounding marble columns and arched colonnades. What begins as a low, guttural rumble gradually escalates into a piercing vibration, causing the camera to tilt downward. The view focuses on the floor, where intricate mosaic tiles, featuring geometric patterns of deep lapis lazuli, terracotta, and polished ivory, begin to fracture. A network of fine hairline cracks rapidly spiderweb across the polished surface, before the tiles buckle upward and part with a series of sharp, thrilling resonations. Immediately afterwards, a deep surface fissure opens, decisively cutting through the stone floor, revealing a dark, jagged fissure in the raw earth beneath. The entire catastrophic event is captured in a surreal and clean style, rendering every shattered fragment and rising dust mote with razor-sharp photographic detail, emphasizing the severe, irreversible damage inflicted on the beautiful historic structure and creating an atmosphere of sudden, terrifying destruction.", + "seed": 8722, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_6896_seed_8722_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": -0.12220733687281608, + "vlm_logprob": -0.12220733687281608, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.12220733687281608, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.41884395480155945 + ], + "opposite_logprob_tokens": [ + -1.1688439846038818 + ], + "logprob_reward": -0.41884395480155945, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.41884395480155945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18665753304958344, + 0.057592179626226425, + 0.04650372266769409, + 0.20430395007133484, + 0.08955716341733932, + 0.053092341870069504, + 0.0451672226190567, + 0.03288022056221962, + 0.2518085241317749, + 0.03243711218237877 + ], + "credit_map": [ + 0.009298311546444893, + 0.0588541254401207, + 0.02123590186238289, + 0.00835518166422844, + 0.012864633463323116, + 0.05437745153903961, + 0.012448376975953579, + 0.00922354869544506, + 0.007924427278339863, + 0.007555798627436161, + 0.01115594245493412, + 0.009164531715214252, + 0.006432292517274618, + 0.005016116425395012, + 0.005880118813365698, + 0.004462950862944126, + 0.00774884270504117, + 0.00435571139678359, + 0.004072129260748625, + 0.004972473252564669, + 0.00903201475739479, + 0.006925566121935844, + 0.0038495296612381935, + 0.005547457840293646, + 0.014456035569310188, + 0.03171404078602791, + 0.014221735298633575, + 0.02065122313797474, + 0.014295305125415325, + 0.037300992757081985, + 0.035234153270721436, + 0.03643046319484711, + 0.006662359461188316, + 0.007443176582455635, + 0.0078801941126585, + 0.010575606487691402, + 0.008254471234977245, + 0.011235914193093777, + 0.017500052228569984, + 0.02000538632273674, + 0.003515868680551648, + 0.005426853895187378, + 0.006448694504797459, + 0.010950655676424503, + 0.00480929110199213, + 0.0045530227944254875, + 0.010949366725981236, + 0.006438587326556444, + 0.003699949476867914, + 0.00641104718670249, + 0.0064034671522676945, + 0.010037564672529697, + 0.0034386757761240005, + 0.004003075882792473, + 0.005727481096982956, + 0.005445958115160465, + 0.004882829263806343, + 0.003729737363755703, + 0.003916648682206869, + 0.004201395437121391, + 0.004179683513939381, + 0.004626478999853134, + 0.00393645791336894, + 0.003406988922506571, + 0.01129095908254385, + 0.013103308156132698, + 0.008129305206239223, + 0.01508916262537241, + 0.030060352757573128, + 0.06216541677713394, + 0.046668410301208496, + 0.06530162692070007, + 0.0033969657961279154, + 0.004802519455552101, + 0.00419565849006176, + 0.0032475008629262447, + 0.003724522888660431, + 0.0037645259872078896, + 0.006256979890167713, + 0.0030484397429972887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.41884395480155945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800530970096588 + ], + "opposite_logprob_tokens": [ + -4.634800434112549 + ], + "logprob_reward": -0.009800530970096588, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800530970096588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1823475956916809, + 0.22289830446243286, + 0.1044398620724678, + 0.09724092483520508, + 0.08981780707836151, + 0.06479906290769577, + 0.06006966531276703, + 0.05858388543128967, + 0.04071209579706192, + 0.07909085601568222 + ], + "credit_map": [ + 0.007993427105247974, + 0.028784045949578285, + 0.032469943165779114, + 0.009495535865426064, + 0.012977611273527145, + 0.06549407541751862, + 0.014775431714951992, + 0.010357514023780823, + 0.013430599123239517, + 0.012617841362953186, + 0.08204586058855057, + 0.0545508936047554, + 0.015624403022229671, + 0.009121201001107693, + 0.02953553944826126, + 0.005971971899271011, + 0.018577584996819496, + 0.00807140301913023, + 0.008407026529312134, + 0.013371079228818417, + 0.021929951384663582, + 0.01629456877708435, + 0.007066613063216209, + 0.010721627622842789, + 0.006113259121775627, + 0.012161067686975002, + 0.011034960858523846, + 0.012903504073619843, + 0.007915383204817772, + 0.01823521964251995, + 0.013761327601969242, + 0.015116200782358646, + 0.006190185900777578, + 0.005879753734916449, + 0.007094293367117643, + 0.013550841249525547, + 0.008820552378892899, + 0.008235827088356018, + 0.025902237743139267, + 0.014144111424684525, + 0.005026209633797407, + 0.0066133481450378895, + 0.009890870191156864, + 0.009261168539524078, + 0.006059695966541767, + 0.006277038715779781, + 0.012692099437117577, + 0.008978627622127533, + 0.00573250325396657, + 0.00658212136477232, + 0.005910574924200773, + 0.008939535357058048, + 0.010831598192453384, + 0.010042516514658928, + 0.005614512134343386, + 0.006416305433958769, + 0.008127031847834587, + 0.006053585093468428, + 0.009437589906156063, + 0.006840228568762541, + 0.006129508838057518, + 0.005857066251337528, + 0.008075150661170483, + 0.008063724264502525, + 0.0028408875223249197, + 0.003291103756055236, + 0.004018828272819519, + 0.0116416085511446, + 0.002842564834281802, + 0.003005683422088623, + 0.005055323243141174, + 0.00801609642803669, + 0.009735346771776676, + 0.008231144398450851, + 0.013607022352516651, + 0.009768171235918999, + 0.005486463196575642, + 0.009136849083006382, + 0.008575084619224072, + 0.01455078087747097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.009800530970096588 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.15679119527339935, + 0.1655663102865219, + 0.02682896889746189, + 0.08378966897726059, + 0.06361377984285355, + 0.0807640329003334, + 0.055635806173086166, + 0.061528775840997696, + 0.16022317111492157, + 0.14525826275348663 + ], + "credit_map": [ + 0.008015734143555164, + 0.0485437773168087, + 0.016412116587162018, + 0.006733571644872427, + 0.01245407946407795, + 0.0502912700176239, + 0.0068956720642745495, + 0.00744497450068593, + 0.0091853067278862, + 0.011905333958566189, + 0.06597305834293365, + 0.03722139447927475, + 0.009846502915024757, + 0.005144222639501095, + 0.022468876093626022, + 0.0038216046523302794, + 0.00433899974450469, + 0.0023976503871381283, + 0.0029566334560513496, + 0.0037589508574455976, + 0.004564131144434214, + 0.003727926639840007, + 0.002212602412328124, + 0.002872074255719781, + 0.0042868913151323795, + 0.008264178410172462, + 0.007452052552253008, + 0.00794333964586258, + 0.0054388041608035564, + 0.011481761932373047, + 0.01976666785776615, + 0.019155973568558693, + 0.004600544460117817, + 0.004345794208347797, + 0.006059554405510426, + 0.010632083751261234, + 0.006953872740268707, + 0.006306287366896868, + 0.016256941482424736, + 0.008458695374429226, + 0.005084254313260317, + 0.008055263198912144, + 0.00885266624391079, + 0.012511755339801311, + 0.00909434724599123, + 0.0074827647767961025, + 0.01780587248504162, + 0.011877108365297318, + 0.006307247560471296, + 0.006841153837740421, + 0.006795985158532858, + 0.009886127896606922, + 0.007266581524163485, + 0.007679914124310017, + 0.004992576315999031, + 0.005866219289600849, + 0.007593793794512749, + 0.0057265241630375385, + 0.010081415064632893, + 0.010210111737251282, + 0.005684969946742058, + 0.0041084433905780315, + 0.00932912528514862, + 0.008794396184384823, + 0.011212337762117386, + 0.011327996850013733, + 0.011052656918764114, + 0.045692987740039825, + 0.016628094017505646, + 0.01954716444015503, + 0.02889985591173172, + 0.015862081199884415, + 0.019202839583158493, + 0.016964679583907127, + 0.023700745776295662, + 0.01810416392982006, + 0.009560728445649147, + 0.016006221994757652, + 0.014219352044165134, + 0.02749953418970108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 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 camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362799108028412 + ], + "opposite_logprob_tokens": [ + -3.651362895965576 + ], + "logprob_reward": -0.026362799108028412, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362799108028412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0791563168168068, + 0.10746932029724121, + 0.03537244349718094, + 0.1374947428703308, + 0.09182420372962952, + 0.05965026840567589, + 0.07101795822381973, + 0.07306909561157227, + 0.10118360817432404, + 0.24376200139522552 + ], + "credit_map": [ + 0.005643357057124376, + 0.019532347097992897, + 0.02035951055586338, + 0.00570671958848834, + 0.005411393009126186, + 0.007521425373852253, + 0.009839273057878017, + 0.005142296198755503, + 0.0071107433177530766, + 0.006423909682780504, + 0.03882691264152527, + 0.026698054745793343, + 0.007353853899985552, + 0.0058659035712480545, + 0.011847302317619324, + 0.003342635929584503, + 0.006016621366143227, + 0.002844200935214758, + 0.0030561364255845547, + 0.0035244461614638567, + 0.006325142458081245, + 0.0062761977314949036, + 0.0038073041941970587, + 0.0035223974846303463, + 0.006238030269742012, + 0.011807890608906746, + 0.010642745532095432, + 0.013197286985814571, + 0.007653574924916029, + 0.013716290704905987, + 0.04493416100740433, + 0.029304763302206993, + 0.006154184695333242, + 0.005991261452436447, + 0.008184440433979034, + 0.014116664417088032, + 0.007987229153513908, + 0.008217686787247658, + 0.02686746045947075, + 0.014305275864899158, + 0.004355527460575104, + 0.005742300301790237, + 0.008712077513337135, + 0.009065861813724041, + 0.005649454426020384, + 0.005204461049288511, + 0.012015522457659245, + 0.008905063383281231, + 0.007376058492809534, + 0.008118958212435246, + 0.008690830320119858, + 0.013932266272604465, + 0.008558092638850212, + 0.008232494816184044, + 0.008292538113892078, + 0.007816717028617859, + 0.012300264090299606, + 0.007683405186980963, + 0.00960011687129736, + 0.008741144090890884, + 0.00939095951616764, + 0.00840004626661539, + 0.009491031058132648, + 0.0074621238745749, + 0.0078010158613324165, + 0.006387725938111544, + 0.006950201466679573, + 0.01804216578602791, + 0.010180611163377762, + 0.023194661363959312, + 0.011477574706077576, + 0.017149651423096657, + 0.025225283578038216, + 0.02752043679356575, + 0.04088537022471428, + 0.025263048708438873, + 0.022028915584087372, + 0.023494403809309006, + 0.04410812258720398, + 0.035236429423093796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.026362799108028412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294224858283997 + ], + "opposite_logprob_tokens": [ + -2.2379422187805176 + ], + "logprob_reward": -0.11294224858283997, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294224858283997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29154810309410095, + 0.0693056508898735, + 0.036312274634838104, + 0.11921340972185135, + 0.04369224235415459, + 0.027070283889770508, + 0.17407077550888062, + 0.0699729397892952, + 0.10242713987827301, + 0.06638715416193008 + ], + "credit_map": [ + 0.012690714560449123, + 0.06727688014507294, + 0.0370650552213192, + 0.01153874583542347, + 0.01998876966536045, + 0.11412861943244934, + 0.013299280777573586, + 0.015560026280581951, + 0.006246792618185282, + 0.006328684743493795, + 0.02381046675145626, + 0.011699854396283627, + 0.005117592867463827, + 0.004149198066443205, + 0.009169619530439377, + 0.0027834458742290735, + 0.006506350357085466, + 0.002526563825085759, + 0.002256498672068119, + 0.004278908018022776, + 0.00665296008810401, + 0.005867901258170605, + 0.002993183908984065, + 0.005229907110333443, + 0.009170949459075928, + 0.014914843253791332, + 0.01209369208663702, + 0.014417361468076706, + 0.012457754462957382, + 0.02027367241680622, + 0.023639721795916557, + 0.012245415709912777, + 0.004546197596937418, + 0.004218118265271187, + 0.006840595975518227, + 0.0057881614193320274, + 0.003572257235646248, + 0.005187426693737507, + 0.008229371160268784, + 0.0053101107478141785, + 0.0016564478864893317, + 0.002877708524465561, + 0.0034090718254446983, + 0.005082006100565195, + 0.0020464302506297827, + 0.0025649648159742355, + 0.0058174231089651585, + 0.003616232657805085, + 0.024352319538593292, + 0.027209768071770668, + 0.020912857726216316, + 0.022479090839624405, + 0.022393452003598213, + 0.020316673442721367, + 0.018332060426473618, + 0.018074562773108482, + 0.010338311083614826, + 0.010853560641407967, + 0.009310360066592693, + 0.010042637586593628, + 0.007665398996323347, + 0.006145184393972158, + 0.009086406789720058, + 0.006531085819005966, + 0.00509015005081892, + 0.00636264169588685, + 0.006383741274476051, + 0.020636627450585365, + 0.009895875118672848, + 0.020843155682086945, + 0.011073717847466469, + 0.022141234949231148, + 0.006086281035095453, + 0.01022916566580534, + 0.007892253808677197, + 0.009063026867806911, + 0.006846953649073839, + 0.00695938803255558, + 0.010897899977862835, + 0.008412179537117481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.11294224858283997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a precarious, handheld, first-person perspective, the viewer is positioned within an ancient, ornate palace courtyard, where the ground suddenly trembles and heaves with violent force. The scene is bathed in midday sunlight, which casts sharp, deep shadows across the surrounding marble columns and arched colonnades. What begins as a low, guttural rumble gradually escalates into a piercing vibration, causing the camera to tilt downward. The view focuses on the floor, where intricate mosaic tiles, featuring geometric patterns of deep lapis lazuli, terracotta, and polished ivory, begin to fracture. A network of fine hairline cracks rapidly spiderweb across the polished surface, before the tiles buckle upward and part with a series of sharp, thrilling resonations. Immediately afterwards, a deep surface fissure opens, decisively cutting through the stone floor, revealing a dark, jagged fissure in the raw earth beneath. The entire catastrophic event is captured in a surreal and clean style, rendering every shattered fragment and rising dust mote with razor-sharp photographic detail, emphasizing the severe, irreversible damage inflicted on the beautiful historic structure and creating an atmosphere of sudden, terrifying destruction.", + "seed": 8723, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_6896_seed_8723_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": -0.13312729652971028, + "vlm_logprob": -0.13312729652971028, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.13312729652971028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42381715774536133 + ], + "opposite_logprob_tokens": [ + -1.1738171577453613 + ], + "logprob_reward": -0.42381715774536133, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.42381715774536133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10256212949752808, + 0.061960089951753616, + 0.19331200420856476, + 0.10882217437028885, + 0.05709325149655342, + 0.06999832391738892, + 0.04651865363121033, + 0.10869474709033966, + 0.14956612884998322, + 0.10147251188755035 + ], + "credit_map": [ + 0.010094617493450642, + 0.01363748125731945, + 0.015360290184617043, + 0.019460726529359818, + 0.00990546029061079, + 0.015008389949798584, + 0.01100468821823597, + 0.008090470917522907, + 0.006341270636767149, + 0.008955750614404678, + 0.006256613414734602, + 0.009683271870017052, + 0.006076835561543703, + 0.0072899493388831615, + 0.01148536242544651, + 0.00587103608995676, + 0.01201543491333723, + 0.015147187747061253, + 0.006191619206219912, + 0.005788089707493782, + 0.08008749783039093, + 0.0316692590713501, + 0.006210833787918091, + 0.0362020768225193, + 0.009753778576850891, + 0.008960423991084099, + 0.008708873763680458, + 0.015779953449964523, + 0.010671229101717472, + 0.014592388644814491, + 0.025526855140924454, + 0.01482867356389761, + 0.008420530706644058, + 0.019268544390797615, + 0.003704247297719121, + 0.0032176098320633173, + 0.011990583501756191, + 0.005061632953584194, + 0.0032555884681642056, + 0.0021745122503489256, + 0.005611881613731384, + 0.006577483378350735, + 0.013983196578919888, + 0.009940583258867264, + 0.0077870506793260574, + 0.005920022260397673, + 0.01144352462142706, + 0.008734584785997868, + 0.005217551253736019, + 0.007139754481613636, + 0.00585442828014493, + 0.00530814565718174, + 0.004588665906339884, + 0.005868127569556236, + 0.008127115666866302, + 0.004414869472384453, + 0.008650621399283409, + 0.008629814721643925, + 0.018817652016878128, + 0.021219972521066666, + 0.006261136382818222, + 0.007728481665253639, + 0.023709269240498543, + 0.013677801936864853, + 0.015228389762341976, + 0.018400654196739197, + 0.017164794728159904, + 0.016431814059615135, + 0.02429967187345028, + 0.01755390875041485, + 0.026849692687392235, + 0.013637198135256767, + 0.00900057889521122, + 0.01125103048980236, + 0.011470462195575237, + 0.014893114566802979, + 0.010762695223093033, + 0.0185294970870018, + 0.013655679300427437, + 0.011909454129636288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.42381715774536133 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014213516376912594 + ], + "opposite_logprob_tokens": [ + -4.264213562011719 + ], + "logprob_reward": -0.014213516376912594, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014213516376912594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18394042551517487, + 0.07478411495685577, + 0.08632861077785492, + 0.053011875599622726, + 0.08880124241113663, + 0.10630658268928528, + 0.058307211846113205, + 0.03361959010362625, + 0.23236648738384247, + 0.08253379911184311 + ], + "credit_map": [ + 0.020597681403160095, + 0.015465592965483665, + 0.031134068965911865, + 0.02815902978181839, + 0.01899830251932144, + 0.02560381218791008, + 0.02600165456533432, + 0.01798029989004135, + 0.006802351679652929, + 0.008163263089954853, + 0.011267813853919506, + 0.013527312316000462, + 0.006252884864807129, + 0.013075029477477074, + 0.007327153347432613, + 0.008368304930627346, + 0.006022550165653229, + 0.006209794897586107, + 0.005193162243813276, + 0.005820233840495348, + 0.03212960809469223, + 0.014612961560487747, + 0.003340129042044282, + 0.013000172562897205, + 0.004323085304349661, + 0.004864698741585016, + 0.006721312180161476, + 0.005401843227446079, + 0.006747063715010881, + 0.009633585810661316, + 0.008137467317283154, + 0.007182822097092867, + 0.011035647243261337, + 0.03082386404275894, + 0.005057253409177065, + 0.00512261176481843, + 0.019969971850514412, + 0.006690487265586853, + 0.0055437698028981686, + 0.004557635169476271, + 0.009205585345625877, + 0.007414495572447777, + 0.015609909780323505, + 0.016886800527572632, + 0.010809077881276608, + 0.008982574567198753, + 0.021328303962945938, + 0.01606983132660389, + 0.005316019989550114, + 0.006803775671869516, + 0.007068126928061247, + 0.008601684123277664, + 0.004489191807806492, + 0.006912856362760067, + 0.011422798037528992, + 0.007692760322242975, + 0.0029967441223561764, + 0.003619458759203553, + 0.0047560748644173145, + 0.004953980445861816, + 0.002717656549066305, + 0.0034550109412521124, + 0.007205995265394449, + 0.003914668690413237, + 0.01997537538409233, + 0.030662652105093002, + 0.02738979458808899, + 0.03440287709236145, + 0.033157140016555786, + 0.028653794899582863, + 0.027026044204831123, + 0.03109881654381752, + 0.00851453561335802, + 0.009164726361632347, + 0.00899890623986721, + 0.013838596642017365, + 0.009171436540782452, + 0.01255061849951744, + 0.011730087921023369, + 0.008564895018935204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.014213516376912594 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498713254928589 + ], + "opposite_logprob_tokens": [ + -2.9299871921539307 + ], + "logprob_reward": -0.05498713254928589, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498713254928589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06699604541063309, + 0.28810253739356995, + 0.036535751074552536, + 0.18621979653835297, + 0.10748802125453949, + 0.05226204916834831, + 0.05262839049100876, + 0.07754945009946823, + 0.10627587139606476, + 0.02594209648668766 + ], + "credit_map": [ + 0.007545890752226114, + 0.007474916521459818, + 0.008472079411149025, + 0.011680694296956062, + 0.0075899651274085045, + 0.009629311040043831, + 0.008671825751662254, + 0.005931357853114605, + 0.03244565427303314, + 0.03404344990849495, + 0.07122847437858582, + 0.0537126362323761, + 0.0379224568605423, + 0.02449868805706501, + 0.016926001757383347, + 0.01732517033815384, + 0.0026746673975139856, + 0.0029552236665040255, + 0.004186795558780432, + 0.003184386994689703, + 0.010722573846578598, + 0.006667440291494131, + 0.00291319377720356, + 0.0032314674463123083, + 0.015281346626579762, + 0.015454355627298355, + 0.024111047387123108, + 0.02801177278161049, + 0.01841062679886818, + 0.023385439068078995, + 0.04303295537829399, + 0.018532266840338707, + 0.009301970712840557, + 0.027751578018069267, + 0.006707590539008379, + 0.012348784133791924, + 0.018641257658600807, + 0.009847358800470829, + 0.014888954348862171, + 0.008000527508556843, + 0.007389011327177286, + 0.004528924357146025, + 0.008163884282112122, + 0.006449453067034483, + 0.0058052632957696915, + 0.005206447094678879, + 0.009237051010131836, + 0.005482011940330267, + 0.004904622212052345, + 0.005708631593734026, + 0.006933572702109814, + 0.00825630221515894, + 0.0037498425226658583, + 0.004707010462880135, + 0.010467023588716984, + 0.007901385426521301, + 0.005856069270521402, + 0.0067826746962964535, + 0.013313479721546173, + 0.015615773387253284, + 0.0036965201143175364, + 0.004517816007137299, + 0.01843850128352642, + 0.00932861678302288, + 0.009073221124708652, + 0.012551077641546726, + 0.013606802560389042, + 0.017425281926989555, + 0.013651746325194836, + 0.012469395063817501, + 0.013267002999782562, + 0.014231340028345585, + 0.0018923950847238302, + 0.0024950788356363773, + 0.0035786000080406666, + 0.003831947222352028, + 0.0030306652188301086, + 0.002443142933771014, + 0.004048449918627739, + 0.004621817264705896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.05498713254928589 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029817182570695877 + ], + "opposite_logprob_tokens": [ + -3.5298171043395996 + ], + "logprob_reward": -0.029817182570695877, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029817182570695877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04490736871957779, + 0.19099971652030945, + 0.05343304201960564, + 0.07401890307664871, + 0.023301299661397934, + 0.054129790514707565, + 0.028051832690835, + 0.05228981748223305, + 0.32039904594421387, + 0.15846921503543854 + ], + "credit_map": [ + 0.004769755993038416, + 0.004867409821599722, + 0.005963974632322788, + 0.004822766408324242, + 0.00636742589995265, + 0.006761558819562197, + 0.005923797842115164, + 0.0054306816309690475, + 0.027862632647156715, + 0.019857823848724365, + 0.05843990668654442, + 0.03489897400140762, + 0.019740819931030273, + 0.012415789067745209, + 0.008664156310260296, + 0.009119605645537376, + 0.006148852873593569, + 0.003622581949457526, + 0.009130534715950489, + 0.00895603559911251, + 0.005152531899511814, + 0.006299702916294336, + 0.005759299732744694, + 0.008363500237464905, + 0.005901367869228125, + 0.005414736922830343, + 0.008380105718970299, + 0.006758280098438263, + 0.005728201474994421, + 0.018025437369942665, + 0.009326095692813396, + 0.014484675601124763, + 0.002464037388563156, + 0.0025387245696038008, + 0.0029482028912752867, + 0.003789669368416071, + 0.002896298421546817, + 0.002318484243005514, + 0.0042227208614349365, + 0.002123161917552352, + 0.00670958124101162, + 0.004462763201445341, + 0.010228986851871014, + 0.007271276321262121, + 0.0046927109360694885, + 0.004286706447601318, + 0.009857221506536007, + 0.006620544474571943, + 0.002952447161078453, + 0.0028976525645703077, + 0.003668184857815504, + 0.004279563203454018, + 0.0017726562218740582, + 0.002881648950278759, + 0.0057874638587236404, + 0.0038122155237942934, + 0.005072636995464563, + 0.004733780398964882, + 0.008780835196375847, + 0.0101268095895648, + 0.002966096857562661, + 0.003479932900518179, + 0.012060011737048626, + 0.005069711711257696, + 0.023838844150304794, + 0.038242775946855545, + 0.042210329324007034, + 0.059079695492982864, + 0.03262312337756157, + 0.03691427782177925, + 0.030388664454221725, + 0.05710133910179138, + 0.01328355073928833, + 0.02116132341325283, + 0.01722254417836666, + 0.02774955704808235, + 0.018338441848754883, + 0.01687789335846901, + 0.025865644216537476, + 0.017970262095332146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.029817182570695877 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective handheld and first-person?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in an ancient, ornate palace courtyard with marble columns and arched colonnades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the floor tiles feature geometric patterns of deep lapis lazuli, terracotta, and polished ivory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the floor fracture with spiderweb cracks, buckle upward, and open a deep fissure revealing raw earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal and clean style with razor-sharp photographic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280149340629578 + ], + "opposite_logprob_tokens": [ + -2.017801523208618 + ], + "logprob_reward": -0.14280149340629578, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280149340629578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1297163963317871, + 0.08543544262647629, + 0.16421939432621002, + 0.08109865337610245, + 0.10226571559906006, + 0.06309352070093155, + 0.05544180050492287, + 0.04545453563332558, + 0.15786407887935638, + 0.1154104471206665 + ], + "credit_map": [ + 0.013498404063284397, + 0.012535549700260162, + 0.024260567501187325, + 0.02849012427031994, + 0.011490209959447384, + 0.013796615414321423, + 0.01457980927079916, + 0.011065111495554447, + 0.00950565468519926, + 0.008280495181679726, + 0.010228870436549187, + 0.01290020439773798, + 0.011939932592213154, + 0.013102634809911251, + 0.012064085341989994, + 0.007413562852889299, + 0.010688732378184795, + 0.010655411519110203, + 0.010861474089324474, + 0.010598239488899708, + 0.059984318912029266, + 0.02529236674308777, + 0.007514661177992821, + 0.02862420491874218, + 0.00672113848850131, + 0.007633616682142019, + 0.009788457304239273, + 0.01277912687510252, + 0.007871692068874836, + 0.011369533836841583, + 0.016345985233783722, + 0.008589103817939758, + 0.010372870601713657, + 0.033285677433013916, + 0.010858853347599506, + 0.0067260246723890305, + 0.022652678191661835, + 0.008714419789612293, + 0.005259277764707804, + 0.004395906813442707, + 0.006847671698778868, + 0.006797384936362505, + 0.012566710822284222, + 0.00862579233944416, + 0.004839749541133642, + 0.0067875864915549755, + 0.010351670905947685, + 0.006276953965425491, + 0.003999028354883194, + 0.007161244750022888, + 0.008816475979983807, + 0.008801723830401897, + 0.0036092293448746204, + 0.006487686652690172, + 0.009890605695545673, + 0.006675802171230316, + 0.0036210466641932726, + 0.0031979570630937815, + 0.0065607549622654915, + 0.006424555554986, + 0.002425452461466193, + 0.003463843371719122, + 0.011817666701972485, + 0.007943259552121162, + 0.01298284251242876, + 0.02062550187110901, + 0.017628628760576248, + 0.01889652945101261, + 0.02786540426313877, + 0.017884353175759315, + 0.017405910417437553, + 0.024574916809797287, + 0.013828642666339874, + 0.015630384907126427, + 0.01440874207764864, + 0.01818404160439968, + 0.013127707876265049, + 0.015991507098078728, + 0.01092703640460968, + 0.013312389142811298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.14280149340629578 + } + ], + "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/0000025/group_8/index_tensor([891, 891])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_8/index_tensor([891, 891])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2ae51e08f9a83ddbf18e99149b5dab046a7eacc4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_8/index_tensor([891, 891])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle panning shot, moving steadily to the right, closely follows a bright yellow electric forklift as it smoothly navigates a narrow aisle between two towering, industrial-grade metal shelves. These massive shelves stretch from the smooth gray concrete floor to the high warehouse ceiling, stacked with neatly organized cartons and wooden pallets tightly wrapped in transparent plastic. The forklift bears the slightest traces of industrial wear, its metal forks held slightly raised for safe maneuvering. The rotating orange warning light on its roof flashes rhythmically, casting dynamic circles of amber light across the cold metal shelves and the shimmering floor. The forklift moves forward steadily and deliberately, its sturdy black rubber tires turning silently on the ground, accompanied only by the gentle, constant hum of the motor. The entire scene is illuminated evenly by the cool glow of rows of white fluorescent tubes on the ceiling, contrasting sharply with the warm, pulsing glow of the forklift's warning lights, creating an atmosphere of orderly industrial operation that is both quiet and full of potential power.", + "seed": 33472, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_8/index_891_seed_33472_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": -0.04434658158570528, + "vlm_logprob": -0.04434658158570528, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04434658158570528, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.012904939241707325 + ], + "opposite_logprob_tokens": [ + -5.262905120849609 + ], + "logprob_reward": -0.012904939241707325, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.012904939241707325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18972715735435486, + 0.07399734854698181, + 0.0565774068236351, + 0.028631584718823433, + 0.037959616631269455, + 0.035118646919727325, + 0.24062016606330872, + 0.21454443037509918, + 0.06989433616399765, + 0.052929311990737915 + ], + "credit_map": [ + 0.029376301914453506, + 0.02305689826607704, + 0.0248915646225214, + 0.03352327272295952, + 0.01600402034819126, + 0.022441208362579346, + 0.02045358344912529, + 0.01998029462993145, + 0.014739726670086384, + 0.00819226261228323, + 0.014169024303555489, + 0.009654159657657146, + 0.00618181424215436, + 0.006389440502971411, + 0.008973336778581142, + 0.005697577726095915, + 0.008428308181464672, + 0.007328657899051905, + 0.00702855596318841, + 0.0060203117318451405, + 0.00574483722448349, + 0.006738217547535896, + 0.00791451521217823, + 0.007373997010290623, + 0.004536093678325415, + 0.002691275905817747, + 0.0034803664311766624, + 0.0030428599566221237, + 0.0033725860994309187, + 0.004788884427398443, + 0.004299923311918974, + 0.0024195904843509197, + 0.006556919310241938, + 0.003507613902911544, + 0.004985468462109566, + 0.004796827677637339, + 0.004252742975950241, + 0.006201723124831915, + 0.004241958260536194, + 0.0034163608215749264, + 0.006701185368001461, + 0.0033621033653616905, + 0.005463274661451578, + 0.004815435968339443, + 0.003470818977802992, + 0.004776699934154749, + 0.00331489322707057, + 0.003214230528101325, + 0.050503965467214584, + 0.022874778136610985, + 0.03945933282375336, + 0.03529923036694527, + 0.02233514003455639, + 0.027946317568421364, + 0.021636568009853363, + 0.020564822480082512, + 0.025171108543872833, + 0.023887362331151962, + 0.01762075535953045, + 0.023547712713479996, + 0.022565867751836777, + 0.05141449347138405, + 0.021846430376172066, + 0.028490686789155006, + 0.008437472395598888, + 0.008595726452767849, + 0.010013522580265999, + 0.01113985013216734, + 0.009062484838068485, + 0.008552885614335537, + 0.008150249719619751, + 0.00594213604927063, + 0.005332014989107847, + 0.006147240754216909, + 0.007017804309725761, + 0.009616797789931297, + 0.0047973995096981525, + 0.009899101220071316, + 0.005593020934611559, + 0.0045259264297783375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.012904939241707325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.04867726191878319 + ], + "opposite_logprob_tokens": [ + -3.0486772060394287 + ], + "logprob_reward": -0.04867726191878319, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867726191878319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1978732943534851, + 0.09936648607254028, + 0.054103292524814606, + 0.04000534489750862, + 0.07364526391029358, + 0.055534686893224716, + 0.13939601182937622, + 0.19286265969276428, + 0.07658279687166214, + 0.07063013315200806 + ], + "credit_map": [ + 0.025058608502149582, + 0.027745360508561134, + 0.023853236809372902, + 0.047444168478250504, + 0.017702169716358185, + 0.020430922508239746, + 0.016387738287448883, + 0.019251087680459023, + 0.011139401234686375, + 0.009922338649630547, + 0.013441495597362518, + 0.008549626916646957, + 0.011257932521402836, + 0.021405141800642014, + 0.014454861171543598, + 0.009195692837238312, + 0.005813883617520332, + 0.005054127890616655, + 0.009147755801677704, + 0.006210299674421549, + 0.005936739034950733, + 0.009145650081336498, + 0.00774145033210516, + 0.005053384695202112, + 0.005843969527631998, + 0.0035721363965421915, + 0.006202197168022394, + 0.005055687855929136, + 0.004231040831655264, + 0.006897222250699997, + 0.003938333131372929, + 0.004264759365469217, + 0.006935661658644676, + 0.008893641643226147, + 0.006961306557059288, + 0.007173504680395126, + 0.014987437054514885, + 0.012559914030134678, + 0.008999494835734367, + 0.007134308107197285, + 0.007696256507188082, + 0.005597847513854504, + 0.00707207852974534, + 0.006452532950788736, + 0.005780654028058052, + 0.012497030198574066, + 0.005706882569938898, + 0.004731406457722187, + 0.014563348144292831, + 0.010851267725229263, + 0.021044274792075157, + 0.017241429537534714, + 0.013214223086833954, + 0.03503842279314995, + 0.0139928562566638, + 0.013450195081532001, + 0.022400222718715668, + 0.016750935465097427, + 0.015020024962723255, + 0.01402700413018465, + 0.01994428038597107, + 0.060276247560977936, + 0.020494647324085236, + 0.023949308320879936, + 0.00798131711781025, + 0.005796285346150398, + 0.007985069416463375, + 0.009178387932479382, + 0.0077709718607366085, + 0.02246943674981594, + 0.0076699587516486645, + 0.007731366436928511, + 0.011701211333274841, + 0.009271007031202316, + 0.008058193139731884, + 0.010360730811953545, + 0.005511626601219177, + 0.011524911969900131, + 0.006569478195160627, + 0.0076329708099365234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.04867726191878319 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.07896608859300613 + ], + "opposite_logprob_tokens": [ + -2.5789661407470703 + ], + "logprob_reward": -0.07896608859300613, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896608859300613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2195153385400772, + 0.14029507339000702, + 0.08246030658483505, + 0.040875982493162155, + 0.05511021614074707, + 0.048722025007009506, + 0.11902420967817307, + 0.08720014989376068, + 0.13818737864494324, + 0.06860940903425217 + ], + "credit_map": [ + 0.02480008266866207, + 0.017490368336439133, + 0.022985272109508514, + 0.04868912324309349, + 0.020683374255895615, + 0.019985457882285118, + 0.01944664493203163, + 0.04543498158454895, + 0.019244853407144547, + 0.010908206924796104, + 0.01636279746890068, + 0.012325228191912174, + 0.016722993925213814, + 0.03144233673810959, + 0.018653621897101402, + 0.014635018073022366, + 0.009423618204891682, + 0.005493583623319864, + 0.008743537589907646, + 0.007233963813632727, + 0.010187627747654915, + 0.019814735278487206, + 0.010601726360619068, + 0.010961503721773624, + 0.005822834558784962, + 0.003105450887233019, + 0.004549363628029823, + 0.0064332312904298306, + 0.0043515292927622795, + 0.007746885996311903, + 0.003811037167906761, + 0.005055645480751991, + 0.012241249904036522, + 0.0068682036362588406, + 0.006468074396252632, + 0.007209497969597578, + 0.006173761561512947, + 0.006241709925234318, + 0.005586458835750818, + 0.004321252927184105, + 0.00941468682140112, + 0.004979859571903944, + 0.005346149206161499, + 0.0051528639160096645, + 0.004599985666573048, + 0.010250680148601532, + 0.004497517831623554, + 0.004480278585106134, + 0.01864314079284668, + 0.011210178956389427, + 0.020250501111149788, + 0.017596516758203506, + 0.013326036743819714, + 0.014365422539412975, + 0.01375191006809473, + 0.009880483150482178, + 0.008922400884330273, + 0.005995613057166338, + 0.006935596466064453, + 0.008382230997085571, + 0.00874162558466196, + 0.02852701023221016, + 0.007705214433372021, + 0.011990446597337723, + 0.014064551331102848, + 0.012076660059392452, + 0.015703290700912476, + 0.012682529166340828, + 0.018353469669818878, + 0.032685596495866776, + 0.014162895269691944, + 0.018458368256688118, + 0.009199883788824081, + 0.007487719412893057, + 0.0069678546860814095, + 0.010840240865945816, + 0.006676557473838329, + 0.012461066246032715, + 0.009377684444189072, + 0.0055983904749155045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.07896608859300613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.043076302856206894 + ], + "opposite_logprob_tokens": [ + -3.168076276779175 + ], + "logprob_reward": -0.043076302856206894, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043076302856206894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3448379337787628, + 0.09012985974550247, + 0.085411436855793, + 0.03453409671783447, + 0.0835447758436203, + 0.049673356115818024, + 0.06039704009890556, + 0.07487434893846512, + 0.06141822785139084, + 0.11517895013093948 + ], + "credit_map": [ + 0.03926606848835945, + 0.03196857497096062, + 0.035490404814481735, + 0.10115650296211243, + 0.025437813252210617, + 0.02484227530658245, + 0.03513222932815552, + 0.05154405161738396, + 0.00980864092707634, + 0.008964295499026775, + 0.01098540797829628, + 0.00911449920386076, + 0.014917158521711826, + 0.014643521048128605, + 0.010948406532406807, + 0.010747929103672504, + 0.012627833522856236, + 0.007937856949865818, + 0.01003674603998661, + 0.00928568933159113, + 0.008681798353791237, + 0.015570142306387424, + 0.009496593847870827, + 0.011774776503443718, + 0.0046885148622095585, + 0.003105179639533162, + 0.005853378679603338, + 0.004099178593605757, + 0.0034436227288097143, + 0.005673194769769907, + 0.004561213310807943, + 0.0031098132021725178, + 0.013582752086222172, + 0.009445879608392715, + 0.015673253685235977, + 0.009646089747548103, + 0.008216436952352524, + 0.00967775471508503, + 0.010113583877682686, + 0.007189023774117231, + 0.012248590588569641, + 0.006300138309597969, + 0.00726997060701251, + 0.006828233599662781, + 0.005828870460391045, + 0.003611934371292591, + 0.0042196777649223804, + 0.0033659397158771753, + 0.013580185361206532, + 0.005032776389271021, + 0.009279792197048664, + 0.00776266772300005, + 0.0051919082179665565, + 0.00821359921246767, + 0.006339952349662781, + 0.00499616377055645, + 0.008032185956835747, + 0.006506312172859907, + 0.009590350091457367, + 0.009127883240580559, + 0.006667737849056721, + 0.02023657038807869, + 0.008238580077886581, + 0.006474726367741823, + 0.005733216647058725, + 0.004446893464773893, + 0.006923101842403412, + 0.009324993938207626, + 0.005802310071885586, + 0.014978598803281784, + 0.006037381011992693, + 0.008171729743480682, + 0.018187006935477257, + 0.020300762727856636, + 0.01558746863156557, + 0.017645912244915962, + 0.008875904604792595, + 0.010647973045706749, + 0.012242925353348255, + 0.011691001243889332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.043076302856206894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.03810831531882286 + ], + "opposite_logprob_tokens": [ + -3.2881083488464355 + ], + "logprob_reward": -0.03810831531882286, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810831531882286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3684854805469513, + 0.1205325722694397, + 0.03357673063874245, + 0.04571986570954323, + 0.08134526014328003, + 0.03596760332584381, + 0.10474081337451935, + 0.11155442893505096, + 0.028545688837766647, + 0.06953152269124985 + ], + "credit_map": [ + 0.031762078404426575, + 0.025204241275787354, + 0.03889451175928116, + 0.061097051948308945, + 0.030032722279429436, + 0.07726097106933594, + 0.05130400136113167, + 0.05292992666363716, + 0.014609945937991142, + 0.008530313149094582, + 0.013838764280080795, + 0.010215680114924908, + 0.015994474291801453, + 0.02623595856130123, + 0.015405764803290367, + 0.015701688826084137, + 0.003957749810069799, + 0.0029644037131220102, + 0.004995568189769983, + 0.003464963985607028, + 0.004359173122793436, + 0.005141911096870899, + 0.005446211900562048, + 0.003246748587116599, + 0.0045365761034190655, + 0.002943865256384015, + 0.004337009508162737, + 0.00378595688380301, + 0.004112200811505318, + 0.01525731012225151, + 0.005798475816845894, + 0.004948473069816828, + 0.010603484697639942, + 0.009439036250114441, + 0.009556868113577366, + 0.007929258979856968, + 0.013438865542411804, + 0.013870982453227043, + 0.009318561293184757, + 0.007188210729509592, + 0.0031246233265846968, + 0.0029714482370764017, + 0.006732773967087269, + 0.0037832229863852262, + 0.00375548773445189, + 0.0069938586093485355, + 0.0056425053626298904, + 0.002963688690215349, + 0.019124649465084076, + 0.00984695553779602, + 0.014300644397735596, + 0.014787077903747559, + 0.010661128908395767, + 0.016138285398483276, + 0.010189304128289223, + 0.009692779742181301, + 0.01579384319484234, + 0.013636953197419643, + 0.010806858539581299, + 0.01274939626455307, + 0.01053705159574747, + 0.022932611405849457, + 0.011390549130737782, + 0.013707183301448822, + 0.0043637691996991634, + 0.003064085030928254, + 0.0031818547286093235, + 0.003436483209952712, + 0.0025213330518454313, + 0.006578167900443077, + 0.002563256537541747, + 0.002836742205545306, + 0.007320868317037821, + 0.0068785701878368855, + 0.005237257573753595, + 0.0068241930566728115, + 0.005230908282101154, + 0.025102639570832253, + 0.005901910364627838, + 0.0070351832546293736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.03810831531882286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle panning shot, moving steadily to the right, closely follows a bright yellow electric forklift as it smoothly navigates a narrow aisle between two towering, industrial-grade metal shelves. These massive shelves stretch from the smooth gray concrete floor to the high warehouse ceiling, stacked with neatly organized cartons and wooden pallets tightly wrapped in transparent plastic. The forklift bears the slightest traces of industrial wear, its metal forks held slightly raised for safe maneuvering. The rotating orange warning light on its roof flashes rhythmically, casting dynamic circles of amber light across the cold metal shelves and the shimmering floor. The forklift moves forward steadily and deliberately, its sturdy black rubber tires turning silently on the ground, accompanied only by the gentle, constant hum of the motor. The entire scene is illuminated evenly by the cool glow of rows of white fluorescent tubes on the ceiling, contrasting sharply with the warm, pulsing glow of the forklift's warning lights, creating an atmosphere of orderly industrial operation that is both quiet and full of potential power.", + "seed": 33473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_8/index_891_seed_33473_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": -0.01655438207089901, + "vlm_logprob": -0.01655438207089901, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.01655438207089901, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.010060672648251057 + ], + "opposite_logprob_tokens": [ + -5.510060787200928 + ], + "logprob_reward": -0.010060672648251057, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010060672648251057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12414122372865677, + 0.17467160522937775, + 0.17157723009586334, + 0.07336903363466263, + 0.07024363428354263, + 0.05395471677184105, + 0.0773850753903389, + 0.043872199952602386, + 0.1382955014705658, + 0.07248974591493607 + ], + "credit_map": [ + 0.02143515832722187, + 0.014450267888605595, + 0.022317860275506973, + 0.01390816643834114, + 0.008050301112234592, + 0.011817792430520058, + 0.017704080790281296, + 0.01445759553462267, + 0.03859874978661537, + 0.014896515756845474, + 0.028292862698435783, + 0.015381609089672565, + 0.011406213976442814, + 0.024584513157606125, + 0.018653642386198044, + 0.022857505828142166, + 0.028240444138646126, + 0.01502638403326273, + 0.017039425671100616, + 0.012911495752632618, + 0.02714482694864273, + 0.027913518249988556, + 0.01823229342699051, + 0.025068849325180054, + 0.00879512820392847, + 0.008900546468794346, + 0.007430552504956722, + 0.00903669185936451, + 0.009099540300667286, + 0.008238726295530796, + 0.010403623804450035, + 0.011464232578873634, + 0.00922456569969654, + 0.011138342320919037, + 0.009019048884510994, + 0.01074174977838993, + 0.00740572577342391, + 0.008710474707186222, + 0.005849349312484264, + 0.008154385723173618, + 0.005074928514659405, + 0.0047814082354307175, + 0.006314604543149471, + 0.007796383462846279, + 0.006528643425554037, + 0.009606744162738323, + 0.006394708529114723, + 0.007457299157977104, + 0.010993133299052715, + 0.005269131623208523, + 0.006947074551135302, + 0.01481106597930193, + 0.006899525411427021, + 0.007002830971032381, + 0.006476515904068947, + 0.018985798582434654, + 0.004361058119684458, + 0.005595981143414974, + 0.00456091295927763, + 0.006997522432357073, + 0.005811756942421198, + 0.007768532726913691, + 0.003936667460948229, + 0.00483977235853672, + 0.009976057335734367, + 0.03737962618470192, + 0.011244107037782669, + 0.049513354897499084, + 0.007913803681731224, + 0.007375288289040327, + 0.008586723357439041, + 0.00630655512213707, + 0.005762822926044464, + 0.005697683896869421, + 0.01045982912182808, + 0.007962867617607117, + 0.014320739544928074, + 0.011588078923523426, + 0.010569152422249317, + 0.006128580775111914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.010060672648251057 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.012557132169604301 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012557132169604301, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557132169604301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10822770744562149, + 0.15735593438148499, + 0.15035384893417358, + 0.12868912518024445, + 0.04660211503505707, + 0.08990315347909927, + 0.07026432454586029, + 0.0702548548579216, + 0.14054790139198303, + 0.03780106455087662 + ], + "credit_map": [ + 0.018890319392085075, + 0.00982274953275919, + 0.012542907148599625, + 0.0103739183396101, + 0.010210598818957806, + 0.02244293875992298, + 0.00873348768800497, + 0.015210787765681744, + 0.04238088056445122, + 0.010985616594552994, + 0.02660266123712063, + 0.011064711026847363, + 0.011619983240962029, + 0.019610581919550896, + 0.011106549762189388, + 0.023984936997294426, + 0.01612905226647854, + 0.021592872217297554, + 0.019692312926054, + 0.016102222725749016, + 0.019587358459830284, + 0.014025690965354443, + 0.02016863040626049, + 0.023055709898471832, + 0.01671253703534603, + 0.01566004753112793, + 0.013722902163863182, + 0.01514657773077488, + 0.009999399073421955, + 0.017037605866789818, + 0.018345406278967857, + 0.022064652293920517, + 0.005383674520999193, + 0.003830421483144164, + 0.005958814173936844, + 0.007791896350681782, + 0.003651427337899804, + 0.007476348429918289, + 0.004992608446627855, + 0.007516925223171711, + 0.008422816172242165, + 0.00977256242185831, + 0.008625433780252934, + 0.014489153400063515, + 0.00875517912209034, + 0.014360623434185982, + 0.013029895722866058, + 0.012447492219507694, + 0.00631239777430892, + 0.005191847216337919, + 0.006576766725629568, + 0.006291392724961042, + 0.008353028446435928, + 0.014337553642690182, + 0.009478066116571426, + 0.013723273761570454, + 0.012542884796857834, + 0.007124563213437796, + 0.0100678950548172, + 0.008756015449762344, + 0.007654731627553701, + 0.00957436952739954, + 0.00515228183940053, + 0.009382111020386219, + 0.02408508211374283, + 0.04561339318752289, + 0.01077128667384386, + 0.01691121980547905, + 0.007606618572026491, + 0.009514250792562962, + 0.01617760770022869, + 0.009868450462818146, + 0.0027411894407123327, + 0.00409531919285655, + 0.0049926405772566795, + 0.005880511365830898, + 0.006972013507038355, + 0.004988288041204214, + 0.004723976831883192, + 0.0034071241971105337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.012557132169604301 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.029790453612804413 + ], + "opposite_logprob_tokens": [ + -3.5297904014587402 + ], + "logprob_reward": -0.029790453612804413, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790453612804413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04220499470829964, + 0.020799584686756134, + 0.08733855187892914, + 0.04242126643657684, + 0.040036819875240326, + 0.05710112676024437, + 0.051129963248968124, + 0.09345842152833939, + 0.5231448411941528, + 0.04236448183655739 + ], + "credit_map": [ + 0.0080694779753685, + 0.002964849816635251, + 0.005747494287788868, + 0.003501391736790538, + 0.0035757592413574457, + 0.008271797560155392, + 0.004071562085300684, + 0.006002660375088453, + 0.0036872257478535175, + 0.0020749589893966913, + 0.0022185638081282377, + 0.0019679348915815353, + 0.0015301610110327601, + 0.0047472091391682625, + 0.0017992827342823148, + 0.0027742476668208838, + 0.012287518940865993, + 0.008556976914405823, + 0.010910996235907078, + 0.0081761684268713, + 0.013023032806813717, + 0.009026356972754002, + 0.010189393535256386, + 0.015168110840022564, + 0.007446812931448221, + 0.006651435513049364, + 0.004671081434935331, + 0.0033809992019087076, + 0.004841295536607504, + 0.004957367666065693, + 0.005057486705482006, + 0.005414786748588085, + 0.004512254148721695, + 0.006027094088494778, + 0.0047059268690645695, + 0.006101254373788834, + 0.004371520597487688, + 0.003960573114454746, + 0.004357989877462387, + 0.006000207737088203, + 0.004440500866621733, + 0.004783037584275007, + 0.005621515680104494, + 0.009335354901850224, + 0.007152027916163206, + 0.008440363220870495, + 0.009117837995290756, + 0.008210484869778156, + 0.00451613450422883, + 0.004797255154699087, + 0.005981628783047199, + 0.007905900478363037, + 0.00526809599250555, + 0.005801124963909388, + 0.004901793319731951, + 0.011958029121160507, + 0.016839347779750824, + 0.012627608142793179, + 0.011162987910211086, + 0.012619959190487862, + 0.009099879302084446, + 0.008667917922139168, + 0.008535784669220448, + 0.013904937542974949, + 0.030894722789525986, + 0.17752553522586823, + 0.03832974284887314, + 0.21044285595417023, + 0.022807614877820015, + 0.008770457468926907, + 0.023173732683062553, + 0.011200153268873692, + 0.0031121568754315376, + 0.0028544520027935505, + 0.005234424490481615, + 0.004696154501289129, + 0.007153462152928114, + 0.006480949930846691, + 0.008272086270153522, + 0.004560795612633228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.029790453612804413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.009786483831703663 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786483831703663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786483831703663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10470446199178696, + 0.11346990615129471, + 0.13738055527210236, + 0.07956890761852264, + 0.04080065339803696, + 0.1577877402305603, + 0.07900384068489075, + 0.12938112020492554, + 0.0862046480178833, + 0.07169819623231888 + ], + "credit_map": [ + 0.025969836860895157, + 0.007842949591577053, + 0.02181214839220047, + 0.008880840614438057, + 0.007285203784704208, + 0.010713987052440643, + 0.00928106252104044, + 0.012918432243168354, + 0.026941226795315742, + 0.009695730172097683, + 0.01905408687889576, + 0.008147821761667728, + 0.007011039648205042, + 0.018191911280155182, + 0.009942963719367981, + 0.0144851328805089, + 0.01835644617676735, + 0.017192237079143524, + 0.019095443189144135, + 0.014706951566040516, + 0.020584270358085632, + 0.012321996502578259, + 0.0151631198823452, + 0.019960083067417145, + 0.0074386619962751865, + 0.00849728100001812, + 0.006444369908422232, + 0.007899860851466656, + 0.012302967719733715, + 0.012819929048418999, + 0.013963233679533005, + 0.010202601552009583, + 0.0043896459974348545, + 0.0031246296130120754, + 0.004583556670695543, + 0.005560317542403936, + 0.0043842908926308155, + 0.007071022409945726, + 0.006842521950602531, + 0.004844667389988899, + 0.014346195384860039, + 0.013535471633076668, + 0.01214347593486309, + 0.01336676999926567, + 0.02649890072643757, + 0.03883909061551094, + 0.015920288860797882, + 0.023137550801038742, + 0.007170998491346836, + 0.006015484686940908, + 0.00783601589500904, + 0.013074944727122784, + 0.006084183696657419, + 0.00781343225389719, + 0.006265053991228342, + 0.024743730202317238, + 0.00895262137055397, + 0.012767021544277668, + 0.010459205135703087, + 0.01872939243912697, + 0.015858398750424385, + 0.021808186545968056, + 0.011383024975657463, + 0.029423262923955917, + 0.01142131257802248, + 0.023482417687773705, + 0.010922059416770935, + 0.00756809301674366, + 0.005945930723100901, + 0.007099087815731764, + 0.013607732020318508, + 0.006158015225082636, + 0.005926764104515314, + 0.005272901151329279, + 0.01111019216477871, + 0.008146025240421295, + 0.01341091189533472, + 0.010009540244936943, + 0.011742914095520973, + 0.006078946869820356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.009786483831703663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.020577168092131615 + ], + "opposite_logprob_tokens": [ + -3.8955771923065186 + ], + "logprob_reward": -0.020577168092131615, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577168092131615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3499146103858948, + 0.07717657089233398, + 0.07979420572519302, + 0.06104707345366478, + 0.0327182300388813, + 0.06274017691612244, + 0.04913930967450142, + 0.06304490566253662, + 0.16418924927711487, + 0.060235682874917984 + ], + "credit_map": [ + 0.10742399096488953, + 0.017445746809244156, + 0.08330054581165314, + 0.016727417707443237, + 0.020105693489313126, + 0.03150429204106331, + 0.026811759918928146, + 0.04659518599510193, + 0.012323983944952488, + 0.00812013354152441, + 0.014194168150424957, + 0.010105717927217484, + 0.008399850688874722, + 0.00885683298110962, + 0.006978452671319246, + 0.008197428658604622, + 0.009436025284230709, + 0.01140451617538929, + 0.011107194237411022, + 0.008641378954052925, + 0.011784892529249191, + 0.007046697661280632, + 0.007845557294785976, + 0.012527942657470703, + 0.006975415162742138, + 0.00613925838842988, + 0.005699307192116976, + 0.005481641739606857, + 0.0053954338654875755, + 0.012927905656397343, + 0.009793180041015148, + 0.008634933270514011, + 0.0027966881170868874, + 0.0041418313048779964, + 0.0024424393195658922, + 0.0037730506155639887, + 0.003238497767597437, + 0.006032565608620644, + 0.004685978405177593, + 0.005607177037745714, + 0.007174905389547348, + 0.009168319404125214, + 0.005522958468645811, + 0.006372327916324139, + 0.00782336201518774, + 0.011470131576061249, + 0.00706171290948987, + 0.008146457374095917, + 0.004246794618666172, + 0.004231301601976156, + 0.005226687993854284, + 0.005610384047031403, + 0.006849201861768961, + 0.007530702743679285, + 0.005626166705042124, + 0.009818070568144321, + 0.005158405285328627, + 0.005117539782077074, + 0.0065687005408108234, + 0.009787355549633503, + 0.006818869616836309, + 0.013804261572659016, + 0.007841352373361588, + 0.007948421873152256, + 0.01742447167634964, + 0.060411881655454636, + 0.01102550607174635, + 0.046565067023038864, + 0.006975800730288029, + 0.0050954269245266914, + 0.012391732074320316, + 0.004299357533454895, + 0.003973511513322592, + 0.003769964911043644, + 0.009579941630363464, + 0.006162199657410383, + 0.011017626151442528, + 0.007712929975241423, + 0.013985595665872097, + 0.004033914767205715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.020577168092131615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle panning shot, moving steadily to the right, closely follows a bright yellow electric forklift as it smoothly navigates a narrow aisle between two towering, industrial-grade metal shelves. These massive shelves stretch from the smooth gray concrete floor to the high warehouse ceiling, stacked with neatly organized cartons and wooden pallets tightly wrapped in transparent plastic. The forklift bears the slightest traces of industrial wear, its metal forks held slightly raised for safe maneuvering. The rotating orange warning light on its roof flashes rhythmically, casting dynamic circles of amber light across the cold metal shelves and the shimmering floor. The forklift moves forward steadily and deliberately, its sturdy black rubber tires turning silently on the ground, accompanied only by the gentle, constant hum of the motor. The entire scene is illuminated evenly by the cool glow of rows of white fluorescent tubes on the ceiling, contrasting sharply with the warm, pulsing glow of the forklift's warning lights, creating an atmosphere of orderly industrial operation that is both quiet and full of potential power.", + "seed": 33474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_8/index_891_seed_33474_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": -0.026557157561182976, + "vlm_logprob": -0.026557157561182976, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.026557157561182976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.012010518461465836 + ], + "opposite_logprob_tokens": [ + -5.01201057434082 + ], + "logprob_reward": -0.012010518461465836, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012010518461465836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23118191957473755, + 0.20683234930038452, + 0.03874966874718666, + 0.05647043511271477, + 0.052319180220365524, + 0.1396220177412033, + 0.08115150779485703, + 0.0941237285733223, + 0.06431271880865097, + 0.035236481577157974 + ], + "credit_map": [ + 0.024728627875447273, + 0.034995924681425095, + 0.019602471962571144, + 0.03784632310271263, + 0.02075916901230812, + 0.031156087294220924, + 0.03868784382939339, + 0.023405462503433228, + 0.020283685997128487, + 0.021374186500906944, + 0.03437070548534393, + 0.027717381715774536, + 0.02672780677676201, + 0.021979186683893204, + 0.03438401594758034, + 0.019995369017124176, + 0.004854502622038126, + 0.004190434236079454, + 0.004167824052274227, + 0.004711013287305832, + 0.004664548207074404, + 0.006184972357004881, + 0.006195289548486471, + 0.003781082108616829, + 0.006076774559915066, + 0.009520924650132656, + 0.006266928277909756, + 0.007872330024838448, + 0.008589531295001507, + 0.009110515005886555, + 0.004803376737982035, + 0.004230049904435873, + 0.007549193222075701, + 0.006712887901812792, + 0.00539771281182766, + 0.00657907547429204, + 0.006635312922298908, + 0.007607836741954088, + 0.006186435930430889, + 0.005650721490383148, + 0.012270701117813587, + 0.016622232273221016, + 0.021601542830467224, + 0.021281346678733826, + 0.016506221145391464, + 0.015520133078098297, + 0.017416097223758698, + 0.01840374246239662, + 0.008447707630693913, + 0.009969174861907959, + 0.0126015804708004, + 0.014033684507012367, + 0.013402185402810574, + 0.008042585104703903, + 0.008016739040613174, + 0.006637839134782553, + 0.00786194484680891, + 0.015763795003294945, + 0.007951546460390091, + 0.009992982260882854, + 0.01878419704735279, + 0.011473619379103184, + 0.010104318149387836, + 0.012191321700811386, + 0.008446461521089077, + 0.00691341795027256, + 0.01081033144146204, + 0.008049117401242256, + 0.007090426050126553, + 0.007516872137784958, + 0.00918457843363285, + 0.006301506422460079, + 0.0032886825501918793, + 0.003742735832929611, + 0.004480163101106882, + 0.002815584186464548, + 0.00923140812665224, + 0.0045249988324940205, + 0.0037881170865148306, + 0.0033647918608039618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.012010518461465836 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.0142159853130579 + ], + "opposite_logprob_tokens": [ + -4.26421594619751 + ], + "logprob_reward": -0.0142159853130579, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0142159853130579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1990995556116104, + 0.1307726353406906, + 0.07130863517522812, + 0.05694616958498955, + 0.04951462522149086, + 0.11169753223657608, + 0.03591933846473694, + 0.1807253062725067, + 0.09286082535982132, + 0.07115545868873596 + ], + "credit_map": [ + 0.01786228083074093, + 0.024454807862639427, + 0.021628789603710175, + 0.022268852218985558, + 0.021874265745282173, + 0.030461449176073074, + 0.039235763251781464, + 0.021313326433300972, + 0.008853072300553322, + 0.012043531984090805, + 0.018862485885620117, + 0.012437108904123306, + 0.023875664919614792, + 0.02624259516596794, + 0.01708376407623291, + 0.011374397203326225, + 0.009235587902367115, + 0.006528904661536217, + 0.009049241431057453, + 0.009657223708927631, + 0.009660103358328342, + 0.008037935942411423, + 0.012433547526597977, + 0.006706082262098789, + 0.005259785335510969, + 0.007943917997181416, + 0.0070420377887785435, + 0.007368043530732393, + 0.009598487056791782, + 0.010334025137126446, + 0.004883888177573681, + 0.004515978042036295, + 0.006445035338401794, + 0.005849034991115332, + 0.006972774397581816, + 0.00477655278518796, + 0.004745334852486849, + 0.012199551798403263, + 0.0032071880996227264, + 0.005319149699062109, + 0.011060505174100399, + 0.013165146112442017, + 0.01585557870566845, + 0.014667708426713943, + 0.013841916806995869, + 0.021873075515031815, + 0.00879055168479681, + 0.012443031184375286, + 0.004009212367236614, + 0.004120885394513607, + 0.0049799042753875256, + 0.006029949523508549, + 0.004885893315076828, + 0.004890294745564461, + 0.003626876510679722, + 0.0033763176761567593, + 0.01739666610956192, + 0.02527288906276226, + 0.018022578209638596, + 0.014478394761681557, + 0.03164013847708702, + 0.03208126872777939, + 0.01670052483677864, + 0.02513282373547554, + 0.012847306206822395, + 0.00896511785686016, + 0.014797024428844452, + 0.009017673321068287, + 0.013729434460401535, + 0.011562705971300602, + 0.01265158411115408, + 0.009289964102208614, + 0.008352218195796013, + 0.006143799517303705, + 0.010328841395676136, + 0.005470689851790667, + 0.016897259280085564, + 0.010206236504018307, + 0.008057811297476292, + 0.005698598455637693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0142159853130579 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.04304855689406395 + ], + "opposite_logprob_tokens": [ + -3.168048620223999 + ], + "logprob_reward": -0.04304855689406395, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304855689406395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2339550256729126, + 0.09687092155218124, + 0.1840120553970337, + 0.054725442081689835, + 0.05516131594777107, + 0.13757026195526123, + 0.08348552137613297, + 0.10774967819452286, + 0.02246536873281002, + 0.024004438892006874 + ], + "credit_map": [ + 0.02266044169664383, + 0.03350171074271202, + 0.029594674706459045, + 0.0494508221745491, + 0.019078554585576057, + 0.03088633343577385, + 0.030852336436510086, + 0.017930135130882263, + 0.008698209188878536, + 0.012245427817106247, + 0.01170223206281662, + 0.011698989197611809, + 0.012257714755833149, + 0.012348830699920654, + 0.01914503425359726, + 0.008774465881288052, + 0.019075095653533936, + 0.018884025514125824, + 0.016243908554315567, + 0.018854200839996338, + 0.040247734636068344, + 0.030777696520090103, + 0.022374054417014122, + 0.017555316910147667, + 0.004383525811135769, + 0.006462922785431147, + 0.006177972070872784, + 0.008019784465432167, + 0.008692598901689053, + 0.01072633359581232, + 0.005762556102126837, + 0.004499743692576885, + 0.0062872362323105335, + 0.006773775443434715, + 0.006753908470273018, + 0.007165747229009867, + 0.0071619474329054356, + 0.011088103987276554, + 0.004002561792731285, + 0.005928033031523228, + 0.011733618564903736, + 0.017253803089261055, + 0.020304027944803238, + 0.019352443516254425, + 0.018759818747639656, + 0.019988594576716423, + 0.013187587261199951, + 0.016990359872579575, + 0.007969237864017487, + 0.0075564635917544365, + 0.021302953362464905, + 0.0117868110537529, + 0.009281270205974579, + 0.0062033566646277905, + 0.006520627066493034, + 0.012864788994193077, + 0.012873292900621891, + 0.01667950302362442, + 0.01083856076002121, + 0.011437035165727139, + 0.016512785106897354, + 0.013155137188732624, + 0.010923348367214203, + 0.015330011956393719, + 0.004886121954768896, + 0.0024374290369451046, + 0.0041115256026387215, + 0.0018807084998115897, + 0.0019599641673266888, + 0.00281540141440928, + 0.0020002375822514296, + 0.002373979426920414, + 0.001740887644700706, + 0.0029777290765196085, + 0.0025956910103559494, + 0.002027832670137286, + 0.005793402902781963, + 0.003486670320853591, + 0.0031245185527950525, + 0.002257704036310315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.04304855689406395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.03370916470885277 + ], + "opposite_logprob_tokens": [ + -3.4087090492248535 + ], + "logprob_reward": -0.03370916470885277, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370916470885277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1614982783794403, + 0.3984942138195038, + 0.10767734050750732, + 0.06869568675756454, + 0.04834664240479469, + 0.06618152558803558, + 0.04121467471122742, + 0.03768943250179291, + 0.03228917345404625, + 0.037913065403699875 + ], + "credit_map": [ + 0.01763046532869339, + 0.026149017736315727, + 0.01750953122973442, + 0.02119392715394497, + 0.02137249894440174, + 0.02803138829767704, + 0.017570575699210167, + 0.012040875852108002, + 0.026863861829042435, + 0.038947757333517075, + 0.04886561632156372, + 0.03199722617864609, + 0.08929518610239029, + 0.09273423999547958, + 0.040685586631298065, + 0.02910471521317959, + 0.012607662938535213, + 0.010149591602385044, + 0.014247365295886993, + 0.017330864444375038, + 0.015623646788299084, + 0.014952590689063072, + 0.012483346275985241, + 0.010282265953719616, + 0.005955629050731659, + 0.00989613588899374, + 0.006601959932595491, + 0.0068656159564852715, + 0.008508505299687386, + 0.01844114437699318, + 0.006557242013514042, + 0.005869452841579914, + 0.006441547069698572, + 0.005754953715950251, + 0.007294056471437216, + 0.0046676271595060825, + 0.005193083547055721, + 0.011556628160178661, + 0.0035010993015021086, + 0.003937648609280586, + 0.005852597765624523, + 0.007402633782476187, + 0.009346273727715015, + 0.00808996893465519, + 0.009872045367956161, + 0.010097866877913475, + 0.0066320826299488544, + 0.008888055570423603, + 0.004288090392947197, + 0.005577018018811941, + 0.006465058773756027, + 0.006783969234675169, + 0.004410500172525644, + 0.004995220340788364, + 0.005058472044765949, + 0.003636348992586136, + 0.00528978044167161, + 0.005449355114251375, + 0.004687571432441473, + 0.003781759412959218, + 0.004675517790019512, + 0.004174810368567705, + 0.0036636649165302515, + 0.005966972094029188, + 0.0066430107690393925, + 0.0029958621598780155, + 0.006449766457080841, + 0.003246963955461979, + 0.0027065265458077192, + 0.003601227654144168, + 0.003235491458326578, + 0.0034103253856301308, + 0.004558716434985399, + 0.005798273254185915, + 0.004143087659031153, + 0.003272572997957468, + 0.007724385242909193, + 0.004864808171987534, + 0.004259265959262848, + 0.0032919577788561583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.03370916470885277 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.029801562428474426 + ], + "opposite_logprob_tokens": [ + -3.529801607131958 + ], + "logprob_reward": -0.029801562428474426, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801562428474426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2398151308298111, + 0.12890630960464478, + 0.07910731434822083, + 0.3477790653705597, + 0.053951557725667953, + 0.05000833049416542, + 0.019407490268349648, + 0.029686488211154938, + 0.02729746885597706, + 0.024040870368480682 + ], + "credit_map": [ + 0.018667759373784065, + 0.02809884585440159, + 0.019324442371726036, + 0.03697330877184868, + 0.02877645194530487, + 0.031540531665086746, + 0.04704764112830162, + 0.029386142268776894, + 0.011396058835089207, + 0.014699726365506649, + 0.017641926184296608, + 0.014070844277739525, + 0.022091297432780266, + 0.022898094728589058, + 0.014783570542931557, + 0.011324787512421608, + 0.009736200794577599, + 0.010005831718444824, + 0.008760987780988216, + 0.009838071651756763, + 0.01200654823333025, + 0.014183659106492996, + 0.007989807985723019, + 0.006586204748600721, + 0.02775263413786888, + 0.033292461186647415, + 0.034210335463285446, + 0.029599327594041824, + 0.04667322710156441, + 0.11190235614776611, + 0.027600347995758057, + 0.03674836456775665, + 0.006432495545595884, + 0.0050317938439548016, + 0.0076606678776443005, + 0.005217444151639938, + 0.006363122258335352, + 0.013995901681482792, + 0.0037574530579149723, + 0.005492677446454763, + 0.0047028944827616215, + 0.006130150519311428, + 0.007807418238371611, + 0.007617506664246321, + 0.00472173560410738, + 0.007614099886268377, + 0.0042811292223632336, + 0.007133394945412874, + 0.0019748893100768328, + 0.002440396463498473, + 0.002726898528635502, + 0.002928494242951274, + 0.002453709254041314, + 0.002377110533416271, + 0.0022778292186558247, + 0.0022281636483967304, + 0.004171217791736126, + 0.004840840119868517, + 0.0037656782660633326, + 0.002830508165061474, + 0.004235663916915655, + 0.002967853331938386, + 0.00293349358253181, + 0.003941233269870281, + 0.004603849723935127, + 0.0026525938883423805, + 0.006719761062413454, + 0.0022601208183914423, + 0.0026812998112291098, + 0.002804618561640382, + 0.002569311996921897, + 0.003005912061780691, + 0.0024811800103634596, + 0.003474052529782057, + 0.003187313675880432, + 0.0022222353145480156, + 0.005465763621032238, + 0.002648192225024104, + 0.002323088701814413, + 0.0022390468511730433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.029801562428474426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle panning shot, moving steadily to the right, closely follows a bright yellow electric forklift as it smoothly navigates a narrow aisle between two towering, industrial-grade metal shelves. These massive shelves stretch from the smooth gray concrete floor to the high warehouse ceiling, stacked with neatly organized cartons and wooden pallets tightly wrapped in transparent plastic. The forklift bears the slightest traces of industrial wear, its metal forks held slightly raised for safe maneuvering. The rotating orange warning light on its roof flashes rhythmically, casting dynamic circles of amber light across the cold metal shelves and the shimmering floor. The forklift moves forward steadily and deliberately, its sturdy black rubber tires turning silently on the ground, accompanied only by the gentle, constant hum of the motor. The entire scene is illuminated evenly by the cool glow of rows of white fluorescent tubes on the ceiling, contrasting sharply with the warm, pulsing glow of the forklift's warning lights, creating an atmosphere of orderly industrial operation that is both quiet and full of potential power.", + "seed": 33475, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_8/index_891_seed_33475_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": -0.05572188124060631, + "vlm_logprob": -0.05572188124060631, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.05572188124060631, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.05550985410809517 + ], + "opposite_logprob_tokens": [ + -3.1805098056793213 + ], + "logprob_reward": -0.05550985410809517, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05550985410809517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0813843384385109, + 0.07396598905324936, + 0.18310736119747162, + 0.07816802710294724, + 0.14218027889728546, + 0.10798537731170654, + 0.1084047332406044, + 0.04892599582672119, + 0.1595885455608368, + 0.016289331018924713 + ], + "credit_map": [ + 0.008626033551990986, + 0.008312389254570007, + 0.008611970581114292, + 0.013556737452745438, + 0.009446213021874428, + 0.015163530595600605, + 0.01055602915585041, + 0.007111446466296911, + 0.009831649251282215, + 0.0064885118044912815, + 0.009714731015264988, + 0.005729368422180414, + 0.01346828043460846, + 0.014274520799517632, + 0.007110909558832645, + 0.007348025217652321, + 0.020711466670036316, + 0.014889279380440712, + 0.023620065301656723, + 0.017474809661507607, + 0.052057791501283646, + 0.01635928824543953, + 0.025759465992450714, + 0.012235202826559544, + 0.009506918489933014, + 0.008649087511003017, + 0.0069280946627259254, + 0.010662093758583069, + 0.01087047066539526, + 0.011409327387809753, + 0.00905284471809864, + 0.011089195497334003, + 0.027596376836299896, + 0.008480747230350971, + 0.01770303212106228, + 0.015603648498654366, + 0.008772759698331356, + 0.01824701391160488, + 0.024207398295402527, + 0.02156931534409523, + 0.01505254115909338, + 0.020861389115452766, + 0.008989513851702213, + 0.014166124165058136, + 0.014484674669802189, + 0.012252965942025185, + 0.013906216248869896, + 0.008271967060863972, + 0.011261386796832085, + 0.004736960865557194, + 0.009557230398058891, + 0.007033372763544321, + 0.00874148216098547, + 0.02696998044848442, + 0.027661586180329323, + 0.012442735955119133, + 0.004658908117562532, + 0.004570143762975931, + 0.004719811491668224, + 0.00579387741163373, + 0.005309827625751495, + 0.009410436265170574, + 0.006932320073246956, + 0.007530678063631058, + 0.0203861091285944, + 0.02164703980088234, + 0.01323883980512619, + 0.014896176755428314, + 0.022707805037498474, + 0.02006092108786106, + 0.026406576856970787, + 0.020245101302862167, + 0.0016924585215747356, + 0.0030823461711406708, + 0.0016525904648005962, + 0.001985996961593628, + 0.0017191238002851605, + 0.002256410662084818, + 0.0017339085461571813, + 0.0021664968226104975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.05550985410809517 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.07006516307592392 + ], + "opposite_logprob_tokens": [ + -2.6950652599334717 + ], + "logprob_reward": -0.07006516307592392, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006516307592392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10286228358745575, + 0.05376778170466423, + 0.5398789644241333, + 0.16823653876781464, + 0.016047559678554535, + 0.011338412761688232, + 0.026578592136502266, + 0.020623862743377686, + 0.046368349343538284, + 0.014297750778496265 + ], + "credit_map": [ + 0.015243601985275745, + 0.00978800281882286, + 0.013963290490210056, + 0.01151197124272585, + 0.01409944612532854, + 0.01780923269689083, + 0.011292930692434311, + 0.009153801947832108, + 0.006321960128843784, + 0.004910022486001253, + 0.006960784085094929, + 0.004694445990025997, + 0.011547544039785862, + 0.006877020932734013, + 0.006066369358450174, + 0.0063896263018250465, + 0.06054268032312393, + 0.038843877613544464, + 0.07754675298929214, + 0.047158509492874146, + 0.15811468660831451, + 0.048511065542697906, + 0.07580593973398209, + 0.03335539624094963, + 0.026297805830836296, + 0.017771907150745392, + 0.015058633871376514, + 0.01723152957856655, + 0.03063127025961876, + 0.027608279138803482, + 0.0223226435482502, + 0.01131444051861763, + 0.0020821872167289257, + 0.0015037731500342488, + 0.0016472918214276433, + 0.0023292391560971737, + 0.0013237670063972473, + 0.0026818225160241127, + 0.0023974147625267506, + 0.0020820617210119963, + 0.001383829046972096, + 0.002122266450896859, + 0.0010282397270202637, + 0.001360693946480751, + 0.0014165517641231418, + 0.0015214988961815834, + 0.0013927390100434422, + 0.001112592639401555, + 0.0022250195033848286, + 0.002327806781977415, + 0.003356973873451352, + 0.002021646359935403, + 0.0020443026442080736, + 0.006313234101980925, + 0.004993669223040342, + 0.003295939415693283, + 0.002333201002329588, + 0.001324679353274405, + 0.0018580470932647586, + 0.002651696326211095, + 0.0018280234653502703, + 0.003641196759417653, + 0.0032691366504877806, + 0.003717882325872779, + 0.0073625072836875916, + 0.007499197497963905, + 0.003134192666038871, + 0.004485048819333315, + 0.0064399223774671555, + 0.005249241832643747, + 0.006382979918271303, + 0.005815255455672741, + 0.001234002411365509, + 0.0016148202121257782, + 0.0016812138492241502, + 0.0018006317550316453, + 0.0012585015501827002, + 0.0029675120022147894, + 0.0014437512727454305, + 0.002297317376360297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.07006516307592392 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.026346195489168167 + ], + "opposite_logprob_tokens": [ + -3.651346206665039 + ], + "logprob_reward": -0.026346195489168167, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346195489168167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1123589351773262, + 0.06011061370372772, + 0.11932259052991867, + 0.11444966495037079, + 0.06473031640052795, + 0.11909174174070358, + 0.19935540854930878, + 0.11227099597454071, + 0.06100005283951759, + 0.037309661507606506 + ], + "credit_map": [ + 0.012429263442754745, + 0.013980759307742119, + 0.013475814834237099, + 0.010677569545805454, + 0.01854354329407215, + 0.019265972077846527, + 0.013268242590129375, + 0.01071777381002903, + 0.009111085906624794, + 0.004709586501121521, + 0.013004251755774021, + 0.0066856639459729195, + 0.0110923545435071, + 0.004994645714759827, + 0.005850939080119133, + 0.00466208765283227, + 0.011648126877844334, + 0.010730047710239887, + 0.016080403700470924, + 0.01268335897475481, + 0.03189200907945633, + 0.010268784128129482, + 0.015381945297122002, + 0.010637912899255753, + 0.015561016276478767, + 0.013131354935467243, + 0.011524751782417297, + 0.012796364724636078, + 0.020060302689671516, + 0.017555000260472298, + 0.0138963358476758, + 0.00992454495280981, + 0.010652988217771053, + 0.004183595534414053, + 0.00816704798489809, + 0.006452225148677826, + 0.0055797821842134, + 0.011436420492827892, + 0.008730082772672176, + 0.009528175927698612, + 0.016516821458935738, + 0.016080817207694054, + 0.014363016001880169, + 0.012167593464255333, + 0.016229864209890366, + 0.016550853848457336, + 0.014408137649297714, + 0.01277463510632515, + 0.01738685369491577, + 0.014899409376084805, + 0.015260076150298119, + 0.012853535823523998, + 0.016068950295448303, + 0.05131584405899048, + 0.046575579792261124, + 0.024995163083076477, + 0.008687453344464302, + 0.01510740164667368, + 0.009418082423508167, + 0.013576057739555836, + 0.013406243175268173, + 0.019229546189308167, + 0.01591416820883751, + 0.016932053491473198, + 0.006107908673584461, + 0.006306564435362816, + 0.006097849458456039, + 0.006237661000341177, + 0.007359711453318596, + 0.009192472323775291, + 0.010467425920069218, + 0.009230460040271282, + 0.0029998584650456905, + 0.0033480266574770212, + 0.0037245897110551596, + 0.0054402584210038185, + 0.003980247303843498, + 0.007620861288160086, + 0.004608494695276022, + 0.005587325897067785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.026346195489168167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.02636094018816948 + ], + "opposite_logprob_tokens": [ + -3.6513609886169434 + ], + "logprob_reward": -0.02636094018816948, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636094018816948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18666988611221313, + 0.08078467845916748, + 0.326638400554657, + 0.12041941285133362, + 0.03449559956789017, + 0.03349226340651512, + 0.04555520787835121, + 0.09470157325267792, + 0.05334538593888283, + 0.023897597566246986 + ], + "credit_map": [ + 0.020023543387651443, + 0.020794110372662544, + 0.028896180912852287, + 0.03176203370094299, + 0.02692514844238758, + 0.019995978102087975, + 0.023746313527226448, + 0.01452656090259552, + 0.012054385617375374, + 0.0038611693307757378, + 0.016714749857783318, + 0.004831002093851566, + 0.016742154955863953, + 0.010642755776643753, + 0.007429748773574829, + 0.008508704602718353, + 0.03811004012823105, + 0.026276355609297752, + 0.03893309459090233, + 0.031290203332901, + 0.09668201953172684, + 0.029124056920409203, + 0.04374094307422638, + 0.022481679916381836, + 0.016618259251117706, + 0.014129213988780975, + 0.009185525588691235, + 0.01339542306959629, + 0.022160084918141365, + 0.01979525201022625, + 0.01610296219587326, + 0.009032675996422768, + 0.005520758684724569, + 0.003122876398265362, + 0.00366177037358284, + 0.003345574950799346, + 0.003403075970709324, + 0.005487967748194933, + 0.004782491829246283, + 0.005171080585569143, + 0.00430170027539134, + 0.0028050255496054888, + 0.0027873667422682047, + 0.003922173287719488, + 0.004313149023801088, + 0.005044970661401749, + 0.005247292574495077, + 0.005070579703897238, + 0.0035942993126809597, + 0.0036008439492434263, + 0.005131265614181757, + 0.003992117941379547, + 0.004264307674020529, + 0.010603231377899647, + 0.007944784127175808, + 0.006424351595342159, + 0.008331271819770336, + 0.009382219053804874, + 0.010906546376645565, + 0.010017847642302513, + 0.0097130062058568, + 0.017798295244574547, + 0.015342678874731064, + 0.013209698721766472, + 0.006607421208173037, + 0.0064374106004834175, + 0.005352957174181938, + 0.00499527994543314, + 0.006681447383016348, + 0.005549791734665632, + 0.010434495285153389, + 0.007286576088517904, + 0.0022118340712040663, + 0.003356384579092264, + 0.002711363136768341, + 0.003891212632879615, + 0.0024927351623773575, + 0.002902229083701968, + 0.0032989608589559793, + 0.0030328771099448204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.02636094018816948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow electric forklift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the forklift navigate a narrow aisle between towering industrial-grade metal shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle and panning steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rotating orange warning light on the forklift's roof visible and flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool white fluorescent tubes contrasting with the warm amber glow of the warning 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": [ + -0.1003272533416748 + ], + "opposite_logprob_tokens": [ + -2.350327253341675 + ], + "logprob_reward": -0.1003272533416748, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003272533416748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09450313448905945, + 0.3288182020187378, + 0.047905221581459045, + 0.17177675664424896, + 0.07142620533704758, + 0.038642965257167816, + 0.09619192034006119, + 0.06559491902589798, + 0.049051783978939056, + 0.03608895093202591 + ], + "credit_map": [ + 0.012924063950777054, + 0.013854462653398514, + 0.012836671434342861, + 0.010413865558803082, + 0.012403418309986591, + 0.009849913418293, + 0.010296386666595936, + 0.011924347840249538, + 0.05012008920311928, + 0.01588156633079052, + 0.05655001476407051, + 0.01926148310303688, + 0.05943156033754349, + 0.06380831450223923, + 0.029298383742570877, + 0.034466780722141266, + 0.004414204508066177, + 0.004805073607712984, + 0.005461090710014105, + 0.006159864366054535, + 0.008369917050004005, + 0.00493428111076355, + 0.007327873259782791, + 0.006432916969060898, + 0.024168433621525764, + 0.019384704530239105, + 0.013310550712049007, + 0.019923005253076553, + 0.030619533732533455, + 0.025207413360476494, + 0.022886216640472412, + 0.016276894137263298, + 0.008195782080292702, + 0.004068682435899973, + 0.014620348811149597, + 0.005119506735354662, + 0.004861443303525448, + 0.011651045642793179, + 0.01100168190896511, + 0.011907713487744331, + 0.004906268324702978, + 0.004330395255237818, + 0.003830295754596591, + 0.004721538629382849, + 0.00545485457405448, + 0.005016260780394077, + 0.0056853280402719975, + 0.004698025528341532, + 0.008472364395856857, + 0.006299580447375774, + 0.00681378273293376, + 0.005972962826490402, + 0.007975995540618896, + 0.027193987742066383, + 0.023830000311136246, + 0.009633241221308708, + 0.007216640282422304, + 0.006136796902865171, + 0.0057270582765340805, + 0.006931676995009184, + 0.007411076687276363, + 0.0121756074950099, + 0.010541609488427639, + 0.009454445913434029, + 0.003975193481892347, + 0.005222567357122898, + 0.005720723420381546, + 0.004919860977679491, + 0.00694469315931201, + 0.007507369387894869, + 0.008373726159334183, + 0.006387651432305574, + 0.002868668409064412, + 0.003114342223852873, + 0.0031415452249348164, + 0.003937085159122944, + 0.004037659615278244, + 0.008626551367342472, + 0.004763500764966011, + 0.005599601194262505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.1003272533416748 + } + ], + "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/0000025/group_9/index_tensor([6028, 6028])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_9/index_tensor([6028, 6028])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3ca81d1cc458bbf43ca8d737172e9f6de9f71909 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_9/index_tensor([6028, 6028])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro close-up, panning slowly and steadily to the right, locks the viewer's gaze squarely on a pair of hands, etched with deep wrinkles and light brown age spots. Softly illuminated by a side-lit desk lamp, every wrinkle and vein is clearly visible, as if bearing witness to the passage of time. One of the larger, more robust hands (male), with slightly thicker knuckles, firmly holds the base of a dark, softly patinated antique wood. The other, a more delicate hand (female), with its rounded, neatly manicured nails, gently and slowly traces the intricate and delicate carvings on the object's surface with an almost reverent gesture. As the camera pans, we see the fingertips delicately dig into the carved grooves, feeling the subtle undulations of the wood's grain. The light, as it passes over the fingertips, casts a fleeting, soft highlight on the smoothed, raised edges. The background is blurred into a fuzzy warm color, and a corner of the study can be vaguely seen. The whole picture is shrouded in a quiet, focused and affectionate nostalgic atmosphere, as if this pair of silent hands are telling a story spanning a long time.", + "seed": 1024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_9/index_6028_seed_1024_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": -0.273275001719594, + "vlm_logprob": -0.273275001719594, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.273275001719594, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04715163633227348 + ], + "opposite_logprob_tokens": [ + -3.172151565551758 + ], + "logprob_reward": -0.04715163633227348, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04715163633227348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17588309943675995, + 0.09288496524095535, + 0.0901058241724968, + 0.06846737116575241, + 0.08636561781167984, + 0.1448836624622345, + 0.1180386021733284, + 0.07056710124015808, + 0.03811703622341156, + 0.11468669772148132 + ], + "credit_map": [ + 0.01837044581770897, + 0.02091987617313862, + 0.023113908246159554, + 0.018431011587381363, + 0.012993830256164074, + 0.042283181101083755, + 0.013953158631920815, + 0.02581770531833172, + 0.008636269718408585, + 0.009427808225154877, + 0.018510084599256516, + 0.010762832127511501, + 0.009235173463821411, + 0.02055247128009796, + 0.0075338054448366165, + 0.008226531557738781, + 0.009534933604300022, + 0.008381185121834278, + 0.012884111143648624, + 0.01254439726471901, + 0.007668418809771538, + 0.021768739446997643, + 0.007601460907608271, + 0.009722585789859295, + 0.005350295454263687, + 0.007772000972181559, + 0.009827690199017525, + 0.011784074828028679, + 0.007306832354515791, + 0.0071109747514128685, + 0.009126151911914349, + 0.010189354419708252, + 0.008432039991021156, + 0.008612871170043945, + 0.020728087052702904, + 0.01545210462063551, + 0.0056837499141693115, + 0.00880521722137928, + 0.007740914821624756, + 0.010910638608038425, + 0.01346306037157774, + 0.011686391197144985, + 0.03743185102939606, + 0.019384073093533516, + 0.008661117404699326, + 0.014149641618132591, + 0.013386616483330727, + 0.026720933616161346, + 0.011523444205522537, + 0.012334034778177738, + 0.014893214218318462, + 0.02656722627580166, + 0.009065099060535431, + 0.01598292589187622, + 0.012051752768456936, + 0.015620915219187737, + 0.005073160864412785, + 0.009573747403919697, + 0.010311434045433998, + 0.011310835368931293, + 0.008577604778110981, + 0.008805441670119762, + 0.009333329275250435, + 0.007581557147204876, + 0.0033131763339042664, + 0.003680664347484708, + 0.010827267542481422, + 0.004508086945861578, + 0.00348231615498662, + 0.005183140281587839, + 0.003401997033506632, + 0.0037203894462436438, + 0.007774829864501953, + 0.012318815104663372, + 0.008780173026025295, + 0.012916949577629566, + 0.02588500641286373, + 0.009843410924077034, + 0.015693366527557373, + 0.021474162116646767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.04715163633227348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00765107199549675 + ], + "opposite_logprob_tokens": [ + -4.882650852203369 + ], + "logprob_reward": -0.00765107199549675, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00765107199549675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12622997164726257, + 0.1395520120859146, + 0.29723814129829407, + 0.06618370115756989, + 0.034115374088287354, + 0.1178973987698555, + 0.057269275188446045, + 0.03742746636271477, + 0.021586181595921516, + 0.10250046849250793 + ], + "credit_map": [ + 0.016629593446850777, + 0.01310147624462843, + 0.01103159412741661, + 0.01315904501825571, + 0.008786666207015514, + 0.03620367869734764, + 0.009967249818146229, + 0.017350677400827408, + 0.013524558395147324, + 0.0063906204886734486, + 0.046415988355875015, + 0.013863381929695606, + 0.008748389780521393, + 0.03047293983399868, + 0.008779509924352169, + 0.011356620118021965, + 0.0219391118735075, + 0.0344528928399086, + 0.022903762757778168, + 0.02581721358001232, + 0.0239896047860384, + 0.0817624107003212, + 0.026232201606035233, + 0.0601409450173378, + 0.004687432199716568, + 0.005299482494592667, + 0.012480826117098331, + 0.011418460868299007, + 0.005725070368498564, + 0.007220729254186153, + 0.01072544977068901, + 0.008626255206763744, + 0.0029514781199395657, + 0.002870909171178937, + 0.006639275699853897, + 0.005627370439469814, + 0.002427891129627824, + 0.0034886221401393414, + 0.00537894107401371, + 0.004730887711048126, + 0.00938002672046423, + 0.010547715239226818, + 0.026044359430670738, + 0.017498357221484184, + 0.009346659295260906, + 0.010739012621343136, + 0.01466844417154789, + 0.019672827795147896, + 0.005663507618010044, + 0.006095342803746462, + 0.007403785362839699, + 0.013203606009483337, + 0.004284835420548916, + 0.007526187226176262, + 0.005682691000401974, + 0.0074093216098845005, + 0.0028855169657617807, + 0.004686891101300716, + 0.006503106560558081, + 0.004401339218020439, + 0.004550369456410408, + 0.004966540727764368, + 0.005536861252039671, + 0.0038968415465205908, + 0.0018580651376396418, + 0.003040141658857465, + 0.0036125851329416037, + 0.0028487155213952065, + 0.002920803613960743, + 0.0016391067765653133, + 0.0032602972351014614, + 0.002406463958323002, + 0.0062528993003070354, + 0.012694785371422768, + 0.010866334661841393, + 0.01084016915410757, + 0.018041683360934258, + 0.007535629440099001, + 0.012894565239548683, + 0.023374395444989204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00765107199549675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603558212518692 + ], + "opposite_logprob_tokens": [ + -1.910355806350708 + ], + "logprob_reward": -0.1603558212518692, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603558212518692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034759484231472015, + 0.32668647170066833, + 0.10125347226858139, + 0.0864226371049881, + 0.04777985066175461, + 0.04838322103023529, + 0.12849904596805573, + 0.08728276193141937, + 0.11457094550132751, + 0.02436213195323944 + ], + "credit_map": [ + 0.004362939856946468, + 0.0038879746571183205, + 0.004321079235523939, + 0.0038466814439743757, + 0.003711183089762926, + 0.007459578569978476, + 0.003264796920120716, + 0.0039052481297403574, + 0.030249420553445816, + 0.01516914926469326, + 0.10988578200340271, + 0.03253697231411934, + 0.01957021839916706, + 0.07271751016378403, + 0.02122815139591694, + 0.025329260155558586, + 0.008990952745079994, + 0.005551865324378014, + 0.012298292480409145, + 0.00725938007235527, + 0.004358053673058748, + 0.02734372206032276, + 0.007447720970958471, + 0.02800348773598671, + 0.0074891517870128155, + 0.010209625586867332, + 0.016670964658260345, + 0.01049466896802187, + 0.006503112614154816, + 0.012928287498652935, + 0.007374068722128868, + 0.014752754010260105, + 0.004079245030879974, + 0.004905628506094217, + 0.011571765877306461, + 0.009556657634675503, + 0.0029046954587101936, + 0.004591676872223616, + 0.00401989696547389, + 0.006150282919406891, + 0.004219559486955404, + 0.005347112659364939, + 0.008445710875093937, + 0.0074884300120174885, + 0.0044948565773665905, + 0.005317972972989082, + 0.006326834671199322, + 0.006742742378264666, + 0.011923731304705143, + 0.012892135418951511, + 0.017712295055389404, + 0.028109030798077583, + 0.009324689395725727, + 0.017698997631669044, + 0.012611176818609238, + 0.01822698675096035, + 0.006922923028469086, + 0.012252109125256538, + 0.013608180917799473, + 0.012049558572471142, + 0.011197912506759167, + 0.010400352068245411, + 0.011894677765667439, + 0.008957046084105968, + 0.00832823384553194, + 0.007345772814005613, + 0.033981386572122574, + 0.011533142067492008, + 0.010934881865978241, + 0.011270160786807537, + 0.014137095771729946, + 0.017040282487869263, + 0.001719785388559103, + 0.0029959825333207846, + 0.002123098587617278, + 0.002395707182586193, + 0.004479369148612022, + 0.002503808354958892, + 0.0036035324446856976, + 0.004540849011391401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.1603558212518692 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.1370208263397217 + ], + "opposite_logprob_tokens": [ + -0.3870207965373993 + ], + "logprob_reward": -1.1370208263397217, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370208263397217, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04448278620839119, + 0.05728111043572426, + 0.2491624653339386, + 0.17666274309158325, + 0.05824906378984451, + 0.17014196515083313, + 0.10001731663942337, + 0.03632473573088646, + 0.0557386614382267, + 0.05193911865353584 + ], + "credit_map": [ + 0.005818378180265427, + 0.005902937613427639, + 0.005963167175650597, + 0.007247993256896734, + 0.0030078906565904617, + 0.0073380861431360245, + 0.004730010870844126, + 0.0044743213802576065, + 0.006368950475007296, + 0.003799173515290022, + 0.017251739278435707, + 0.005280709825456142, + 0.0033868008758872747, + 0.012973444536328316, + 0.003698528278619051, + 0.004521761555224657, + 0.020479073747992516, + 0.021229544654488564, + 0.027525736019015312, + 0.021999459713697433, + 0.013792752288281918, + 0.07414600998163223, + 0.015100848861038685, + 0.05488903075456619, + 0.015304895117878914, + 0.01882605254650116, + 0.024307504296302795, + 0.02397284470498562, + 0.020410312339663506, + 0.025573061779141426, + 0.017553841695189476, + 0.030714236199855804, + 0.005053873173892498, + 0.006506009958684444, + 0.014162685722112656, + 0.009820637293159962, + 0.0037602968513965607, + 0.004598657134920359, + 0.006355325225740671, + 0.007991579361259937, + 0.012699061073362827, + 0.013506440445780754, + 0.042708903551101685, + 0.022760378196835518, + 0.011771128512918949, + 0.015342620201408863, + 0.021031897515058517, + 0.030321534723043442, + 0.00798650924116373, + 0.008704852312803268, + 0.017444392666220665, + 0.018611347302794456, + 0.007959112524986267, + 0.014819410629570484, + 0.009213311597704887, + 0.015278379432857037, + 0.0029805332887917757, + 0.005396923050284386, + 0.0057959710247814655, + 0.004941651131957769, + 0.005463692359626293, + 0.004051708150655031, + 0.004158383700996637, + 0.003535872558131814, + 0.00475555332377553, + 0.005704066716134548, + 0.014200709760189056, + 0.006339920684695244, + 0.006739432457834482, + 0.006936715915799141, + 0.006497815251350403, + 0.004564445931464434, + 0.003427832154557109, + 0.0042809899896383286, + 0.005455162841826677, + 0.006921194959431887, + 0.00973360426723957, + 0.0035828540567308664, + 0.010944793932139874, + 0.0075926873832941055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -1.1370208263397217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195652678608894 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195652678608894, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195652678608894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04168444126844406, + 0.35367804765701294, + 0.14990808069705963, + 0.14298611879348755, + 0.04419426992535591, + 0.11235315352678299, + 0.06248548626899719, + 0.06289998441934586, + 0.011765790171921253, + 0.01804461143910885 + ], + "credit_map": [ + 0.004762796685099602, + 0.005146906711161137, + 0.006627062801271677, + 0.0056954845786094666, + 0.004179328214377165, + 0.006490746047347784, + 0.004405608866363764, + 0.004376505967229605, + 0.03405575826764107, + 0.020222416147589684, + 0.11652180552482605, + 0.03305279463529587, + 0.021778134629130363, + 0.07867632061243057, + 0.022286824882030487, + 0.027083996683359146, + 0.011694616638123989, + 0.0175889004021883, + 0.014833253808319569, + 0.01425267942249775, + 0.010066651739180088, + 0.04119786620140076, + 0.008549251593649387, + 0.031724851578474045, + 0.012203022837638855, + 0.015303835272789001, + 0.019722728058695793, + 0.020873388275504112, + 0.016408594325184822, + 0.019401049241423607, + 0.014463454484939575, + 0.024610044434666634, + 0.004059391561895609, + 0.0046800272539258, + 0.011028843000531197, + 0.00862953532487154, + 0.00242600217461586, + 0.003680752357468009, + 0.004284523893147707, + 0.0054051922634243965, + 0.008387366309762001, + 0.007361259777098894, + 0.0335727222263813, + 0.012962374836206436, + 0.006679438520222902, + 0.009532245807349682, + 0.013301214203238487, + 0.020556533709168434, + 0.006158283445984125, + 0.0070835198275744915, + 0.007752955425530672, + 0.009347904473543167, + 0.005472966004163027, + 0.010115633718669415, + 0.006741264835000038, + 0.009812957607209682, + 0.005740872584283352, + 0.008970454335212708, + 0.008996979333460331, + 0.007937407121062279, + 0.008910691365599632, + 0.007697302382439375, + 0.00781281664967537, + 0.006833462044596672, + 0.0013399975141510367, + 0.0011523751309141517, + 0.0026276519056409597, + 0.001256815274246037, + 0.0013188202865421772, + 0.001101842848584056, + 0.0016746257897466421, + 0.0012936615385115147, + 0.0012392858043313026, + 0.0014646764611825347, + 0.0019515084568411112, + 0.002393212402239442, + 0.0026597001124173403, + 0.0022330430801957846, + 0.0032483041286468506, + 0.0028548813425004482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.014195652678608894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, panning slowly and steadily to the right, locks the viewer's gaze squarely on a pair of hands, etched with deep wrinkles and light brown age spots. Softly illuminated by a side-lit desk lamp, every wrinkle and vein is clearly visible, as if bearing witness to the passage of time. One of the larger, more robust hands (male), with slightly thicker knuckles, firmly holds the base of a dark, softly patinated antique wood. The other, a more delicate hand (female), with its rounded, neatly manicured nails, gently and slowly traces the intricate and delicate carvings on the object's surface with an almost reverent gesture. As the camera pans, we see the fingertips delicately dig into the carved grooves, feeling the subtle undulations of the wood's grain. The light, as it passes over the fingertips, casts a fleeting, soft highlight on the smoothed, raised edges. The background is blurred into a fuzzy warm color, and a corner of the study can be vaguely seen. The whole picture is shrouded in a quiet, focused and affectionate nostalgic atmosphere, as if this pair of silent hands are telling a story spanning a long time.", + "seed": 1025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_9/index_6028_seed_1025_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": -0.15891168285161256, + "vlm_logprob": -0.15891168285161256, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.15891168285161256, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08356603980064392 + ], + "opposite_logprob_tokens": [ + -2.5835659503936768 + ], + "logprob_reward": -0.08356603980064392, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08356603980064392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10602648556232452, + 0.01824132539331913, + 0.06289082020521164, + 0.10704925656318665, + 0.08597312122583389, + 0.07978616654872894, + 0.217638298869133, + 0.2347782850265503, + 0.02393771894276142, + 0.06367851793766022 + ], + "credit_map": [ + 0.01163170300424099, + 0.010444937273859978, + 0.010960428044199944, + 0.019278882071375847, + 0.013976209796965122, + 0.013762317597866058, + 0.012131740339100361, + 0.013840273022651672, + 0.0024977740831673145, + 0.0015720276860520244, + 0.0019928456749767065, + 0.0034922331105917692, + 0.0014752533752471209, + 0.00281248870305717, + 0.0021643079817295074, + 0.0022343958262354136, + 0.00765987578779459, + 0.005976088345050812, + 0.009713222272694111, + 0.009094229899346828, + 0.006459986325353384, + 0.0060501997359097, + 0.009320564568042755, + 0.008616652339696884, + 0.03518996387720108, + 0.004852024372667074, + 0.00675678625702858, + 0.03691098466515541, + 0.005032903980463743, + 0.007048469968140125, + 0.004734765272587538, + 0.006523363757878542, + 0.02121627703309059, + 0.006123360246419907, + 0.009813463315367699, + 0.021229537203907967, + 0.007935131900012493, + 0.008030848577618599, + 0.006270821671932936, + 0.0053536780178546906, + 0.01530085876584053, + 0.005803610198199749, + 0.011397652328014374, + 0.014078487642109394, + 0.008291534148156643, + 0.00962161272764206, + 0.007036666385829449, + 0.008255738765001297, + 0.03135228157043457, + 0.0239228755235672, + 0.03780624642968178, + 0.02824360877275467, + 0.03373240306973457, + 0.0190458744764328, + 0.023852672427892685, + 0.019682331010699272, + 0.008127564564347267, + 0.006705407984554768, + 0.01007632352411747, + 0.09081543982028961, + 0.008682153187692165, + 0.007988846860826015, + 0.09188848733901978, + 0.010494059883058071, + 0.0021333263721317053, + 0.0025901575572788715, + 0.003417666768655181, + 0.0028534599114209414, + 0.002528470940887928, + 0.0024118341971188784, + 0.004037861246615648, + 0.00396494148299098, + 0.006528765894472599, + 0.008077503181993961, + 0.00923533458262682, + 0.008891984820365906, + 0.007914330810308456, + 0.005969500169157982, + 0.008549407124519348, + 0.008511687628924847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.08356603980064392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020615939050912857 + ], + "opposite_logprob_tokens": [ + -3.895616054534912 + ], + "logprob_reward": -0.020615939050912857, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020615939050912857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0543118380010128, + 0.012901615351438522, + 0.0480794794857502, + 0.02154889516532421, + 0.09167464822530746, + 0.031038397923111916, + 0.17866313457489014, + 0.3382256329059601, + 0.06598199903964996, + 0.15757429599761963 + ], + "credit_map": [ + 0.005248395726084709, + 0.006632313132286072, + 0.0054460144601762295, + 0.011814151890575886, + 0.004882736597210169, + 0.0064999680034816265, + 0.006659414153546095, + 0.0071288482286036015, + 0.002039987128227949, + 0.0011627221247181296, + 0.001963678514584899, + 0.001642818795517087, + 0.0013586487621068954, + 0.0018629193073138595, + 0.0014844293473288417, + 0.0013864112552255392, + 0.00604758458212018, + 0.003927271813154221, + 0.005512535572052002, + 0.007919760420918465, + 0.004562648478895426, + 0.004694016650319099, + 0.006829565856605768, + 0.008586093783378601, + 0.0033094300888478756, + 0.0027490968350321054, + 0.0043528396636247635, + 0.0022872276604175568, + 0.002447552280500531, + 0.0021283351816236973, + 0.0020955714862793684, + 0.0021788417361676693, + 0.022375773638486862, + 0.006230783183127642, + 0.0064955586567521095, + 0.026621798053383827, + 0.009004196152091026, + 0.007933255285024643, + 0.006551194004714489, + 0.0064620813354849815, + 0.0035902983509004116, + 0.001666123396717012, + 0.003614750225096941, + 0.005435519851744175, + 0.002728619845584035, + 0.004326676018536091, + 0.0038735431153327227, + 0.0058028679341077805, + 0.025950707495212555, + 0.02021690085530281, + 0.029335493221879005, + 0.02207886055111885, + 0.029277123510837555, + 0.015602322295308113, + 0.020297076553106308, + 0.01590464450418949, + 0.0105109428986907, + 0.009012382477521896, + 0.013652480207383633, + 0.12877775728702545, + 0.011810408905148506, + 0.010578878223896027, + 0.13495826721191406, + 0.018924525007605553, + 0.00571838254109025, + 0.004445286933332682, + 0.010119026526808739, + 0.006235124543309212, + 0.005076427012681961, + 0.010601396672427654, + 0.01384635642170906, + 0.009940002113580704, + 0.016739368438720703, + 0.021557267755270004, + 0.01931173913180828, + 0.021353960037231445, + 0.018579287454485893, + 0.014672976918518543, + 0.022090405225753784, + 0.0232692938297987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.020615939050912857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233200788497925 + ], + "opposite_logprob_tokens": [ + -0.8983200788497925 + ], + "logprob_reward": -0.5233200788497925, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233200788497925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0587841272354126, + 0.09207519143819809, + 0.05683046951889992, + 0.19975119829177856, + 0.03711019083857536, + 0.04570353776216507, + 0.3118491470813751, + 0.12263264507055283, + 0.05731695890426636, + 0.017946502193808556 + ], + "credit_map": [ + 0.008343063294887543, + 0.009744145907461643, + 0.005095312371850014, + 0.008108196780085564, + 0.005499609746038914, + 0.006125994957983494, + 0.007991797290742397, + 0.007876012474298477, + 0.011708739213645458, + 0.009777470491826534, + 0.010523390956223011, + 0.021586453542113304, + 0.006812026724219322, + 0.010430233553051949, + 0.010659209452569485, + 0.010577673092484474, + 0.006957498844712973, + 0.006635116878896952, + 0.01233671884983778, + 0.00873144157230854, + 0.004963031969964504, + 0.005141290836036205, + 0.006710265763103962, + 0.005355111788958311, + 0.0699317455291748, + 0.008431773632764816, + 0.010330883786082268, + 0.07153751701116562, + 0.008990490809082985, + 0.012087778188288212, + 0.007085700519382954, + 0.0113553237169981, + 0.007442245725542307, + 0.0029241470620036125, + 0.005272426176816225, + 0.006888608913868666, + 0.004500578623265028, + 0.003595019457861781, + 0.0034562137443572283, + 0.0030309553258121014, + 0.005566413514316082, + 0.0027946343179792166, + 0.005063733085989952, + 0.009122845716774464, + 0.004149205982685089, + 0.005102436523884535, + 0.006342361681163311, + 0.007561909966170788, + 0.04089558869600296, + 0.03598778694868088, + 0.050048526376485825, + 0.04260947182774544, + 0.05186506360769272, + 0.0274175014346838, + 0.03561915457248688, + 0.027406098321080208, + 0.004333789926022291, + 0.005077897571027279, + 0.005815035197883844, + 0.04919872432947159, + 0.0037926752120256424, + 0.003426091745495796, + 0.04085928574204445, + 0.010129159316420555, + 0.0060228342190384865, + 0.005688342731446028, + 0.007994167506694794, + 0.007151644676923752, + 0.007579004857689142, + 0.00684441439807415, + 0.00771094998344779, + 0.0083256084471941, + 0.0020319558680057526, + 0.0019618484657257795, + 0.0021335415076464415, + 0.0023732143454253674, + 0.003014145651832223, + 0.0016789969522505999, + 0.002399325603619218, + 0.0023534740321338177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.5233200788497925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032087802886963 + ], + "opposite_logprob_tokens": [ + -1.9103208780288696 + ], + "logprob_reward": -0.16032087802886963, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032087802886963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024736395105719566, + 0.01594443991780281, + 0.024199210107326508, + 0.02938980981707573, + 0.06266049295663834, + 0.03472386673092842, + 0.2252965122461319, + 0.45895519852638245, + 0.064736008644104, + 0.05935805290937424 + ], + "credit_map": [ + 0.002916627563536167, + 0.0026518625672906637, + 0.002538684755563736, + 0.003908214624971151, + 0.003216613782569766, + 0.0030203021597117186, + 0.0030101530719548464, + 0.00347393611446023, + 0.0019912701100111008, + 0.0022617140784859657, + 0.0018719404470175505, + 0.002164354780688882, + 0.002007903065532446, + 0.0021840091794729233, + 0.0018857785034924746, + 0.0015774703351780772, + 0.003801986575126648, + 0.0017498087836429477, + 0.002545117400586605, + 0.004877045284956694, + 0.001688331482000649, + 0.002364855259656906, + 0.0033705816604197025, + 0.0038014836609363556, + 0.005858553573489189, + 0.004967078100889921, + 0.0024612208362668753, + 0.0050170947797596455, + 0.0028050800319761038, + 0.004552675411105156, + 0.0017694254638627172, + 0.001958682434633374, + 0.0096193952485919, + 0.0048277899622917175, + 0.008396456018090248, + 0.011650959961116314, + 0.008257001638412476, + 0.0066523002460598946, + 0.007354507222771645, + 0.005902078468352556, + 0.0051008714362978935, + 0.0027135186828672886, + 0.005067057907581329, + 0.004574173595756292, + 0.004416036419570446, + 0.004046327900141478, + 0.004254079889506102, + 0.0045517985709011555, + 0.03317150101065636, + 0.025378145277500153, + 0.037811845541000366, + 0.027380000799894333, + 0.03787589445710182, + 0.019283421337604523, + 0.025235837325453758, + 0.019159866496920586, + 0.021155577152967453, + 0.00958547554910183, + 0.014473283663392067, + 0.12538009881973267, + 0.01743585430085659, + 0.018613819032907486, + 0.18536287546157837, + 0.06694822013378143, + 0.005840481724590063, + 0.0029480180237442255, + 0.014323427341878414, + 0.005647677928209305, + 0.003418510314077139, + 0.006812272593379021, + 0.017106924206018448, + 0.008638698607683182, + 0.00759025476872921, + 0.0060845729894936085, + 0.007466299459338188, + 0.006526843644678593, + 0.006533284205943346, + 0.006182375829666853, + 0.008262677118182182, + 0.010711746290326118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.16032087802886963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.10599221289157867, + 0.07774417847394943, + 0.1128692775964737, + 0.22999097406864166, + 0.18215958774089813, + 0.07689637690782547, + 0.04584037885069847, + 0.08725512027740479, + 0.03469356149435043, + 0.04655839875340462 + ], + "credit_map": [ + 0.012828810140490532, + 0.012949290685355663, + 0.009810741059482098, + 0.02094295248389244, + 0.011956144124269485, + 0.010311431251466274, + 0.013614923693239689, + 0.013577905483543873, + 0.008587184362113476, + 0.01615932397544384, + 0.008077101781964302, + 0.014687214978039265, + 0.004790800157934427, + 0.0077819605357944965, + 0.009414596483111382, + 0.008245989680290222, + 0.014426032081246376, + 0.011744575574994087, + 0.020389188081026077, + 0.016880106180906296, + 0.00929498951882124, + 0.010141505859792233, + 0.01707048900425434, + 0.012922386638820171, + 0.07989140599966049, + 0.009414870291948318, + 0.011187649331986904, + 0.08117181062698364, + 0.010629833675920963, + 0.014940773136913776, + 0.008859649300575256, + 0.013894964009523392, + 0.02075239084661007, + 0.019261378794908524, + 0.019603407010436058, + 0.024448717013001442, + 0.03307882696390152, + 0.02193089947104454, + 0.02435496635735035, + 0.018728984519839287, + 0.012625057250261307, + 0.005783481989055872, + 0.012117328122258186, + 0.008206330239772797, + 0.006618738640099764, + 0.010746932588517666, + 0.008366834372282028, + 0.012431665323674679, + 0.008369959890842438, + 0.002898282138630748, + 0.008634887635707855, + 0.007893639616668224, + 0.005310055799782276, + 0.003259511198848486, + 0.005459361243993044, + 0.004014675971120596, + 0.0032114661298692226, + 0.0033132717944681644, + 0.0038551390171051025, + 0.0348636731505394, + 0.0029182322323322296, + 0.0024663766380399466, + 0.031573474407196045, + 0.005053482484072447, + 0.003853675676509738, + 0.003924311138689518, + 0.0058774990029633045, + 0.004652080126106739, + 0.005474267993122339, + 0.0032365950755774975, + 0.004138988442718983, + 0.0035361438058316708, + 0.004046647809445858, + 0.004893431439995766, + 0.00715258065611124, + 0.0060178013518452644, + 0.0068405806086957455, + 0.004649946931749582, + 0.007042492274194956, + 0.005914913024753332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.006735478527843952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, panning slowly and steadily to the right, locks the viewer's gaze squarely on a pair of hands, etched with deep wrinkles and light brown age spots. Softly illuminated by a side-lit desk lamp, every wrinkle and vein is clearly visible, as if bearing witness to the passage of time. One of the larger, more robust hands (male), with slightly thicker knuckles, firmly holds the base of a dark, softly patinated antique wood. The other, a more delicate hand (female), with its rounded, neatly manicured nails, gently and slowly traces the intricate and delicate carvings on the object's surface with an almost reverent gesture. As the camera pans, we see the fingertips delicately dig into the carved grooves, feeling the subtle undulations of the wood's grain. The light, as it passes over the fingertips, casts a fleeting, soft highlight on the smoothed, raised edges. The background is blurred into a fuzzy warm color, and a corner of the study can be vaguely seen. The whole picture is shrouded in a quiet, focused and affectionate nostalgic atmosphere, as if this pair of silent hands are telling a story spanning a long time.", + "seed": 1026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_9/index_6028_seed_1026_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": -0.23515468016266822, + "vlm_logprob": -0.23515468016266822, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.23515468016266822, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043361492455005646 + ], + "opposite_logprob_tokens": [ + -3.2933614253997803 + ], + "logprob_reward": -0.043361492455005646, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.043361492455005646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144433856010437, + 0.03721125051379204, + 0.025855785235762596, + 0.1055556908249855, + 0.3416450023651123, + 0.0900789350271225, + 0.09556132555007935, + 0.06242328882217407, + 0.04005922004580498, + 0.057175684720277786 + ], + "credit_map": [ + 0.018012631684541702, + 0.044434212148189545, + 0.013250705786049366, + 0.009639807045459747, + 0.014649015851318836, + 0.013170123100280762, + 0.015055845491588116, + 0.01622152142226696, + 0.004081251565366983, + 0.004872526507824659, + 0.0035365165676921606, + 0.006126268766820431, + 0.0031494498252868652, + 0.0038900477811694145, + 0.0055481079034507275, + 0.006007079966366291, + 0.002648615511134267, + 0.0046890126541256905, + 0.002974727423861623, + 0.003416038816794753, + 0.0027735887560993433, + 0.0030398149974644184, + 0.0033059429842978716, + 0.0030080422293394804, + 0.008881010115146637, + 0.019887013360857964, + 0.01205314602702856, + 0.009083818644285202, + 0.01027591247111559, + 0.012909763492643833, + 0.017528725787997246, + 0.014936295337975025, + 0.04926461726427078, + 0.023165389895439148, + 0.016275715082883835, + 0.15263207256793976, + 0.007398820016533136, + 0.007813407108187675, + 0.06695479154586792, + 0.018140185624361038, + 0.01586695946753025, + 0.005084461998194456, + 0.008266007527709007, + 0.03661840036511421, + 0.004269282333552837, + 0.004881950560957193, + 0.008868824690580368, + 0.0062230415642261505, + 0.013066168874502182, + 0.005686250049620867, + 0.014434237964451313, + 0.02047259546816349, + 0.007150097284466028, + 0.010452750138938427, + 0.0112055828794837, + 0.013093631714582443, + 0.00826194416731596, + 0.007541975472122431, + 0.011695971712470055, + 0.009907701052725315, + 0.00895476434379816, + 0.005658133886754513, + 0.005327520426362753, + 0.0050752735696733, + 0.005002512596547604, + 0.00375684117898345, + 0.006897525861859322, + 0.0070972382090985775, + 0.003811273956671357, + 0.0031585379038006067, + 0.005244080908596516, + 0.005091205704957247, + 0.006376585923135281, + 0.005789943039417267, + 0.010251482017338276, + 0.007648447062820196, + 0.0072660986334085464, + 0.004134085029363632, + 0.00807446800172329, + 0.007634571287781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.043361492455005646 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822289079427719 + ], + "opposite_logprob_tokens": [ + -4.018222808837891 + ], + "logprob_reward": -0.01822289079427719, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822289079427719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05110269412398338, + 0.021675599738955498, + 0.013517877086997032, + 0.04403230547904968, + 0.3651677668094635, + 0.052676063030958176, + 0.09906107187271118, + 0.031805094331502914, + 0.06237088516354561, + 0.25859060883522034 + ], + "credit_map": [ + 0.006613748613744974, + 0.0154051398858428, + 0.0050521609373390675, + 0.0035436300095170736, + 0.004204987548291683, + 0.004274214152246714, + 0.006480108480900526, + 0.005528703797608614, + 0.0025750279892235994, + 0.0023330587428063154, + 0.0019656396470963955, + 0.002868024166673422, + 0.002296985359862447, + 0.00260373973287642, + 0.003676712978631258, + 0.0033564120531082153, + 0.0013760968577116728, + 0.003573640715330839, + 0.001396378269419074, + 0.0014949076576158404, + 0.001208876259624958, + 0.0017395350150763988, + 0.0012133439304307103, + 0.0015150981489568949, + 0.003968187607824802, + 0.010630316101014614, + 0.0044403052888810635, + 0.003442885586991906, + 0.004798359237611294, + 0.004760285839438438, + 0.00653085857629776, + 0.005461107939481735, + 0.028190655633807182, + 0.03230540081858635, + 0.05688414350152016, + 0.16001561284065247, + 0.010371760465204716, + 0.009969726204872131, + 0.048136740922927856, + 0.01929372549057007, + 0.011551136150956154, + 0.002265678020194173, + 0.0030827135778963566, + 0.020404640585184097, + 0.0022718231193721294, + 0.0028674297500401735, + 0.0051232087425887585, + 0.005109434016048908, + 0.013608858920633793, + 0.005847185850143433, + 0.014239241369068623, + 0.022753503173589706, + 0.007794549223035574, + 0.010928310453891754, + 0.011616283096373081, + 0.012273131869733334, + 0.0037488301750272512, + 0.003085558535531163, + 0.0037475191056728363, + 0.004572284873574972, + 0.005039895884692669, + 0.003706634510308504, + 0.004223623778671026, + 0.0036807493306696415, + 0.006888718344271183, + 0.005343225784599781, + 0.008255756460130215, + 0.013737315312027931, + 0.004803606774657965, + 0.0061242408119142056, + 0.008259333670139313, + 0.008958692662417889, + 0.035242240875959396, + 0.02511904574930668, + 0.05188552662730217, + 0.03496392071247101, + 0.03203697130084038, + 0.017278321087360382, + 0.03164736554026604, + 0.030417222529649734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.01822289079427719 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8261656165122986 + ], + "opposite_logprob_tokens": [ + -0.5761656165122986 + ], + "logprob_reward": -0.8261656165122986, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261656165122986, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10997291654348373, + 0.027626533061265945, + 0.063121497631073, + 0.1358090043067932, + 0.3458169996738434, + 0.0291733555495739, + 0.1558779627084732, + 0.10984932631254196, + 0.01193311344832182, + 0.0108192702755332 + ], + "credit_map": [ + 0.01460578665137291, + 0.03794471547007561, + 0.009272814728319645, + 0.006503233220428228, + 0.010838286019861698, + 0.007968754507601261, + 0.013714193366467953, + 0.009125126525759697, + 0.003712048288434744, + 0.00478885555639863, + 0.00255067041143775, + 0.0028431089594960213, + 0.0033484133891761303, + 0.0035573572386056185, + 0.0033022253774106503, + 0.0035238531418144703, + 0.006811846047639847, + 0.009748757816851139, + 0.0069218589924275875, + 0.010840225033462048, + 0.007064810022711754, + 0.007202113978564739, + 0.00743957469239831, + 0.007092310115695, + 0.011471237055957317, + 0.023623600602149963, + 0.015039500780403614, + 0.011570173315703869, + 0.014218956232070923, + 0.01710464432835579, + 0.02266876958310604, + 0.020112119615077972, + 0.059420693665742874, + 0.01879986748099327, + 0.013979185372591019, + 0.14698489010334015, + 0.007364547345787287, + 0.008313465863466263, + 0.07330751419067383, + 0.017646832391619682, + 0.0026381872594356537, + 0.003983124624937773, + 0.0038782497867941856, + 0.004592645913362503, + 0.005263327620923519, + 0.0033603468909859657, + 0.0032385033555328846, + 0.002218967769294977, + 0.025622671470046043, + 0.011128192767500877, + 0.025487283244729042, + 0.03305041417479515, + 0.012942507863044739, + 0.01641201414167881, + 0.01645806059241295, + 0.014776813797652721, + 0.01293256040662527, + 0.014311401173472404, + 0.018229089677333832, + 0.016863686963915825, + 0.013449421152472496, + 0.010296555235981941, + 0.011148158460855484, + 0.012618456967175007, + 0.0016511037247255445, + 0.00104471982922405, + 0.0025164741091430187, + 0.0017281228210777044, + 0.0011414078762754798, + 0.000956136267632246, + 0.0014082150300964713, + 0.0014869325095787644, + 0.001158019993454218, + 0.001068417215719819, + 0.0013834780547767878, + 0.0016906942473724484, + 0.0015195583691820502, + 0.0008972096256911755, + 0.001455756020732224, + 0.0016461360501125455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.8261656165122986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812899351119995 + ], + "opposite_logprob_tokens": [ + -1.4062899351119995 + ], + "logprob_reward": -0.2812899351119995, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812899351119995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09051763266324997, + 0.03943730518221855, + 0.02471870370209217, + 0.043728768825531006, + 0.4884011149406433, + 0.026889344677329063, + 0.07477039098739624, + 0.1298709511756897, + 0.0651848167181015, + 0.016480935737490654 + ], + "credit_map": [ + 0.011034278199076653, + 0.02342286892235279, + 0.008011877536773682, + 0.0056654540821909904, + 0.010532454587519169, + 0.012423762120306492, + 0.008606402203440666, + 0.010820535011589527, + 0.004482574760913849, + 0.006595204584300518, + 0.003678729524835944, + 0.003994491416960955, + 0.003501210594549775, + 0.005357567220926285, + 0.006496039684861898, + 0.005331491120159626, + 0.0015346446307376027, + 0.006167554762214422, + 0.0021997445728629827, + 0.005579788703471422, + 0.002229854930192232, + 0.0030230069532990456, + 0.0022140038199722767, + 0.0017701073084026575, + 0.0038953283801674843, + 0.008647104725241661, + 0.0042936778627336025, + 0.00337145384401083, + 0.008655221201479435, + 0.005242902785539627, + 0.005495927762240171, + 0.004127154126763344, + 0.08785965293645859, + 0.03020775131881237, + 0.048582274466753006, + 0.16812945902347565, + 0.013559394516050816, + 0.014492864720523357, + 0.10186471790075302, + 0.023705031722784042, + 0.0030261094216257334, + 0.002685362007468939, + 0.005085944663733244, + 0.004576010629534721, + 0.00223339325748384, + 0.0022358654532581568, + 0.0036595307756215334, + 0.003387132892385125, + 0.00958408135920763, + 0.007072546053677797, + 0.011062593199312687, + 0.015778513625264168, + 0.004758960101753473, + 0.007101499009877443, + 0.008734321221709251, + 0.010677879676222801, + 0.015929972752928734, + 0.012717143632471561, + 0.014206054620444775, + 0.018024465069174767, + 0.017577314749360085, + 0.016316238790750504, + 0.017838379368185997, + 0.017261384055018425, + 0.007784760557115078, + 0.006623122375458479, + 0.01230965182185173, + 0.012609455734491348, + 0.005752103868871927, + 0.005737569183111191, + 0.007812662050127983, + 0.006555497646331787, + 0.002221908653154969, + 0.002473569242283702, + 0.0018877920228987932, + 0.0022196376230567694, + 0.0020997701212763786, + 0.0012214250164106488, + 0.0016968029085546732, + 0.0026600316632539034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.2812899351119995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733465939760208 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733465939760208, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733465939760208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37135204672813416, + 0.03573494032025337, + 0.02930331602692604, + 0.03959763050079346, + 0.15940715372562408, + 0.036785755306482315, + 0.05539996176958084, + 0.20458582043647766, + 0.040247395634651184, + 0.027585972100496292 + ], + "credit_map": [ + 0.048641644418239594, + 0.1258774697780609, + 0.027432512491941452, + 0.02211134508252144, + 0.045694731175899506, + 0.02636716701090336, + 0.045906055718660355, + 0.029321134090423584, + 0.004199519287794828, + 0.005954674445092678, + 0.0032938087824732065, + 0.003964186180382967, + 0.0048238723538815975, + 0.0031925635412335396, + 0.005028499290347099, + 0.0052778166718780994, + 0.0040917727164924145, + 0.003884107107296586, + 0.003295740345492959, + 0.004208426456898451, + 0.0027394965291023254, + 0.003199612256139517, + 0.004086540546268225, + 0.0037976191379129887, + 0.00473370123654604, + 0.006478769704699516, + 0.004177070688456297, + 0.0028393531683832407, + 0.008326984941959381, + 0.004848852287977934, + 0.004925656598061323, + 0.0032672409433871508, + 0.023155469447374344, + 0.014938637614250183, + 0.00970856100320816, + 0.0624285489320755, + 0.003681875765323639, + 0.004718796815723181, + 0.03320462629199028, + 0.007570636458694935, + 0.004532091319561005, + 0.0031094381120055914, + 0.00798891857266426, + 0.008488887920975685, + 0.0029807628598064184, + 0.002663585590198636, + 0.004194674547761679, + 0.002827398246154189, + 0.011190485209226608, + 0.004050281830132008, + 0.01107112132012844, + 0.007275355979800224, + 0.00484612537547946, + 0.005031182896345854, + 0.0057326508685946465, + 0.006202758755534887, + 0.022542214021086693, + 0.018929332494735718, + 0.028622185811400414, + 0.03087616339325905, + 0.03154614940285683, + 0.021974675357341766, + 0.024264037609100342, + 0.025831084698438644, + 0.006574490573257208, + 0.00285772024653852, + 0.008684343658387661, + 0.007437576074153185, + 0.0026490576565265656, + 0.003206541994586587, + 0.004458963871002197, + 0.004378700628876686, + 0.0034722385462373495, + 0.0030741849914193153, + 0.003203649539500475, + 0.006036350503563881, + 0.002434455556795001, + 0.002211114391684532, + 0.003458803053945303, + 0.003695173654705286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.006733465939760208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, panning slowly and steadily to the right, locks the viewer's gaze squarely on a pair of hands, etched with deep wrinkles and light brown age spots. Softly illuminated by a side-lit desk lamp, every wrinkle and vein is clearly visible, as if bearing witness to the passage of time. One of the larger, more robust hands (male), with slightly thicker knuckles, firmly holds the base of a dark, softly patinated antique wood. The other, a more delicate hand (female), with its rounded, neatly manicured nails, gently and slowly traces the intricate and delicate carvings on the object's surface with an almost reverent gesture. As the camera pans, we see the fingertips delicately dig into the carved grooves, feeling the subtle undulations of the wood's grain. The light, as it passes over the fingertips, casts a fleeting, soft highlight on the smoothed, raised edges. The background is blurred into a fuzzy warm color, and a corner of the study can be vaguely seen. The whole picture is shrouded in a quiet, focused and affectionate nostalgic atmosphere, as if this pair of silent hands are telling a story spanning a long time.", + "seed": 1027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_9/index_6028_seed_1027_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": -0.18910846188664437, + "vlm_logprob": -0.18910846188664437, + "vlm_yes_no_margin": 2.700000011920929, + "vlm_token_credit": -0.18910846188664437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05272490158677101 + ], + "opposite_logprob_tokens": [ + -3.052724838256836 + ], + "logprob_reward": -0.05272490158677101, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05272490158677101, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09153662621974945, + 0.027924025431275368, + 0.07589440047740936, + 0.342529833316803, + 0.10416964441537857, + 0.05457331985235214, + 0.03584552928805351, + 0.17628984153270721, + 0.051769766956567764, + 0.03946699947118759 + ], + "credit_map": [ + 0.00664226058870554, + 0.034423355013132095, + 0.004727181978523731, + 0.013147279620170593, + 0.01235306914895773, + 0.006811175495386124, + 0.007908842526376247, + 0.005523460917174816, + 0.0038235620595514774, + 0.003580731339752674, + 0.0022690813057124615, + 0.004081806167960167, + 0.00334833562374115, + 0.0026244078762829304, + 0.002730970736593008, + 0.005465130787342787, + 0.008218558505177498, + 0.00911868829280138, + 0.01228255219757557, + 0.01990516670048237, + 0.005447913892567158, + 0.004824982024729252, + 0.007574053481221199, + 0.008522490039467812, + 0.117740698158741, + 0.08036673069000244, + 0.015240649692714214, + 0.06602045148611069, + 0.015382097102701664, + 0.016633039340376854, + 0.017946314066648483, + 0.013199838809669018, + 0.00951287243515253, + 0.006692334543913603, + 0.032057251781225204, + 0.03145122900605202, + 0.00394509406760335, + 0.0056913867592811584, + 0.0071549867279827595, + 0.007664485368877649, + 0.006198479328304529, + 0.007303979247808456, + 0.007506938185542822, + 0.007467254996299744, + 0.005103251431137323, + 0.0048289792612195015, + 0.008659092709422112, + 0.007505341432988644, + 0.002813127590343356, + 0.004706773441284895, + 0.00456692511215806, + 0.004610075615346432, + 0.0039040048141032457, + 0.005686130840331316, + 0.004582026973366737, + 0.004976463969796896, + 0.007411010097712278, + 0.0067696766927838326, + 0.06306707113981247, + 0.06822331249713898, + 0.00354043603874743, + 0.006696484517306089, + 0.011613838374614716, + 0.008968009613454342, + 0.00514636654406786, + 0.0055836946703493595, + 0.0072777168825268745, + 0.00864490307867527, + 0.0048179966397583485, + 0.004221267532557249, + 0.008599012158811092, + 0.007478808984160423, + 0.004563863389194012, + 0.003054344793781638, + 0.007932953536510468, + 0.009932137094438076, + 0.0025452144909650087, + 0.0031029577367007732, + 0.005618880968540907, + 0.0027166453655809164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.05272490158677101 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216102659702301 + ], + "opposite_logprob_tokens": [ + -4.26421594619751 + ], + "logprob_reward": -0.014216102659702301, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216102659702301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09943459182977676, + 0.019120443612337112, + 0.05883903056383133, + 0.130394846200943, + 0.15036965906620026, + 0.022159328684210777, + 0.03583328798413277, + 0.36197933554649353, + 0.06942299008369446, + 0.05244646966457367 + ], + "credit_map": [ + 0.007004009559750557, + 0.027823995798826218, + 0.004865357186645269, + 0.027227697893977165, + 0.008806522935628891, + 0.005854933522641659, + 0.008938304148614407, + 0.008913770318031311, + 0.0025928188115358353, + 0.0026518432423472404, + 0.0023617721162736416, + 0.0023431708104908466, + 0.003076513297855854, + 0.0017666694475337863, + 0.0019260544795542955, + 0.00240160059183836, + 0.008788452483713627, + 0.006920835934579372, + 0.007963655516505241, + 0.01776808872818947, + 0.004282929003238678, + 0.0031024699565023184, + 0.005043254233896732, + 0.004969344008713961, + 0.044585220515728, + 0.026275403797626495, + 0.0058286190032958984, + 0.028713980689644814, + 0.005338112823665142, + 0.005965461488813162, + 0.007815875113010406, + 0.005872171372175217, + 0.011286512017250061, + 0.01059051789343357, + 0.04810271039605141, + 0.04329722747206688, + 0.007328573614358902, + 0.007334602065384388, + 0.01183705311268568, + 0.010592455975711346, + 0.0021499469876289368, + 0.0028241141699254513, + 0.0031931791454553604, + 0.003221663646399975, + 0.0018423076253384352, + 0.002058489713817835, + 0.003164061112329364, + 0.003705565817654133, + 0.003337970469146967, + 0.0035975072532892227, + 0.005535414908081293, + 0.00439666910097003, + 0.003773615462705493, + 0.003901517018675804, + 0.004644195083528757, + 0.006646397523581982, + 0.013968052342534065, + 0.01221144013106823, + 0.13689644634723663, + 0.14670316874980927, + 0.005466805770993233, + 0.009567433968186378, + 0.023209789767861366, + 0.013956193812191486, + 0.007152595091611147, + 0.007501483894884586, + 0.012090570293366909, + 0.012382292188704014, + 0.006619011051952839, + 0.004410779103636742, + 0.011528965085744858, + 0.007737293839454651, + 0.005934212822467089, + 0.004617704544216394, + 0.00939698051661253, + 0.011288217268884182, + 0.0036403306294232607, + 0.005357223562896252, + 0.007408779580146074, + 0.004803018644452095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.014216102659702301 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34859246015548706 + ], + "opposite_logprob_tokens": [ + -1.2235925197601318 + ], + "logprob_reward": -0.34859246015548706, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34859246015548706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11867571622133255, + 0.09562382102012634, + 0.04817421734333038, + 0.054113756865262985, + 0.2807459533214569, + 0.14644503593444824, + 0.03231988474726677, + 0.1088026836514473, + 0.08251544088125229, + 0.03258344903588295 + ], + "credit_map": [ + 0.008763152174651623, + 0.03270536661148071, + 0.006899463012814522, + 0.02626628614962101, + 0.007262105122208595, + 0.009415576234459877, + 0.012264576740562916, + 0.01509920135140419, + 0.008203045465052128, + 0.017727183178067207, + 0.01411394402384758, + 0.01918192394077778, + 0.009101098403334618, + 0.006130245514214039, + 0.01152003463357687, + 0.009646345861256123, + 0.005768036004155874, + 0.005622461903840303, + 0.006781397853046656, + 0.013685972429811954, + 0.004411620553582907, + 0.002713784109801054, + 0.004758862778544426, + 0.004432088229805231, + 0.013176926411688328, + 0.0033843391574919224, + 0.004935331642627716, + 0.013171201571822166, + 0.0036073196679353714, + 0.0030710739083588123, + 0.0070592197589576244, + 0.005708349868655205, + 0.02449834905564785, + 0.015965845435857773, + 0.09853289276361465, + 0.09395632892847061, + 0.007589058484882116, + 0.013818349689245224, + 0.016774248331785202, + 0.009610908105969429, + 0.01812640391290188, + 0.021763835102319717, + 0.018810303881764412, + 0.016820091754198074, + 0.012566485442221165, + 0.015076452866196632, + 0.02026955597102642, + 0.02301192097365856, + 0.0036771015729755163, + 0.0035940781235694885, + 0.006542841903865337, + 0.005407851189374924, + 0.00260121445171535, + 0.0028219292871654034, + 0.003739722538739443, + 0.00393514521420002, + 0.004211399704217911, + 0.004863623529672623, + 0.0395483560860157, + 0.042630501091480255, + 0.002445581601932645, + 0.003374372376129031, + 0.007362111937254667, + 0.004366752225905657, + 0.004673087038099766, + 0.007059610914438963, + 0.028857175260782242, + 0.02647279016673565, + 0.003210949944332242, + 0.003162492299452424, + 0.005252045579254627, + 0.0038272980600595474, + 0.0035788328386843204, + 0.0026471868623048067, + 0.006463157944381237, + 0.009229318238794804, + 0.002094479277729988, + 0.0026013192255049944, + 0.0037822406738996506, + 0.002186919329687953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.34859246015548706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5232759714126587 + ], + "opposite_logprob_tokens": [ + -0.8982759714126587 + ], + "logprob_reward": -0.5232759714126587, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5232759714126587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07478722184896469, + 0.012504443526268005, + 0.035063017159700394, + 0.18210779130458832, + 0.27297866344451904, + 0.022498229518532753, + 0.01890883781015873, + 0.21973101794719696, + 0.12674088776111603, + 0.034679848700761795 + ], + "credit_map": [ + 0.005018197000026703, + 0.023230424150824547, + 0.005185124929994345, + 0.017457695677876472, + 0.0053831529803574085, + 0.00429318705573678, + 0.006484336219727993, + 0.007735102903097868, + 0.0013472664868459105, + 0.001325407880358398, + 0.0014597506960853934, + 0.0014816292095929384, + 0.0014278143644332886, + 0.00118454999756068, + 0.001769542577676475, + 0.0025084828957915306, + 0.0039826687425374985, + 0.004000470973551273, + 0.005914508365094662, + 0.007265884894877672, + 0.0031930950935930014, + 0.0023802025243639946, + 0.0037791416980326176, + 0.004547045566141605, + 0.06789099425077438, + 0.030118348076939583, + 0.007732077036052942, + 0.04704641178250313, + 0.005734443664550781, + 0.007980669848620892, + 0.00996280089020729, + 0.0056420559994876385, + 0.023773569613695145, + 0.016686055809259415, + 0.09445970505475998, + 0.0912499874830246, + 0.007026710547506809, + 0.013007136061787605, + 0.016292160376906395, + 0.010483314283192158, + 0.0033194462303072214, + 0.002434922382235527, + 0.0023695421405136585, + 0.0036953873932361603, + 0.00206141360104084, + 0.0019376231357455254, + 0.003177384380251169, + 0.0035025100223720074, + 0.0018025204772129655, + 0.001912125269882381, + 0.00251848716288805, + 0.0017668192740529776, + 0.002924765693023801, + 0.002699267351999879, + 0.0022532944567501545, + 0.003031559754163027, + 0.008789291605353355, + 0.008546397089958191, + 0.08204802125692368, + 0.08728896826505661, + 0.00464367913082242, + 0.006826515309512615, + 0.014049780555069447, + 0.007538361009210348, + 0.0074417563155293465, + 0.00926665123552084, + 0.038087621331214905, + 0.03308609500527382, + 0.006836778484284878, + 0.008319330401718616, + 0.011449326761066914, + 0.012253347784280777, + 0.004280471708625555, + 0.0034650126472115517, + 0.007301683072000742, + 0.00962140318006277, + 0.002623822772875428, + 0.001989301759749651, + 0.00304565136320889, + 0.0023525019641965628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.5232759714126587 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two hands visible in the video, one appearing male with thicker knuckles and one appearing female with rounded nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand firmly holding the base of a dark, antique wood object while the other traces its carved surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands and object clearly illuminated by a side-lit desk lamp, revealing deep wrinkles, veins, and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly and steadily to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a fuzzy warm color with only a vague corner of a study visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732873618602753 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732873618602753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732873618602753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11816468089818954, + 0.0760396346449852, + 0.018032150343060493, + 0.3630481958389282, + 0.2665136158466339, + 0.024653062224388123, + 0.01554189808666706, + 0.060316652059555054, + 0.039879538118839264, + 0.01781056635081768 + ], + "credit_map": [ + 0.010122355073690414, + 0.0175645649433136, + 0.007866913452744484, + 0.02718602679669857, + 0.01984482631087303, + 0.009754586033523083, + 0.012538556009531021, + 0.013286850415170193, + 0.008652654476463795, + 0.011308431625366211, + 0.009350758045911789, + 0.013799159787595272, + 0.006938835605978966, + 0.006195570342242718, + 0.009944690391421318, + 0.00984952226281166, + 0.0021652677096426487, + 0.0019000967731699347, + 0.0019200083333998919, + 0.004156163427978754, + 0.0021129311062395573, + 0.0013755963882431388, + 0.002139837946742773, + 0.002262248657643795, + 0.10962697863578796, + 0.10533394664525986, + 0.015637369826436043, + 0.05083413049578667, + 0.019360244274139404, + 0.019534554332494736, + 0.021329423412680626, + 0.021391499787569046, + 0.021688207983970642, + 0.016030825674533844, + 0.09247657656669617, + 0.08614778518676758, + 0.009245260618627071, + 0.013430764898657799, + 0.018379440531134605, + 0.009114748798310757, + 0.0026503466069698334, + 0.002763600554317236, + 0.002862260676920414, + 0.0034538956824690104, + 0.00268076010979712, + 0.0027085738256573677, + 0.0036716246977448463, + 0.0038619963452219963, + 0.001627208199352026, + 0.0019417969742789865, + 0.0026662261225283146, + 0.0025956598110497, + 0.0012958936858922243, + 0.0015405976446345448, + 0.0016996143385767937, + 0.0021748992148786783, + 0.002797951688989997, + 0.00395210599526763, + 0.019210049882531166, + 0.022066807374358177, + 0.0023558891844004393, + 0.002519583562389016, + 0.004308121744543314, + 0.003106137039139867, + 0.0021539926528930664, + 0.0025870322715491056, + 0.014494241215288639, + 0.012901431880891323, + 0.0015417581889778376, + 0.0015700558433309197, + 0.0024089019279927015, + 0.0022221191320568323, + 0.0021262068767100573, + 0.0013044020161032677, + 0.003355225082486868, + 0.004758032970130444, + 0.0012752492912113667, + 0.0017461596289649606, + 0.0017234018305316567, + 0.001521887257695198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.006732873618602753 + } + ], + "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/0000028/group_0/index_tensor([7718, 7718])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_0/index_tensor([7718, 7718])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..59a34e9366b404351043ec440bd43cd2a8ed0080 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_0/index_tensor([7718, 7718])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up captures the serene face of a surreal wax figure of a woman, her features reminiscent of classical sculpture. Her dark, wavy hair is meticulously detailed, each strand catching the light. Under a dramatic overhead spotlight, her form is sculpted from a profound black void, her skin possessing a soft, waxy sheen, punctuated by subtle, painterly imperfections that enhance its realism. Then, the camera begins a slow, tight push-in, its movements synchronizing with the figure's decay. The process begins almost imperceptibly, with a tiny crack appearing across her high cheekbone, rapidly spreading across the smooth surface. Fine, sand-like particles begin to drip from the crack, catching the harsh light as they fall. As the camera inexorably moves in, the disintegration accelerates; the bridge of her perfect nose dents and erodes, and a large section of her cheek peels away, dissolving in mid-air into a cloud of fine black dust, revealing a hollow interior. The entire structure of her face begins to fail, the once-peaceful lips drooping and crumbling as the camera inexorably advances, focusing on the catastrophic, silent collapse until the frame is filled with the chaotic movement of disintegrating wax and swirling dust, creating a deeply unsettling atmosphere of tragedy. Beauty transformed into devastation.", + "seed": 23896, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_0/index_7718_seed_23896_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": -0.016260741651058196, + "vlm_logprob": -0.016260741651058196, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.016260741651058196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006439412012696266 + ], + "opposite_logprob_tokens": [ + -6.756439208984375 + ], + "logprob_reward": -0.006439412012696266, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.006439412012696266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12332067638635635, + 0.03748076781630516, + 0.036240849643945694, + 0.0410473607480526, + 0.10615021735429764, + 0.11472029983997345, + 0.0711391344666481, + 0.17235548794269562, + 0.1672431379556656, + 0.1303020864725113 + ], + "credit_map": [ + 0.011695900931954384, + 0.01035569328814745, + 0.017956813797354698, + 0.01718229055404663, + 0.012311923317611217, + 0.01892000436782837, + 0.019330080598592758, + 0.015567968599498272, + 0.0034523617941886187, + 0.0037606218829751015, + 0.007020116318017244, + 0.004344491753727198, + 0.0029275519773364067, + 0.0066217402927577496, + 0.005452062003314495, + 0.003901819931343198, + 0.004808950703591108, + 0.002685069339349866, + 0.006110779009759426, + 0.0035107440780848265, + 0.003989425487816334, + 0.005440059583634138, + 0.0056046065874397755, + 0.004091212991625071, + 0.0043037692084908485, + 0.0034878242295235395, + 0.005237901117652655, + 0.007457195781171322, + 0.004634474404156208, + 0.004871276207268238, + 0.007304638158529997, + 0.003750282572582364, + 0.012762976810336113, + 0.00997732300311327, + 0.016405785456299782, + 0.01508814562112093, + 0.008617784827947617, + 0.018513182178139687, + 0.013417675159871578, + 0.011367355473339558, + 0.013721239753067493, + 0.009323450736701488, + 0.014720692299306393, + 0.013066902756690979, + 0.013793151825666428, + 0.023060092702507973, + 0.015327702276408672, + 0.011707068420946598, + 0.013246078975498676, + 0.004774162080138922, + 0.011747690849006176, + 0.005129692144691944, + 0.004523369483649731, + 0.009287838824093342, + 0.013216299936175346, + 0.009213997982442379, + 0.025392156094312668, + 0.013616004958748817, + 0.021040717139840126, + 0.018067093566060066, + 0.01498719397932291, + 0.035479094833135605, + 0.02046147733926773, + 0.02331174723803997, + 0.02475275844335556, + 0.013384847901761532, + 0.035856351256370544, + 0.015447728335857391, + 0.011903010308742523, + 0.025675788521766663, + 0.023941989988088608, + 0.01628067158162594, + 0.015358849428594112, + 0.012366397306323051, + 0.01298600435256958, + 0.016413303092122078, + 0.016918392851948738, + 0.021116062998771667, + 0.017425166442990303, + 0.017717907205224037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.006439412012696266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078690066933632 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.016078690066933632, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078690066933632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03546049818396568, + 0.7356336712837219, + 0.014654000289738178, + 0.020247850567102432, + 0.0220017246901989, + 0.024447234347462654, + 0.029013603925704956, + 0.031579386442899704, + 0.03289435803890228, + 0.054067641496658325 + ], + "credit_map": [ + 0.0030628684908151627, + 0.003162908600643277, + 0.0053053442388772964, + 0.004469522275030613, + 0.003612576751038432, + 0.00623669009655714, + 0.006386195309460163, + 0.003224393352866173, + 0.08451488614082336, + 0.1252720057964325, + 0.14934372901916504, + 0.07501032948493958, + 0.06679631769657135, + 0.11886104941368103, + 0.06675031781196594, + 0.04908502474427223, + 0.0015968317165970802, + 0.0013531636213883758, + 0.0018207253888249397, + 0.0016491477144882083, + 0.001715010148473084, + 0.0019398917211219668, + 0.00266982801258564, + 0.0019094017334282398, + 0.001979667693376541, + 0.0023400774225592613, + 0.002902199514210224, + 0.002991961082443595, + 0.0018875047098845243, + 0.003785176435485482, + 0.0026063951663672924, + 0.00175486842636019, + 0.002483396790921688, + 0.001775209209881723, + 0.003021932439878583, + 0.0029572313651442528, + 0.0024508917704224586, + 0.003945673815906048, + 0.0029051415622234344, + 0.0024622478522360325, + 0.002412268193438649, + 0.0020551071502268314, + 0.0033522944431751966, + 0.002787295263260603, + 0.002635503886267543, + 0.00495524238795042, + 0.003758795093744993, + 0.0024907286278903484, + 0.003730716183781624, + 0.0026070522144436836, + 0.004227127879858017, + 0.003291754052042961, + 0.0018970322562381625, + 0.00413259444758296, + 0.00590261397883296, + 0.0032247130293399096, + 0.004510033410042524, + 0.0017620495054870844, + 0.0037119865883141756, + 0.002584895584732294, + 0.0025661031249910593, + 0.0082611795514822, + 0.0036849079187959433, + 0.004498229827731848, + 0.005026208236813545, + 0.0023621588479727507, + 0.003173570381477475, + 0.0056745014153420925, + 0.0015827747993171215, + 0.004841854330152273, + 0.006635752506554127, + 0.003597537288442254, + 0.0055978926829993725, + 0.004281496163457632, + 0.00518193980678916, + 0.00753165315836668, + 0.0061750151216983795, + 0.010094555094838142, + 0.007393314503133297, + 0.007811775431036949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.016078690066933632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535235844552517 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.012535235844552517, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535235844552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2706918716430664, + 0.5558377504348755, + 0.01599925011396408, + 0.012107089161872864, + 0.019538553431630135, + 0.025814518332481384, + 0.025690440088510513, + 0.02000437118113041, + 0.018533270806074142, + 0.035782795399427414 + ], + "credit_map": [ + 0.02572493627667427, + 0.04003769904375076, + 0.04714944586157799, + 0.032496578991413116, + 0.023168262094259262, + 0.05069781839847565, + 0.030118918046355247, + 0.02129824459552765, + 0.06216363236308098, + 0.09228198230266571, + 0.11488910019397736, + 0.05335493013262749, + 0.05021555721759796, + 0.09633409231901169, + 0.04984105005860329, + 0.03675747662782669, + 0.0024733995087444782, + 0.001177232596091926, + 0.0025020954199135303, + 0.001570972497574985, + 0.0011781449429690838, + 0.003448848845437169, + 0.0016977128107100725, + 0.0019508455879986286, + 0.0014629444340243936, + 0.0010449826950207353, + 0.0016968284035101533, + 0.002248006872832775, + 0.0009231199510395527, + 0.001755038509145379, + 0.001641189563088119, + 0.0013349803630262613, + 0.002826913958415389, + 0.0012846461031585932, + 0.0027889239136129618, + 0.002736106049269438, + 0.00180709024425596, + 0.003323212033137679, + 0.0023851299192756414, + 0.0023865324910730124, + 0.002780735492706299, + 0.0020673726685345173, + 0.003733704797923565, + 0.002632475458085537, + 0.0025205297861248255, + 0.005283809266984463, + 0.0041248369961977005, + 0.002671056892722845, + 0.004257768392562866, + 0.002261623740196228, + 0.0035991796758025885, + 0.0028505860827863216, + 0.002051396993920207, + 0.0031595174223184586, + 0.004557741340249777, + 0.002952628768980503, + 0.002676289528608322, + 0.0018163585336878896, + 0.003019250463694334, + 0.0028115648310631514, + 0.0015188016695901752, + 0.0035649535711854696, + 0.0024256857577711344, + 0.002171470317989588, + 0.0026829931885004044, + 0.001700537744909525, + 0.0024010343477129936, + 0.001726567861624062, + 0.001672617974691093, + 0.003017980605363846, + 0.002666252199560404, + 0.002665287582203746, + 0.004106052685528994, + 0.002370705595239997, + 0.003437054343521595, + 0.004793317522853613, + 0.004186540376394987, + 0.006487085949629545, + 0.005832090973854065, + 0.004569950979202986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.012535235844552517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033705245703458786 + ], + "opposite_logprob_tokens": [ + -3.408705234527588 + ], + "logprob_reward": -0.033705245703458786, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033705245703458786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15577200055122375, + 0.19177943468093872, + 0.08781413733959198, + 0.032633114606142044, + 0.049815818667411804, + 0.04367418959736824, + 0.160211980342865, + 0.05309859290719032, + 0.10366798937320709, + 0.12153269350528717 + ], + "credit_map": [ + 0.014832050539553165, + 0.021998999640345573, + 0.03022318333387375, + 0.016365455463528633, + 0.01169256679713726, + 0.03209718316793442, + 0.01647864654660225, + 0.012083900161087513, + 0.022431958466768265, + 0.032220687717199326, + 0.036307673901319504, + 0.02101578377187252, + 0.01773861236870289, + 0.02970610186457634, + 0.019008221104741096, + 0.013350388966500759, + 0.011388255283236504, + 0.008060697466135025, + 0.013918302021920681, + 0.009358003735542297, + 0.00797946099191904, + 0.0168202705681324, + 0.010801027528941631, + 0.00948812160640955, + 0.004125300794839859, + 0.0028367803897708654, + 0.005651822313666344, + 0.0029719723388552666, + 0.0026336817536503077, + 0.008039501495659351, + 0.003324958961457014, + 0.003049098188057542, + 0.008043588139116764, + 0.003409984055906534, + 0.0076752640306949615, + 0.006017219740897417, + 0.0038165394216775894, + 0.008107705041766167, + 0.00661654444411397, + 0.006128975655883551, + 0.005697275046259165, + 0.003791103372350335, + 0.005051461514085531, + 0.005444302689284086, + 0.007064242381602526, + 0.004950474016368389, + 0.00749459071084857, + 0.004180739633738995, + 0.03403135761618614, + 0.011297081597149372, + 0.0318397581577301, + 0.01175095234066248, + 0.00764873344451189, + 0.01685614325106144, + 0.026784168556332588, + 0.020003795623779297, + 0.006164887920022011, + 0.004629554692655802, + 0.006121049635112286, + 0.0058875479735434055, + 0.004112389869987965, + 0.01203783880919218, + 0.008037523366510868, + 0.0061078062281012535, + 0.014288539998233318, + 0.008534295484423637, + 0.019006287679076195, + 0.009605728089809418, + 0.0075167091563344, + 0.01848667487502098, + 0.015838434919714928, + 0.010391322895884514, + 0.01852063275873661, + 0.013159560970962048, + 0.01301894336938858, + 0.016557322815060616, + 0.009263838641345501, + 0.02029101364314556, + 0.015335027128458023, + 0.01538635790348053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.033705245703458786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545124627649784 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545124627649784, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545124627649784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09026084840297699, + 0.023512262850999832, + 0.09350112080574036, + 0.06978500634431839, + 0.05190477520227432, + 0.04781487584114075, + 0.07295342534780502, + 0.09985014796257019, + 0.14331968128681183, + 0.30709776282310486 + ], + "credit_map": [ + 0.007445625029504299, + 0.007417469285428524, + 0.014019763097167015, + 0.013146016746759415, + 0.011110905557870865, + 0.012637566775083542, + 0.016512081027030945, + 0.007971431128680706, + 0.0023282321635633707, + 0.0030048242770135403, + 0.0033722505904734135, + 0.002768038772046566, + 0.0021570823155343533, + 0.0042610494419932365, + 0.0029526930302381516, + 0.002668095286935568, + 0.009687335230410099, + 0.007469107862561941, + 0.014518090523779392, + 0.01088614109903574, + 0.010381946340203285, + 0.01668698526918888, + 0.01450760755687952, + 0.009363914839923382, + 0.0053909048438072205, + 0.0062197120860219, + 0.009470215067267418, + 0.010735963471233845, + 0.00908418744802475, + 0.01321383472532034, + 0.010209337808191776, + 0.005460857879370451, + 0.0075235916301608086, + 0.004017706494778395, + 0.009171664714813232, + 0.006323229055851698, + 0.003907349426299334, + 0.007749082054942846, + 0.006824210751801729, + 0.006387945264577866, + 0.005232854746282101, + 0.004805595148354769, + 0.005904088728129864, + 0.005457447841763496, + 0.006069382186979055, + 0.007491378579288721, + 0.00753199914470315, + 0.005322136916220188, + 0.012983029708266258, + 0.005947333760559559, + 0.01443960890173912, + 0.00595566863194108, + 0.005012060981243849, + 0.012277702800929546, + 0.008796454407274723, + 0.007541572209447622, + 0.009031360037624836, + 0.009315425530076027, + 0.017608581110835075, + 0.014519823715090752, + 0.008627241477370262, + 0.016888299956917763, + 0.013728379271924496, + 0.010131043381989002, + 0.014416382648050785, + 0.011183315888047218, + 0.02233525738120079, + 0.01629980467259884, + 0.010558143258094788, + 0.030551956966519356, + 0.025570770725607872, + 0.0124040637165308, + 0.03571665659546852, + 0.019318023696541786, + 0.031058745458722115, + 0.04181409254670143, + 0.03657161816954613, + 0.05517984926700592, + 0.04819072037935257, + 0.03924808278679848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.012545124627649784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures the serene face of a surreal wax figure of a woman, her features reminiscent of classical sculpture. Her dark, wavy hair is meticulously detailed, each strand catching the light. Under a dramatic overhead spotlight, her form is sculpted from a profound black void, her skin possessing a soft, waxy sheen, punctuated by subtle, painterly imperfections that enhance its realism. Then, the camera begins a slow, tight push-in, its movements synchronizing with the figure's decay. The process begins almost imperceptibly, with a tiny crack appearing across her high cheekbone, rapidly spreading across the smooth surface. Fine, sand-like particles begin to drip from the crack, catching the harsh light as they fall. As the camera inexorably moves in, the disintegration accelerates; the bridge of her perfect nose dents and erodes, and a large section of her cheek peels away, dissolving in mid-air into a cloud of fine black dust, revealing a hollow interior. The entire structure of her face begins to fail, the once-peaceful lips drooping and crumbling as the camera inexorably advances, focusing on the catastrophic, silent collapse until the frame is filled with the chaotic movement of disintegrating wax and swirling dust, creating a deeply unsettling atmosphere of tragedy. Beauty transformed into devastation.", + "seed": 23897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_0/index_7718_seed_23897_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": -0.04796196483075619, + "vlm_logprob": -0.04796196483075619, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.04796196483075619, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009913614019751549 + ], + "opposite_logprob_tokens": [ + -5.259913444519043 + ], + "logprob_reward": -0.009913614019751549, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009913614019751549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02347003109753132, + 0.11585620790719986, + 0.02967934124171734, + 0.019964467734098434, + 0.048329222947359085, + 0.4282894730567932, + 0.032857682555913925, + 0.0890733152627945, + 0.18600039184093475, + 0.026479819789528847 + ], + "credit_map": [ + 0.002662265906110406, + 0.00259663094766438, + 0.0026393861044198275, + 0.0036405029240995646, + 0.0018911751685664058, + 0.003050291445106268, + 0.00449244724586606, + 0.002497331937775016, + 0.007698919624090195, + 0.0035531113389879465, + 0.036644816398620605, + 0.0065996539779007435, + 0.004442088771611452, + 0.008819406852126122, + 0.009735241532325745, + 0.03836296871304512, + 0.003987984731793404, + 0.0018680093344300985, + 0.0071659632958471775, + 0.002120981691405177, + 0.0015935705741867423, + 0.0021233411971479654, + 0.0032142107374966145, + 0.007605281192809343, + 0.002185021759942174, + 0.002135779708623886, + 0.002628256333991885, + 0.002875599777325988, + 0.0019487729296088219, + 0.0025521519128233194, + 0.00307479384355247, + 0.0025640917010605335, + 0.005699816159904003, + 0.0043706647120416164, + 0.007959232665598392, + 0.004708786960691214, + 0.003039454109966755, + 0.006328373681753874, + 0.006705451291054487, + 0.009517447091639042, + 0.02645161934196949, + 0.012693114578723907, + 0.17341886460781097, + 0.019687004387378693, + 0.010890338569879532, + 0.017172642052173615, + 0.014886349439620972, + 0.15308955311775208, + 0.0029343690257519484, + 0.0032206447795033455, + 0.006412842310965061, + 0.0031814302783459425, + 0.003100442700088024, + 0.003938201814889908, + 0.003608381375670433, + 0.006461371202021837, + 0.012716757133603096, + 0.011297686956822872, + 0.010755607858300209, + 0.009686985984444618, + 0.016616718843579292, + 0.008332009427249432, + 0.010798105970025063, + 0.008869445882737637, + 0.015390051528811455, + 0.0249741543084383, + 0.019711771979928017, + 0.02608541026711464, + 0.01918453723192215, + 0.021233154460787773, + 0.02873210608959198, + 0.03068920411169529, + 0.002313850214704871, + 0.0028246326837688684, + 0.002863249508664012, + 0.006308841984719038, + 0.0024634813889861107, + 0.002948497422039509, + 0.0032843388617038727, + 0.0034729274921119213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.009913614019751549 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602744072675705 + ], + "opposite_logprob_tokens": [ + -3.8956027030944824 + ], + "logprob_reward": -0.020602744072675705, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602744072675705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05922611802816391, + 0.32987943291664124, + 0.031698957085609436, + 0.018611248582601547, + 0.05634438991546631, + 0.2475196123123169, + 0.0243803933262825, + 0.08882558345794678, + 0.08620685338973999, + 0.05730744078755379 + ], + "credit_map": [ + 0.007366205099970102, + 0.0055321636609733105, + 0.009806440211832523, + 0.007140013854950666, + 0.0037289317697286606, + 0.009352941997349262, + 0.00776305329054594, + 0.008536366745829582, + 0.02351205423474312, + 0.010072528384625912, + 0.10809820890426636, + 0.017480839043855667, + 0.012037639506161213, + 0.026907965540885925, + 0.019496656954288483, + 0.11227355152368546, + 0.004835235886275768, + 0.0018330779857933521, + 0.007231998257339001, + 0.003436426632106304, + 0.0018244190141558647, + 0.0024865055456757545, + 0.0026529342867434025, + 0.007398359943181276, + 0.002755073830485344, + 0.0018255552276968956, + 0.0024384360294789076, + 0.002634636126458645, + 0.0026646640617400408, + 0.0015942638274282217, + 0.0025376637931913137, + 0.0021609561517834663, + 0.005324776284396648, + 0.0038898037746548653, + 0.013556994497776031, + 0.004706840496510267, + 0.0024873036891222, + 0.0060023837722837925, + 0.004784735385328531, + 0.015591551549732685, + 0.016820289194583893, + 0.009282466024160385, + 0.09072762727737427, + 0.014016469940543175, + 0.009107314981520176, + 0.012703072279691696, + 0.011980248615145683, + 0.0828821212053299, + 0.0020940667018294334, + 0.0023983188439160585, + 0.00376247800886631, + 0.0029521912802010775, + 0.002822147449478507, + 0.0018932399107143283, + 0.0030133742839097977, + 0.0054445769637823105, + 0.010805304162204266, + 0.0109688276425004, + 0.013234555721282959, + 0.010428392328321934, + 0.015568696893751621, + 0.007329410407692194, + 0.01150944922119379, + 0.008980950340628624, + 0.006932236719876528, + 0.010671768337488174, + 0.010268911719322205, + 0.011664074845612049, + 0.009873459115624428, + 0.008648124523460865, + 0.01292931754142046, + 0.01521897129714489, + 0.004656454082578421, + 0.0050498745404183865, + 0.006468358915299177, + 0.018425775691866875, + 0.005871565546840429, + 0.004139970522373915, + 0.006332947872579098, + 0.006362492684274912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.020602744072675705 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292329430580139 + ], + "opposite_logprob_tokens": [ + -2.2379233837127686 + ], + "logprob_reward": -0.11292329430580139, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292329430580139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09847693890333176, + 0.16316616535186768, + 0.09562379121780396, + 0.03332168236374855, + 0.024353381246328354, + 0.3354974389076233, + 0.04729720950126648, + 0.03566530719399452, + 0.11909685283899307, + 0.047501277178525925 + ], + "credit_map": [ + 0.011835131794214249, + 0.007319436874240637, + 0.021751312538981438, + 0.010740372352302074, + 0.007694330532103777, + 0.012857677415013313, + 0.013565015979111195, + 0.01271365862339735, + 0.018618499860167503, + 0.00704739335924387, + 0.04525604844093323, + 0.010011091828346252, + 0.006560575682669878, + 0.017051490023732185, + 0.011848659254610538, + 0.04677240177989006, + 0.014937377534806728, + 0.005138513632118702, + 0.024224666878581047, + 0.006767907179892063, + 0.00598854711279273, + 0.006219508126378059, + 0.0067857918329536915, + 0.025561481714248657, + 0.004992016591131687, + 0.002297867788001895, + 0.004097421653568745, + 0.0060152411460876465, + 0.004077705554664135, + 0.0030147614888846874, + 0.005180086474865675, + 0.0036465812008827925, + 0.0019566358532756567, + 0.0018928439822047949, + 0.006206389982253313, + 0.002051894785836339, + 0.001358425710350275, + 0.0018400326371192932, + 0.002745646983385086, + 0.006301512476056814, + 0.02046920359134674, + 0.008521394804120064, + 0.13632121682167053, + 0.015717465430498123, + 0.007955185137689114, + 0.012969802133738995, + 0.011725451797246933, + 0.1218176931142807, + 0.0037463733460754156, + 0.008100088685750961, + 0.005718369502574205, + 0.006414020899683237, + 0.005205025430768728, + 0.003957123029977083, + 0.007370407693088055, + 0.006785799749195576, + 0.0035044143442064524, + 0.00372284185141325, + 0.004476194269955158, + 0.005111952777951956, + 0.00850663986057043, + 0.0030954487156122923, + 0.0036021315027028322, + 0.003645683638751507, + 0.008103839121758938, + 0.010873592458665371, + 0.014056216925382614, + 0.01654685102403164, + 0.010930445976555347, + 0.01413416862487793, + 0.018229378387331963, + 0.026222361251711845, + 0.004563739988952875, + 0.0033213712740689516, + 0.005056498106569052, + 0.016095967963337898, + 0.004110876936465502, + 0.0035467667039483786, + 0.005880956072360277, + 0.00492510199546814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.11292329430580139 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635258063673973 + ], + "opposite_logprob_tokens": [ + -3.651352643966675 + ], + "logprob_reward": -0.02635258063673973, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635258063673973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04335968568921089, + 0.07419013977050781, + 0.05193784460425377, + 0.025859728455543518, + 0.041716814041137695, + 0.6126145124435425, + 0.05851165950298309, + 0.03338039666414261, + 0.04340200498700142, + 0.015027131885290146 + ], + "credit_map": [ + 0.004274530801922083, + 0.002366436179727316, + 0.011707112193107605, + 0.0038995305076241493, + 0.0027802460826933384, + 0.0047723958268761635, + 0.006152807734906673, + 0.007406627293676138, + 0.0056084319949150085, + 0.002740140538662672, + 0.02397487871348858, + 0.004126837011426687, + 0.002450251020491123, + 0.005268370732665062, + 0.006058721803128719, + 0.023962507024407387, + 0.0061778901144862175, + 0.0037770478520542383, + 0.011518857441842556, + 0.006009934935718775, + 0.0035042676609009504, + 0.00510943541303277, + 0.0041450862772762775, + 0.01169532723724842, + 0.003812866285443306, + 0.002528577344492078, + 0.0035279076546430588, + 0.002968935528770089, + 0.002496476750820875, + 0.004089719615876675, + 0.003047182923182845, + 0.0033880623523145914, + 0.009014847688376904, + 0.0033906653989106417, + 0.007013136055320501, + 0.003444130066782236, + 0.002328445902094245, + 0.005679895170032978, + 0.004948900081217289, + 0.005896793678402901, + 0.03700685128569603, + 0.020185628905892372, + 0.24143119156360626, + 0.031051810830831528, + 0.019579395651817322, + 0.027890406548976898, + 0.02273104153573513, + 0.2127382457256317, + 0.004765421152114868, + 0.008283599279820919, + 0.0071220193058252335, + 0.008375946432352066, + 0.006617826409637928, + 0.005348378326743841, + 0.011403785087168217, + 0.006594683974981308, + 0.0041133929044008255, + 0.0033543286845088005, + 0.003582627745345235, + 0.004896725527942181, + 0.005911360494792461, + 0.0026607278268784285, + 0.005038427654653788, + 0.003822806291282177, + 0.00574807683005929, + 0.00590251199901104, + 0.00585600221529603, + 0.006032873876392841, + 0.0041349600069224834, + 0.004225792363286018, + 0.004922810941934586, + 0.006578977685421705, + 0.001472467090934515, + 0.0013331817463040352, + 0.0015344249550253153, + 0.002860309788957238, + 0.0013932521687820554, + 0.0013299299171194434, + 0.0023840509820729494, + 0.0027195154689252377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.02635258063673973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001759111881256 + ], + "opposite_logprob_tokens": [ + -2.6950175762176514 + ], + "logprob_reward": -0.07001759111881256, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001759111881256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04619082435965538, + 0.22211571037769318, + 0.031911950558423996, + 0.017905736342072487, + 0.061833709478378296, + 0.17399604618549347, + 0.06323651224374771, + 0.17862793803215027, + 0.11625557392835617, + 0.08792604506015778 + ], + "credit_map": [ + 0.004984038416296244, + 0.003358236514031887, + 0.01050232071429491, + 0.005095377564430237, + 0.002948608249425888, + 0.00599377416074276, + 0.005372111219912767, + 0.007936352863907814, + 0.016293048858642578, + 0.003563589183613658, + 0.0825689509510994, + 0.007736943196505308, + 0.005032716318964958, + 0.015983249992132187, + 0.006422866601496935, + 0.0845143273472786, + 0.005325030069798231, + 0.0027242035139352083, + 0.006302324589341879, + 0.003303497564047575, + 0.0025001822505146265, + 0.002418579999357462, + 0.0034164213575422764, + 0.005921707022935152, + 0.002410594606772065, + 0.0015276916092261672, + 0.0024743354879319668, + 0.002341890474781394, + 0.001626881305128336, + 0.0022113583981990814, + 0.0030252260621637106, + 0.0022877552546560764, + 0.007752810139209032, + 0.0042136237025260925, + 0.013663935475051403, + 0.004918606951832771, + 0.0029392309952527285, + 0.00696061085909605, + 0.006467417813837528, + 0.014917464926838875, + 0.010921581648290157, + 0.006581458263099194, + 0.06627753376960754, + 0.00891378615051508, + 0.006121547427028418, + 0.00786280632019043, + 0.00848030112683773, + 0.05883701145648956, + 0.004014537204056978, + 0.006636857986450195, + 0.014468178153038025, + 0.006233698222786188, + 0.005231039132922888, + 0.004218657035380602, + 0.005416390486061573, + 0.017017146572470665, + 0.023845776915550232, + 0.0184902623295784, + 0.028650792315602303, + 0.01618340238928795, + 0.03382713347673416, + 0.016531014814972878, + 0.021605798974633217, + 0.01949373446404934, + 0.0072250100784003735, + 0.011293788440525532, + 0.016386229544878006, + 0.017093772068619728, + 0.012631157413125038, + 0.010181575082242489, + 0.012484876438975334, + 0.02895914576947689, + 0.007407551631331444, + 0.008443647995591164, + 0.010812872089445591, + 0.021821120753884315, + 0.006865690462291241, + 0.006594722159206867, + 0.01435831654816866, + 0.01162211038172245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.07001759111881256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures the serene face of a surreal wax figure of a woman, her features reminiscent of classical sculpture. Her dark, wavy hair is meticulously detailed, each strand catching the light. Under a dramatic overhead spotlight, her form is sculpted from a profound black void, her skin possessing a soft, waxy sheen, punctuated by subtle, painterly imperfections that enhance its realism. Then, the camera begins a slow, tight push-in, its movements synchronizing with the figure's decay. The process begins almost imperceptibly, with a tiny crack appearing across her high cheekbone, rapidly spreading across the smooth surface. Fine, sand-like particles begin to drip from the crack, catching the harsh light as they fall. As the camera inexorably moves in, the disintegration accelerates; the bridge of her perfect nose dents and erodes, and a large section of her cheek peels away, dissolving in mid-air into a cloud of fine black dust, revealing a hollow interior. The entire structure of her face begins to fail, the once-peaceful lips drooping and crumbling as the camera inexorably advances, focusing on the catastrophic, silent collapse until the frame is filled with the chaotic movement of disintegrating wax and swirling dust, creating a deeply unsettling atmosphere of tragedy. Beauty transformed into devastation.", + "seed": 23898, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_0/index_7718_seed_23898_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": -0.027143442211672662, + "vlm_logprob": -0.027143442211672662, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.027143442211672662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006365618668496609 + ], + "opposite_logprob_tokens": [ + -6.381365776062012 + ], + "logprob_reward": -0.006365618668496609, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.006365618668496609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07749827206134796, + 0.03778958320617676, + 0.032529354095458984, + 0.04500921070575714, + 0.03720853850245476, + 0.043792564421892166, + 0.04659834876656532, + 0.5385926961898804, + 0.047758884727954865, + 0.09322251379489899 + ], + "credit_map": [ + 0.007871671579778194, + 0.006482558790594339, + 0.015305664390325546, + 0.006444557104259729, + 0.006440045777708292, + 0.013694812543690205, + 0.007115475833415985, + 0.014143478125333786, + 0.003609344130381942, + 0.002815498039126396, + 0.0035806854721158743, + 0.0037776774261146784, + 0.0044092643074691296, + 0.00805648509413004, + 0.003920688293874264, + 0.007619940675795078, + 0.0031692315824329853, + 0.00211680238135159, + 0.0033090899232774973, + 0.002560465130954981, + 0.002951439004391432, + 0.00803932175040245, + 0.002930609043687582, + 0.007452396210283041, + 0.004644124768674374, + 0.002089043380692601, + 0.004659038502722979, + 0.0030169778037816286, + 0.0033343236427754164, + 0.012512058950960636, + 0.0030249380506575108, + 0.011728703044354916, + 0.002834076527506113, + 0.0016899405745789409, + 0.0025266946759074926, + 0.0017160362331196666, + 0.004357377998530865, + 0.01115860790014267, + 0.002555336570367217, + 0.010370469652116299, + 0.004925835411995649, + 0.004463533870875835, + 0.005806501489132643, + 0.005245209205895662, + 0.005252212751656771, + 0.006860563065856695, + 0.003950964193791151, + 0.007287748623639345, + 0.00389427924528718, + 0.003179800696671009, + 0.00936808530241251, + 0.004266906064003706, + 0.004430809058248997, + 0.007007131353020668, + 0.006764727178961039, + 0.007686605677008629, + 0.009126199409365654, + 0.014387085102498531, + 0.02277684397995472, + 0.008638408035039902, + 0.013125140219926834, + 0.2161974310874939, + 0.01554750557988882, + 0.23879407346248627, + 0.005700381007045507, + 0.0034499703906476498, + 0.009407691657543182, + 0.00625983439385891, + 0.003294154768809676, + 0.007709155324846506, + 0.006580600049346685, + 0.005357097368687391, + 0.009974569082260132, + 0.00944379810243845, + 0.014370973221957684, + 0.014736559242010117, + 0.007088951766490936, + 0.013156929053366184, + 0.012606792151927948, + 0.011843943037092686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.006365618668496609 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653285913169384 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.008653285913169384, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653285913169384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051512714475393295, + 0.04639079049229622, + 0.05003261938691139, + 0.0395522266626358, + 0.11499419808387756, + 0.07074100524187088, + 0.07982750982046127, + 0.37308448553085327, + 0.03817218542098999, + 0.13569217920303345 + ], + "credit_map": [ + 0.004779462236911058, + 0.003452698467299342, + 0.007072742562741041, + 0.00381365860812366, + 0.004505647346377373, + 0.009146667085587978, + 0.008125122636556625, + 0.010616715997457504, + 0.004514036234468222, + 0.003952042665332556, + 0.00772135192528367, + 0.0041325571946799755, + 0.0038684450555592775, + 0.007891287095844746, + 0.00624573091045022, + 0.00806533731520176, + 0.004593509249389172, + 0.0025323876179754734, + 0.004322541877627373, + 0.002973807044327259, + 0.0032986230216920376, + 0.014796428382396698, + 0.0038864901289343834, + 0.013628832995891571, + 0.003559757024049759, + 0.002485246630385518, + 0.003992399200797081, + 0.0026150161866098642, + 0.004752241540700197, + 0.009711114689707756, + 0.0030114029068499804, + 0.009425048716366291, + 0.008066898211836815, + 0.004260471556335688, + 0.008355705998837948, + 0.0064345914870500565, + 0.014843634329736233, + 0.033724166452884674, + 0.00879424437880516, + 0.030514489859342575, + 0.008619433268904686, + 0.005983290262520313, + 0.009604210965335369, + 0.004965760745108128, + 0.010714659467339516, + 0.012991733849048615, + 0.004816871602088213, + 0.013045044615864754, + 0.008181368000805378, + 0.007311224937438965, + 0.011696775443851948, + 0.00887809693813324, + 0.009211231954395771, + 0.011297082528471947, + 0.010962465777993202, + 0.01228926982730627, + 0.006505053490400314, + 0.0076485914178192616, + 0.01570635475218296, + 0.005923847667872906, + 0.008806969970464706, + 0.15183338522911072, + 0.00934339314699173, + 0.16731689870357513, + 0.004748781677335501, + 0.0037676736246794462, + 0.00365788908675313, + 0.0060791634023189545, + 0.0036746724508702755, + 0.005472466815263033, + 0.006125907879322767, + 0.004645632114261389, + 0.007712104823440313, + 0.016046371310949326, + 0.02215215004980564, + 0.02524866722524166, + 0.011255228891968727, + 0.015701670199632645, + 0.02019772119820118, + 0.017378270626068115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.008653285913169384 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030493140220642 + ], + "opposite_logprob_tokens": [ + -2.3503048419952393 + ], + "logprob_reward": -0.10030493140220642, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030493140220642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10362212359905243, + 0.3071167767047882, + 0.06528138369321823, + 0.04693600907921791, + 0.02561967819929123, + 0.040573667734861374, + 0.08198187500238419, + 0.19421550631523132, + 0.06633207201957703, + 0.06832093745470047 + ], + "credit_map": [ + 0.010091080330312252, + 0.008654769510030746, + 0.01822061464190483, + 0.007094296161085367, + 0.00704951211810112, + 0.01963149569928646, + 0.011944970116019249, + 0.02093537151813507, + 0.03243095055222511, + 0.01990109495818615, + 0.05061013996601105, + 0.017430083826184273, + 0.02510054036974907, + 0.07280126959085464, + 0.02304232493042946, + 0.06580034643411636, + 0.0038323390763252974, + 0.0027823534328490496, + 0.008080636151134968, + 0.003446298884227872, + 0.0036574576515704393, + 0.019576922059059143, + 0.005609066225588322, + 0.01829630695283413, + 0.004637621808797121, + 0.003727484727278352, + 0.010907302610576153, + 0.0036047250032424927, + 0.003983730915933847, + 0.007211777381598949, + 0.006856204941868782, + 0.0060071577318012714, + 0.0023313267156481743, + 0.0019442548509687185, + 0.002685485640540719, + 0.0019483404466882348, + 0.0033229365944862366, + 0.005470806732773781, + 0.002810212317854166, + 0.005106313619762659, + 0.00472829956561327, + 0.005265843588858843, + 0.004008615389466286, + 0.003884622361510992, + 0.006055078469216824, + 0.006001561414450407, + 0.004587053786963224, + 0.0060425857082009315, + 0.014055988751351833, + 0.0074844020418822765, + 0.012805525213479996, + 0.0064550526440143585, + 0.007343084551393986, + 0.011955044232308865, + 0.008464924059808254, + 0.013417844660580158, + 0.004421864170581102, + 0.005668947473168373, + 0.009971706196665764, + 0.004191161133348942, + 0.005052386317402124, + 0.07558221369981766, + 0.005570161622017622, + 0.08375705778598785, + 0.008607529103755951, + 0.006293495185673237, + 0.009405326098203659, + 0.009424249641597271, + 0.006075827404856682, + 0.008967842906713486, + 0.009226211346685886, + 0.008331590332090855, + 0.006035403348505497, + 0.007872133515775204, + 0.01334421057254076, + 0.010469285771250725, + 0.0057778372429311275, + 0.007872136309742928, + 0.010184003971517086, + 0.006765923462808132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.10030493140220642 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819257251918316 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819257251918316, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819257251918316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07348906248807907, + 0.04905384033918381, + 0.047097258269786835, + 0.048144564032554626, + 0.03374020755290985, + 0.03084186278283596, + 0.29690462350845337, + 0.2567276954650879, + 0.035453129559755325, + 0.12854774296283722 + ], + "credit_map": [ + 0.007928975857794285, + 0.005949496291577816, + 0.014463458210229874, + 0.006326501723378897, + 0.006180084776133299, + 0.011118379421532154, + 0.011517085134983063, + 0.010005082003772259, + 0.003559308825060725, + 0.00305517902597785, + 0.006745221558958292, + 0.0036052456125617027, + 0.00457545043900609, + 0.011698883958160877, + 0.005455066449940205, + 0.010359489358961582, + 0.004090873058885336, + 0.0031131941359490156, + 0.00699643837288022, + 0.0024809855967760086, + 0.0045433081686496735, + 0.0104723135009408, + 0.005203418433666229, + 0.010196728631854057, + 0.004438838921487331, + 0.002497937995940447, + 0.005267698783427477, + 0.002617578487843275, + 0.003187348833307624, + 0.013646574690937996, + 0.003521544858813286, + 0.012967041693627834, + 0.003682538168504834, + 0.003452276811003685, + 0.005355357192456722, + 0.0035902250092476606, + 0.0038521348033100367, + 0.005011553410440683, + 0.003910282161086798, + 0.0048858411610126495, + 0.003425953909754753, + 0.002909622620791197, + 0.0045049008913338184, + 0.0033630849793553352, + 0.004294009879231453, + 0.0044354647397994995, + 0.002804224379360676, + 0.005104602314531803, + 0.04431144893169403, + 0.01794411614537239, + 0.040332891047000885, + 0.018770165741443634, + 0.01994095928966999, + 0.060500435531139374, + 0.017983753234148026, + 0.07712085545063019, + 0.006723281927406788, + 0.0116379139944911, + 0.011893635615706444, + 0.007796759717166424, + 0.008805972523987293, + 0.09465239197015762, + 0.009868402034044266, + 0.1053493395447731, + 0.002942119725048542, + 0.003732198616489768, + 0.006298260763287544, + 0.004045569337904453, + 0.0039680819027125835, + 0.005286247935146093, + 0.005211540963500738, + 0.003969110548496246, + 0.008169371634721756, + 0.017877493053674698, + 0.019614452496170998, + 0.02581367827951908, + 0.012625463306903839, + 0.01718134619295597, + 0.01781558431684971, + 0.009450346231460571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.01819257251918316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200802555307746 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.002200802555307746, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200802555307746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012716906145215034, + 0.008497696369886398, + 0.01175468135625124, + 0.01913684792816639, + 0.024238543584942818, + 0.01698208600282669, + 0.07639678567647934, + 0.7246672511100769, + 0.04974456503987312, + 0.055864688009023666 + ], + "credit_map": [ + 0.0011684279888868332, + 0.0011425298871472478, + 0.002466048812493682, + 0.0011717341840267181, + 0.0010860717156901956, + 0.002266013529151678, + 0.001466222107410431, + 0.0019498582696542144, + 0.0009643097873777151, + 0.0004908054252155125, + 0.0015038361307233572, + 0.0005181542946957052, + 0.000725537771359086, + 0.001828065956942737, + 0.0008453821064904332, + 0.0016216053627431393, + 0.0010799188166856766, + 0.0006070525851100683, + 0.0012605655938386917, + 0.0007257525576278567, + 0.0007929789135232568, + 0.00339238322339952, + 0.0008428900619037449, + 0.0030531389638781548, + 0.0014741045888513327, + 0.0010928938863798976, + 0.0018155929865315557, + 0.000906069646589458, + 0.001244524260982871, + 0.005970137659460306, + 0.0014353551669046283, + 0.005198169499635696, + 0.001799825462512672, + 0.001150790718384087, + 0.0021670772694051266, + 0.0011154264211654663, + 0.002104872604832053, + 0.007474851328879595, + 0.0013735669199377298, + 0.007052133791148663, + 0.0019608130678534508, + 0.0016615766799077392, + 0.00159395020455122, + 0.0013544755056500435, + 0.0020728816743940115, + 0.0034347218461334705, + 0.0013385231141000986, + 0.003565143560990691, + 0.011480498127639294, + 0.004666411783546209, + 0.0094825504347682, + 0.003528451779857278, + 0.006250577513128519, + 0.01592497155070305, + 0.003951901104301214, + 0.02111142687499523, + 0.010399894788861275, + 0.017694523558020592, + 0.031146839261054993, + 0.0095906350761652, + 0.01480447594076395, + 0.2965545654296875, + 0.01603047549724579, + 0.3284458518028259, + 0.004615153651684523, + 0.004710035864263773, + 0.0062875645235180855, + 0.008446668274700642, + 0.004669972695410252, + 0.007633730303496122, + 0.007940121926367283, + 0.005441316869109869, + 0.004289723001420498, + 0.004382641054689884, + 0.008921037428081036, + 0.009932749904692173, + 0.004131294786930084, + 0.008466691710054874, + 0.008709263056516647, + 0.007031290791928768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.002200802555307746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures the serene face of a surreal wax figure of a woman, her features reminiscent of classical sculpture. Her dark, wavy hair is meticulously detailed, each strand catching the light. Under a dramatic overhead spotlight, her form is sculpted from a profound black void, her skin possessing a soft, waxy sheen, punctuated by subtle, painterly imperfections that enhance its realism. Then, the camera begins a slow, tight push-in, its movements synchronizing with the figure's decay. The process begins almost imperceptibly, with a tiny crack appearing across her high cheekbone, rapidly spreading across the smooth surface. Fine, sand-like particles begin to drip from the crack, catching the harsh light as they fall. As the camera inexorably moves in, the disintegration accelerates; the bridge of her perfect nose dents and erodes, and a large section of her cheek peels away, dissolving in mid-air into a cloud of fine black dust, revealing a hollow interior. The entire structure of her face begins to fail, the once-peaceful lips drooping and crumbling as the camera inexorably advances, focusing on the catastrophic, silent collapse until the frame is filled with the chaotic movement of disintegrating wax and swirling dust, creating a deeply unsettling atmosphere of tragedy. Beauty transformed into devastation.", + "seed": 23899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_0/index_7718_seed_23899_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": -0.04870815481990576, + "vlm_logprob": -0.04870815481990576, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04870815481990576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014382976107299328 + ], + "opposite_logprob_tokens": [ + -4.889382839202881 + ], + "logprob_reward": -0.014382976107299328, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.014382976107299328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07871311157941818, + 0.5827484726905823, + 0.03739798069000244, + 0.08270882815122604, + 0.03052779659628868, + 0.0294575747102499, + 0.03481912985444069, + 0.03431496024131775, + 0.023900169879198074, + 0.06541194021701813 + ], + "credit_map": [ + 0.0047537581995129585, + 0.012811321765184402, + 0.008752682246267796, + 0.011791717261075974, + 0.010732397437095642, + 0.005228613968938589, + 0.011266371235251427, + 0.013376262038946152, + 0.03180498257279396, + 0.01448238454759121, + 0.2312585711479187, + 0.040054548531770706, + 0.016271531581878662, + 0.021622352302074432, + 0.02230895310640335, + 0.20494519174098969, + 0.0038153259083628654, + 0.005276438780128956, + 0.004416888579726219, + 0.005919174756854773, + 0.003320554504171014, + 0.0036605431232601404, + 0.004681984428316355, + 0.0063070752657949924, + 0.015921007841825485, + 0.010061008855700493, + 0.00876791775226593, + 0.011769932694733143, + 0.009176425635814667, + 0.007227436173707247, + 0.011728891171514988, + 0.008056215941905975, + 0.0047829304821789265, + 0.002766659948974848, + 0.003417471656575799, + 0.004231382627040148, + 0.003145235823467374, + 0.003978444263339043, + 0.004321662243455648, + 0.003884012345224619, + 0.0038277835119515657, + 0.0033207740634679794, + 0.0025986151304095984, + 0.005023532081395388, + 0.002633156254887581, + 0.004567112307995558, + 0.005138052627444267, + 0.0023485501296818256, + 0.004948717076331377, + 0.003968214616179466, + 0.003461566986516118, + 0.004631579853594303, + 0.005245172418653965, + 0.003441272536292672, + 0.0060698846355080605, + 0.003052724292501807, + 0.004923269618302584, + 0.005848526954650879, + 0.004128796514123678, + 0.0030254870653152466, + 0.0031681477557867765, + 0.00604650005698204, + 0.004219813738018274, + 0.002954424126073718, + 0.0028383792378008366, + 0.0037938428577035666, + 0.002813647734001279, + 0.003681262256577611, + 0.0020878196228295565, + 0.003212460782378912, + 0.003101050155237317, + 0.0023717095609754324, + 0.007395374123007059, + 0.013168209232389927, + 0.007167820353060961, + 0.0062367357313632965, + 0.009576725773513317, + 0.009975556284189224, + 0.006435326766222715, + 0.005456199403852224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.014382976107299328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082093119621277 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.016082093119621277, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082093119621277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10126452893018723, + 0.4771430492401123, + 0.11584636569023132, + 0.09823755919933319, + 0.03316623717546463, + 0.029452810063958168, + 0.028895022347569466, + 0.024281837046146393, + 0.031224234029650688, + 0.06048842519521713 + ], + "credit_map": [ + 0.005252036266028881, + 0.0045967260375618935, + 0.03308790177106857, + 0.008797931484878063, + 0.00557819614186883, + 0.004677914083003998, + 0.005730289965867996, + 0.03354352340102196, + 0.026092693209648132, + 0.011632255278527737, + 0.18893727660179138, + 0.03260714188218117, + 0.013681675307452679, + 0.017681756988167763, + 0.017931107431650162, + 0.1685791313648224, + 0.019489364698529243, + 0.00724192475900054, + 0.02863139845430851, + 0.00995267927646637, + 0.0049402364529669285, + 0.009529626928269863, + 0.007954779081046581, + 0.02810634858906269, + 0.018795955926179886, + 0.00806462299078703, + 0.02486667037010193, + 0.010659176856279373, + 0.005748759489506483, + 0.0086616026237607, + 0.008862435817718506, + 0.012578333728015423, + 0.005536975804716349, + 0.0031398918945342302, + 0.005646359175443649, + 0.004504075273871422, + 0.0027737608179450035, + 0.003915370441973209, + 0.0039003724232316017, + 0.0037494313437491655, + 0.003299375530332327, + 0.00234159454703331, + 0.00452430872246623, + 0.0042191254906356335, + 0.002802846021950245, + 0.004615786485373974, + 0.0036101252771914005, + 0.004039647988975048, + 0.00512693403288722, + 0.004092707298696041, + 0.0033579200971871614, + 0.0034496495500206947, + 0.0033312791492789984, + 0.0033988943323493004, + 0.003269718959927559, + 0.002867918461561203, + 0.0033675734885036945, + 0.0056171840988099575, + 0.002535905921831727, + 0.0026154248043894768, + 0.002535332227125764, + 0.002510752296075225, + 0.00256846914999187, + 0.0025311941280961037, + 0.004139076918363571, + 0.007654713001102209, + 0.0035816326271742582, + 0.0029674635734409094, + 0.0025181013625115156, + 0.004110426176339388, + 0.0030163051560521126, + 0.003236515913158655, + 0.006214024033397436, + 0.008040003478527069, + 0.005837437231093645, + 0.005197511985898018, + 0.008217819966375828, + 0.013684340752661228, + 0.007224727887660265, + 0.006072561256587505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.016082093119621277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292383074760437 + ], + "opposite_logprob_tokens": [ + -2.2379238605499268 + ], + "logprob_reward": -0.11292383074760437, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292383074760437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04736340790987015, + 0.5668084025382996, + 0.05966159328818321, + 0.06827452778816223, + 0.0466514527797699, + 0.11137134581804276, + 0.01398109458386898, + 0.015141895972192287, + 0.04962444305419922, + 0.021121859550476074 + ], + "credit_map": [ + 0.004259082954376936, + 0.005598881281912327, + 0.008802681230008602, + 0.007107329089194536, + 0.003835732350125909, + 0.0028631798923015594, + 0.004888542927801609, + 0.010007977485656738, + 0.030345994979143143, + 0.013912064023315907, + 0.2252548635005951, + 0.03885683789849281, + 0.015277503989636898, + 0.021333269774913788, + 0.02178478240966797, + 0.20004303753376007, + 0.008921812288463116, + 0.004318461287766695, + 0.014721634797751904, + 0.0059739891439676285, + 0.002802768722176552, + 0.004748466890305281, + 0.0047691864892840385, + 0.013405270874500275, + 0.015675723552703857, + 0.006213325075805187, + 0.015544841066002846, + 0.006501468829810619, + 0.004909245762974024, + 0.004548730794340372, + 0.005372237414121628, + 0.009508948773145676, + 0.005497907754033804, + 0.004786853678524494, + 0.009853067807853222, + 0.004863338079303503, + 0.004621750209480524, + 0.006442987825721502, + 0.00439437897875905, + 0.0061911651864647865, + 0.014124594628810883, + 0.007467770949006081, + 0.025606151670217514, + 0.014128460548818111, + 0.007078980561345816, + 0.019121699035167694, + 0.011373817920684814, + 0.012469860725104809, + 0.0025672095362097025, + 0.0010905639501288533, + 0.0022468899842351675, + 0.0018294317414984107, + 0.0013649041065946221, + 0.0015914702089503407, + 0.0015281842788681388, + 0.0017624407773837447, + 0.0018454121891409159, + 0.005269207525998354, + 0.0009391457424499094, + 0.0012730997987091541, + 0.0014859167858958244, + 0.001382547663524747, + 0.0013591054594144225, + 0.001587460283190012, + 0.006310662254691124, + 0.012165389955043793, + 0.005462262779474258, + 0.0053905644454061985, + 0.0036748028360307217, + 0.006551675498485565, + 0.005006367340683937, + 0.005062716547399759, + 0.0023991267662495375, + 0.0025224911514669657, + 0.0026662617456167936, + 0.0022109104320406914, + 0.00263560120947659, + 0.0045456476509571075, + 0.0022769952192902565, + 0.0018648238619789481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.11292383074760437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810739889740944 + ], + "opposite_logprob_tokens": [ + -3.288107395172119 + ], + "logprob_reward": -0.03810739889740944, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810739889740944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13763585686683655, + 0.07047383487224579, + 0.04784521460533142, + 0.11052748560905457, + 0.06043144315481186, + 0.06091831251978874, + 0.035697877407073975, + 0.06218096986413002, + 0.26861560344696045, + 0.14567331969738007 + ], + "credit_map": [ + 0.007930283434689045, + 0.013132541440427303, + 0.041512902826070786, + 0.0123239541426301, + 0.01327254343777895, + 0.008607187308371067, + 0.00975383073091507, + 0.031102605164051056, + 0.005111492704600096, + 0.0043840170837938786, + 0.0240019578486681, + 0.005326015409082174, + 0.0035308978985995054, + 0.0031960245687514544, + 0.003968730103224516, + 0.02095470391213894, + 0.007041232660412788, + 0.004781887400895357, + 0.010351551696658134, + 0.004863552283495665, + 0.002775156171992421, + 0.004858881235122681, + 0.00397747615352273, + 0.009195474907755852, + 0.018048936501145363, + 0.008075154386460781, + 0.02911812998354435, + 0.014347909018397331, + 0.005366980098187923, + 0.0076673272997140884, + 0.011176562868058681, + 0.01672649011015892, + 0.013575015589594841, + 0.005199627950787544, + 0.009943550452589989, + 0.006793079432100058, + 0.005196836777031422, + 0.006311031524091959, + 0.006956946570426226, + 0.006455357186496258, + 0.009897538460791111, + 0.004400845617055893, + 0.009918061085045338, + 0.007048585917800665, + 0.004793735686689615, + 0.012439083307981491, + 0.006691993214190006, + 0.005728465970605612, + 0.0069326674565672874, + 0.003216042183339596, + 0.004545482806861401, + 0.0057097142562270164, + 0.003551469650119543, + 0.003715363796800375, + 0.00428835628554225, + 0.0037387795746326447, + 0.0077879843302071095, + 0.016350867226719856, + 0.004831717815250158, + 0.00630515580996871, + 0.007186894305050373, + 0.007284585852175951, + 0.004963482730090618, + 0.007470286916941404, + 0.033934030681848526, + 0.060119666159152985, + 0.032220929861068726, + 0.029970820993185043, + 0.018059389665722847, + 0.037691693753004074, + 0.02882404625415802, + 0.027795011177659035, + 0.020469294860959053, + 0.012456176802515984, + 0.018237218260765076, + 0.012989701703190804, + 0.016374599188566208, + 0.03811144083738327, + 0.012202966958284378, + 0.014831913635134697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.03810739889740944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wax figure of a woman with features reminiscent of classical sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure's form sculpted from a profound black void with a soft, waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dramatic overhead spotlight illuminating the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as a static, eye-level close-up and then perform a slow, tight push-in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's face disintegrate, showing cracks, peeling wax, and dissolving into black dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062044475227594376 + ], + "opposite_logprob_tokens": [ + -2.812044382095337 + ], + "logprob_reward": -0.062044475227594376, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062044475227594376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08195550739765167, + 0.4767663776874542, + 0.013211612589657307, + 0.04214634373784065, + 0.03903776779770851, + 0.013342877849936485, + 0.12317315489053726, + 0.026534808799624443, + 0.035977136343717575, + 0.14785443246364594 + ], + "credit_map": [ + 0.004695132840424776, + 0.005659434944391251, + 0.026501644402742386, + 0.007632151711732149, + 0.0054785278625786304, + 0.003880169242620468, + 0.005284882616251707, + 0.02282356098294258, + 0.025695234537124634, + 0.012205771170556545, + 0.18917986750602722, + 0.032677311450242996, + 0.013055183924734592, + 0.018365103751420975, + 0.018425481393933296, + 0.16716241836547852, + 0.0029836129397153854, + 0.0013657881645485759, + 0.0022667478770017624, + 0.001431268872693181, + 0.001233791932463646, + 0.0010123737156391144, + 0.0013337070122361183, + 0.0015843207947909832, + 0.005566784646362066, + 0.004455321002751589, + 0.009312538430094719, + 0.006118590012192726, + 0.004153293091803789, + 0.0029187218751758337, + 0.0037752233911305666, + 0.00584586663171649, + 0.004662306047976017, + 0.0031835935078561306, + 0.007788658607751131, + 0.005473343189805746, + 0.0032361657358705997, + 0.004620575811713934, + 0.0055281007662415504, + 0.0045450180768966675, + 0.0016609951853752136, + 0.001183687592856586, + 0.0017366174142807722, + 0.001873560482636094, + 0.0011177350534126163, + 0.0022376293782144785, + 0.0022826669737696648, + 0.0012499858858063817, + 0.018135279417037964, + 0.01694224588572979, + 0.010880865156650543, + 0.013845792040228844, + 0.02149363048374653, + 0.010665004141628742, + 0.019423894584178925, + 0.011786432936787605, + 0.003424104070290923, + 0.007681668270379305, + 0.002364196814596653, + 0.0025361457373946905, + 0.00270779337733984, + 0.0034194951876997948, + 0.0019084492232650518, + 0.002492952859029174, + 0.004813512787222862, + 0.008759591728448868, + 0.0026079663075506687, + 0.004543345887213945, + 0.0030579224694520235, + 0.003914774861186743, + 0.004775077570229769, + 0.003504945896565914, + 0.01401503849774599, + 0.01615339145064354, + 0.014637093059718609, + 0.011817293241620064, + 0.016535529866814613, + 0.04128420352935791, + 0.017465360462665558, + 0.015946518629789352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.062044475227594376 + } + ], + "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/0000028/group_1/index_tensor([6203, 6203])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_1/index_tensor([6203, 6203])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3f2e4f7f9113d4e5ab572d82ceeb07eaa838d123 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_1/index_tensor([6203, 6203])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, low-angle tracking shot glides smoothly above the water, moving parallel to the majestic, silent swan as it navigates the calm, glassy expanse. The swan's plumage is a flawless, brilliant white, its texture meticulously detailed, from a soft downy undercoat to a smooth, water-repellent outer coat. Its long, graceful neck forms a graceful S-curve, its head held with regal tranquility. Its bill, a vibrant orange with a distinctive black bottom knob, provides a single, sharp point of color in the otherwise muted scene. The swan glides forward effortlessly and deliberately, propelled by the powerful, invisible thumps of its webbed feet, leaving behind a subtle, expanding V-shaped wake that gently disturbs the mirror-like surface. The entire scene is shrouded in a thick, low-hanging morning mist, which softens all edges and partially obscures the distant outline of reeds on the shoreline. Soft, diffuse sunlight filters through the mist, casting a pearly sheen on the water and capturing the outline of the swan's form, imbuing the entire frame with a cool, tranquil, and ethereal atmosphere.", + "seed": 17100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_1/index_6203_seed_17100_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": -0.07021887479349971, + "vlm_logprob": -0.07021887479349971, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.07021887479349971, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07239123433828354 + ], + "opposite_logprob_tokens": [ + -3.5723912715911865 + ], + "logprob_reward": -0.07239123433828354, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.07239123433828354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18333694338798523, + 0.0867193266749382, + 0.08049535006284714, + 0.052297692745923996, + 0.07659411430358887, + 0.12966686487197876, + 0.12407005578279495, + 0.11332615464925766, + 0.09556552022695541, + 0.05792799964547157 + ], + "credit_map": [ + 0.023020494729280472, + 0.026861218735575676, + 0.020362617447972298, + 0.021771376952528954, + 0.015906978398561478, + 0.030421694740653038, + 0.018685268238186836, + 0.026307277381420135, + 0.014597361907362938, + 0.017477545887231827, + 0.010976376011967659, + 0.008974450640380383, + 0.0065674446523189545, + 0.008542140945792198, + 0.009821193292737007, + 0.009762812405824661, + 0.010103259235620499, + 0.011937418952584267, + 0.011388960294425488, + 0.011113107204437256, + 0.007992126978933811, + 0.010615894570946693, + 0.008046426810324192, + 0.009298156015574932, + 0.006947691086679697, + 0.007203010842204094, + 0.006723986007273197, + 0.0074991220608353615, + 0.006082253530621529, + 0.005332729779183865, + 0.006005633622407913, + 0.006503264419734478, + 0.014328363351523876, + 0.012265615165233612, + 0.008893207646906376, + 0.007318147458136082, + 0.006731446832418442, + 0.01086501870304346, + 0.006242712959647179, + 0.009949594736099243, + 0.02824804000556469, + 0.029243040829896927, + 0.012301085516810417, + 0.01852276735007763, + 0.010425841435790062, + 0.007229828275740147, + 0.01000090129673481, + 0.013695352710783482, + 0.01839098520576954, + 0.013343431055545807, + 0.015820976346731186, + 0.017179246991872787, + 0.017285538837313652, + 0.015796156600117683, + 0.013019278645515442, + 0.01323444489389658, + 0.01116940751671791, + 0.03317084535956383, + 0.010331236757338047, + 0.01149861328303814, + 0.023676136508584023, + 0.007104634772986174, + 0.006707442924380302, + 0.009667825885117054, + 0.00940169021487236, + 0.016277680173516273, + 0.01560705341398716, + 0.00958334282040596, + 0.01131352223455906, + 0.013911575078964233, + 0.008317992091178894, + 0.011152666062116623, + 0.0053772530518472195, + 0.014522690325975418, + 0.008563145995140076, + 0.004313168115913868, + 0.006227577570825815, + 0.004179093521088362, + 0.004330176394432783, + 0.010414891876280308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.07239123433828354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20164059102535248 + ], + "opposite_logprob_tokens": [ + -1.7016406059265137 + ], + "logprob_reward": -0.20164059102535248, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20164059102535248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2574448883533478, + 0.10805099457502365, + 0.055191658437252045, + 0.036615148186683655, + 0.039299238473176956, + 0.04945452883839607, + 0.26143407821655273, + 0.09994592517614365, + 0.03543370962142944, + 0.0571298748254776 + ], + "credit_map": [ + 0.025742439553141594, + 0.04150513932108879, + 0.03176834434270859, + 0.0356210432946682, + 0.027505546808242798, + 0.02633487619459629, + 0.03709771856665611, + 0.03186976537108421, + 0.01910061202943325, + 0.017846407368779182, + 0.01342739351093769, + 0.01217315997928381, + 0.009304950013756752, + 0.009547950699925423, + 0.012430474162101746, + 0.014220036566257477, + 0.0067136953584849834, + 0.008043761365115643, + 0.007382458541542292, + 0.006900594104081392, + 0.005656285677105188, + 0.0070095364935696125, + 0.007647602818906307, + 0.005837719887495041, + 0.00642799399793148, + 0.0057520875707268715, + 0.004349328111857176, + 0.004219763446599245, + 0.0036139145959168673, + 0.004708509426563978, + 0.0037244828417897224, + 0.0038190691266208887, + 0.007400219328701496, + 0.006995137315243483, + 0.0039050390478223562, + 0.00453848484903574, + 0.004645633976906538, + 0.0037493319250643253, + 0.0032589887268841267, + 0.004806403536349535, + 0.008576583117246628, + 0.005926971323788166, + 0.005529543850570917, + 0.003936591558158398, + 0.006663968320935965, + 0.009784985333681107, + 0.004954436328262091, + 0.004081452265381813, + 0.04637998342514038, + 0.0281561017036438, + 0.027483133599162102, + 0.025387341156601906, + 0.043363057076931, + 0.03414539620280266, + 0.024330737069249153, + 0.03218832239508629, + 0.010631859302520752, + 0.010950046591460705, + 0.015839235857129097, + 0.009566845372319221, + 0.007668229751288891, + 0.015063170343637466, + 0.009136768989264965, + 0.021089771762490273, + 0.003168668830767274, + 0.005254443734884262, + 0.004065671004354954, + 0.0034926861990243196, + 0.004482149612158537, + 0.007179364562034607, + 0.0033987993374466896, + 0.004391928669065237, + 0.006501219235360622, + 0.00829463079571724, + 0.00671294005587697, + 0.007966266013681889, + 0.008390623144805431, + 0.00618390878662467, + 0.007151833735406399, + 0.005928449332714081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.20164059102535248 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052625699900090694 + ], + "opposite_logprob_tokens": [ + -5.25526237487793 + ], + "logprob_reward": -0.0052625699900090694, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052625699900090694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20732980966567993, + 0.10152279585599899, + 0.09727144986391068, + 0.14058031141757965, + 0.11995582282543182, + 0.04283376410603523, + 0.1360083520412445, + 0.07879102230072021, + 0.030715009197592735, + 0.044991690665483475 + ], + "credit_map": [ + 0.02171158790588379, + 0.028653401881456375, + 0.026214774698019028, + 0.029549764469265938, + 0.01990378089249134, + 0.028644675388932228, + 0.03096308931708336, + 0.021688735112547874, + 0.013289247639477253, + 0.013309030793607235, + 0.01000983640551567, + 0.013962323777377605, + 0.008954397402703762, + 0.019123544916510582, + 0.012617756612598896, + 0.010256659239530563, + 0.015929337590932846, + 0.011201001703739166, + 0.009257444180548191, + 0.008315886370837688, + 0.009692979045212269, + 0.026687312871217728, + 0.007855996489524841, + 0.008331487886607647, + 0.023804783821105957, + 0.01953507401049137, + 0.009990567341446877, + 0.012834783643484116, + 0.014214418828487396, + 0.0385621003806591, + 0.010168968699872494, + 0.011469622142612934, + 0.02304873801767826, + 0.020435398444533348, + 0.013677682727575302, + 0.014574456959962845, + 0.009535417892038822, + 0.014322752133011818, + 0.009783879853785038, + 0.014577500522136688, + 0.009111450053751469, + 0.006795954890549183, + 0.0043817320838570595, + 0.004391897469758987, + 0.005380372516810894, + 0.004564933478832245, + 0.0038246174808591604, + 0.004382807295769453, + 0.02299271710216999, + 0.01357399020344019, + 0.014463306404650211, + 0.015149509534239769, + 0.02085222490131855, + 0.018567385151982307, + 0.013752962462604046, + 0.016656262800097466, + 0.00754881277680397, + 0.020124955102801323, + 0.007062406279146671, + 0.006362216081470251, + 0.015489932149648666, + 0.007393347565084696, + 0.006202285643666983, + 0.008607069961726665, + 0.0033062302973121405, + 0.005050395615398884, + 0.003308125538751483, + 0.0030391605105251074, + 0.0028572739101946354, + 0.005588906351476908, + 0.00255189323797822, + 0.005013022571802139, + 0.004390670917928219, + 0.0067994254641234875, + 0.008844382129609585, + 0.005197823978960514, + 0.003923187032341957, + 0.005613226443529129, + 0.005213751457631588, + 0.005009220447391272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0052625699900090694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810510411858559 + ], + "opposite_logprob_tokens": [ + -3.288105010986328 + ], + "logprob_reward": -0.03810510411858559, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810510411858559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07700911909341812, + 0.05991247296333313, + 0.29714837670326233, + 0.03597303107380867, + 0.12509147822856903, + 0.03695231303572655, + 0.06591300666332245, + 0.18272651731967926, + 0.061382606625556946, + 0.05789105221629143 + ], + "credit_map": [ + 0.014959011226892471, + 0.011247466318309307, + 0.008251646533608437, + 0.008913158439099789, + 0.007307105697691441, + 0.008510742336511612, + 0.00889291986823082, + 0.008927070535719395, + 0.00918218307197094, + 0.009740537032485008, + 0.007236949633806944, + 0.007244773209095001, + 0.006146915256977081, + 0.004893688019365072, + 0.007560164667665958, + 0.007907265797257423, + 0.038265954703092575, + 0.047905031591653824, + 0.044065091758966446, + 0.034291207790374756, + 0.024648690596222878, + 0.02591140754520893, + 0.037613868713378906, + 0.04444713145494461, + 0.005519428290426731, + 0.005403832998126745, + 0.005406338255852461, + 0.0042355298064649105, + 0.003980522509664297, + 0.0035492596216499805, + 0.004013597499579191, + 0.0038645246531814337, + 0.022498248144984245, + 0.022233018651604652, + 0.018217984586954117, + 0.012249168939888477, + 0.011656216345727444, + 0.014079861342906952, + 0.011430861428380013, + 0.012726121582090855, + 0.007215690799057484, + 0.006819683592766523, + 0.004159942269325256, + 0.0027627006638795137, + 0.003198768710717559, + 0.00536394352093339, + 0.003939805552363396, + 0.003491778625175357, + 0.011278982274234295, + 0.007608856540173292, + 0.0076761068776249886, + 0.006313270423561335, + 0.011046268045902252, + 0.008793356828391552, + 0.005895253736525774, + 0.007300917990505695, + 0.016280632466077805, + 0.05375872552394867, + 0.015999481081962585, + 0.01575692556798458, + 0.0417361855506897, + 0.014358047395944595, + 0.009754317812621593, + 0.015082203783094883, + 0.00714873056858778, + 0.012103754095733166, + 0.013078146614134312, + 0.005161765031516552, + 0.006000117864459753, + 0.004446597769856453, + 0.004887964576482773, + 0.00855552963912487, + 0.007389531936496496, + 0.01921333745121956, + 0.012787708081305027, + 0.002678523538634181, + 0.006916404701769352, + 0.0026229284703731537, + 0.0030867313034832478, + 0.003195888828486204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.03810510411858559 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369487449526787 + ], + "opposite_logprob_tokens": [ + -3.4086949825286865 + ], + "logprob_reward": -0.03369487449526787, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369487449526787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1390245258808136, + 0.05557569861412048, + 0.10064106434583664, + 0.048476334661245346, + 0.14677812159061432, + 0.03959959000349045, + 0.23144865036010742, + 0.06539204716682434, + 0.051763877272605896, + 0.12130007147789001 + ], + "credit_map": [ + 0.019873496145009995, + 0.019662782549858093, + 0.018481533974409103, + 0.017207622528076172, + 0.015440291725099087, + 0.01571277715265751, + 0.01372601930052042, + 0.01892000064253807, + 0.005694597493857145, + 0.008130655623972416, + 0.008177662268280983, + 0.005068366415798664, + 0.004450297448784113, + 0.009920228272676468, + 0.0077517107129096985, + 0.006382179446518421, + 0.016118928790092468, + 0.021043967455625534, + 0.011287074536085129, + 0.013007877394557, + 0.008099350146949291, + 0.010551913641393185, + 0.009476764127612114, + 0.011055187322199345, + 0.0051231058314442635, + 0.00857259426265955, + 0.00560372369363904, + 0.007155154831707478, + 0.00565818278118968, + 0.005583048798143864, + 0.0057033575139939785, + 0.005077166482806206, + 0.02741614170372486, + 0.022893503308296204, + 0.017822597175836563, + 0.019206775352358818, + 0.012330999597907066, + 0.015215150080621243, + 0.014419778250157833, + 0.017473170533776283, + 0.008127610199153423, + 0.006114869844168425, + 0.003937578294426203, + 0.003996618092060089, + 0.0036620693281292915, + 0.004944836720824242, + 0.004522896837443113, + 0.004293110221624374, + 0.02271958813071251, + 0.027720484882593155, + 0.030705386772751808, + 0.04200173169374466, + 0.02805805578827858, + 0.02065022476017475, + 0.026811204850673676, + 0.032781973481178284, + 0.007747150957584381, + 0.013451875187456608, + 0.008609676733613014, + 0.005062671843916178, + 0.010133109986782074, + 0.005308035295456648, + 0.003853108501061797, + 0.011226415634155273, + 0.005490923300385475, + 0.008830946870148182, + 0.008018122054636478, + 0.005688643082976341, + 0.007866309955716133, + 0.005839373916387558, + 0.0043173376470804214, + 0.005712217651307583, + 0.012848631478846073, + 0.035141438245773315, + 0.022247666493058205, + 0.011183260940015316, + 0.01380179449915886, + 0.012110765092074871, + 0.006658175960183144, + 0.007308333646506071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.03369487449526787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle tracking shot glides smoothly above the water, moving parallel to the majestic, silent swan as it navigates the calm, glassy expanse. The swan's plumage is a flawless, brilliant white, its texture meticulously detailed, from a soft downy undercoat to a smooth, water-repellent outer coat. Its long, graceful neck forms a graceful S-curve, its head held with regal tranquility. Its bill, a vibrant orange with a distinctive black bottom knob, provides a single, sharp point of color in the otherwise muted scene. The swan glides forward effortlessly and deliberately, propelled by the powerful, invisible thumps of its webbed feet, leaving behind a subtle, expanding V-shaped wake that gently disturbs the mirror-like surface. The entire scene is shrouded in a thick, low-hanging morning mist, which softens all edges and partially obscures the distant outline of reeds on the shoreline. Soft, diffuse sunlight filters through the mist, casting a pearly sheen on the water and capturing the outline of the swan's form, imbuing the entire frame with a cool, tranquil, and ethereal atmosphere.", + "seed": 17101, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_1/index_6203_seed_17101_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": -0.15359209505841137, + "vlm_logprob": -0.15359209505841137, + "vlm_yes_no_margin": 2.575000011920929, + "vlm_token_credit": -0.15359209505841137, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1656772345304489 + ], + "opposite_logprob_tokens": [ + -2.290677309036255 + ], + "logprob_reward": -0.1656772345304489, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1656772345304489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07473187148571014, + 0.12781386077404022, + 0.05505601689219475, + 0.13611267507076263, + 0.022108957171440125, + 0.11858426034450531, + 0.18340033292770386, + 0.1181083470582962, + 0.05403320863842964, + 0.11005056649446487 + ], + "credit_map": [ + 0.009231570176780224, + 0.00764361722394824, + 0.00937768816947937, + 0.008667225949466228, + 0.007870099507272243, + 0.008002075366675854, + 0.01364270318299532, + 0.010296881198883057, + 0.012734697200357914, + 0.014333345927298069, + 0.0243480633944273, + 0.015049873851239681, + 0.013743204064667225, + 0.013816066086292267, + 0.016661953181028366, + 0.017126647755503654, + 0.006427428685128689, + 0.0062391497194767, + 0.004505673423409462, + 0.0060346187092363834, + 0.00846884399652481, + 0.01179225742816925, + 0.005616039037704468, + 0.005972003564238548, + 0.020281795412302017, + 0.017104191705584526, + 0.011869445443153381, + 0.009753762744367123, + 0.017624570056796074, + 0.022999437525868416, + 0.01668408140540123, + 0.01979537308216095, + 0.0033467691391706467, + 0.0021303908433765173, + 0.0020214985124766827, + 0.0016792315291240811, + 0.004645769950002432, + 0.0040680342353880405, + 0.001803101389668882, + 0.002414160408079624, + 0.01823657564818859, + 0.01113486010581255, + 0.011578625999391079, + 0.00909691397100687, + 0.01859343983232975, + 0.02683037705719471, + 0.009801814332604408, + 0.01331164501607418, + 0.027428630739450455, + 0.020411191508173943, + 0.021797847002744675, + 0.016383996233344078, + 0.020643794909119606, + 0.043026313185691833, + 0.014491211622953415, + 0.01921733096241951, + 0.013859557919204235, + 0.019199660047888756, + 0.01088381465524435, + 0.008650986477732658, + 0.014780314639210701, + 0.029283666983246803, + 0.009949959814548492, + 0.011500378139317036, + 0.005216381512582302, + 0.008324583992362022, + 0.006683355197310448, + 0.005115360487252474, + 0.0064551630057394505, + 0.008024772629141808, + 0.008528957143425941, + 0.005684631876647472, + 0.01604388654232025, + 0.016868507489562035, + 0.011350227519869804, + 0.011081886477768421, + 0.015516719780862331, + 0.01891387440264225, + 0.009109742008149624, + 0.011165713891386986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.1656772345304489 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909071981906891 + ], + "opposite_logprob_tokens": [ + -2.4640908241271973 + ], + "logprob_reward": -0.08909071981906891, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909071981906891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14865988492965698, + 0.09504418075084686, + 0.15286752581596375, + 0.05848398059606552, + 0.05372650921344757, + 0.11370090395212173, + 0.07302118092775345, + 0.08986169099807739, + 0.052802715450525284, + 0.1618313491344452 + ], + "credit_map": [ + 0.017889566719532013, + 0.023208683356642723, + 0.017616886645555496, + 0.016021819785237312, + 0.018571510910987854, + 0.018465556204319, + 0.018523074686527252, + 0.018362805247306824, + 0.009539403021335602, + 0.01526554673910141, + 0.008512797765433788, + 0.010249190032482147, + 0.014955786056816578, + 0.016514843329787254, + 0.010534889996051788, + 0.009471721015870571, + 0.018267886713147163, + 0.01697496697306633, + 0.01592078059911728, + 0.01295752264559269, + 0.021840523928403854, + 0.0317952036857605, + 0.018166735768318176, + 0.01694391667842865, + 0.006732726935297251, + 0.007068756502121687, + 0.004766169469803572, + 0.006407751236110926, + 0.009664972312748432, + 0.00948298443108797, + 0.007564787287265062, + 0.006795835681259632, + 0.007978512905538082, + 0.00610252795740962, + 0.004674904048442841, + 0.004414156544953585, + 0.007996797561645508, + 0.01002667099237442, + 0.006541528273373842, + 0.005991409998387098, + 0.016944311559200287, + 0.011468425393104553, + 0.010967399924993515, + 0.008758574724197388, + 0.015439615584909916, + 0.026568006724119186, + 0.008933251723647118, + 0.014621313661336899, + 0.011106496676802635, + 0.00841134786605835, + 0.007256274577230215, + 0.006448386237025261, + 0.010448957793414593, + 0.016159122809767723, + 0.0057572354562580585, + 0.007433359511196613, + 0.011157209053635597, + 0.012965735979378223, + 0.010095320641994476, + 0.006351029966026545, + 0.014734074473381042, + 0.013590504415333271, + 0.011533251032233238, + 0.009434567764401436, + 0.005422891117632389, + 0.005766868591308594, + 0.005478479899466038, + 0.004325878340750933, + 0.013659677468240261, + 0.005948085803538561, + 0.006742692552506924, + 0.005458141211420298, + 0.028745386749505997, + 0.025998316705226898, + 0.009346134029328823, + 0.01248889323323965, + 0.033050358295440674, + 0.02394728921353817, + 0.01552592869848013, + 0.012729041278362274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.08909071981906891 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005265416111797094 + ], + "opposite_logprob_tokens": [ + -5.255265235900879 + ], + "logprob_reward": -0.005265416111797094, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265416111797094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0982275903224945, + 0.08512666821479797, + 0.0755654126405716, + 0.09929439425468445, + 0.09299923479557037, + 0.08509472012519836, + 0.12332990765571594, + 0.08732818067073822, + 0.11356884241104126, + 0.13946500420570374 + ], + "credit_map": [ + 0.0078095486387610435, + 0.010474959388375282, + 0.014818909578025341, + 0.011558949947357178, + 0.014507566578686237, + 0.011774605140089989, + 0.014714294113218784, + 0.012568764388561249, + 0.00830940343439579, + 0.009345430880784988, + 0.010795021429657936, + 0.008533073589205742, + 0.017021888867020607, + 0.011000139638781548, + 0.010312101803719997, + 0.009809615090489388, + 0.008733939379453659, + 0.009886065497994423, + 0.0069807711988687515, + 0.007638006005436182, + 0.011907230131328106, + 0.012607529759407043, + 0.009587057866156101, + 0.008224822580814362, + 0.012904445640742779, + 0.013781226240098476, + 0.007207116577774286, + 0.01081688143312931, + 0.0154831288382411, + 0.011708823963999748, + 0.015546700917184353, + 0.011846084147691727, + 0.013772084377706051, + 0.01031180378049612, + 0.00868074782192707, + 0.008903857320547104, + 0.018490130081772804, + 0.017723271623253822, + 0.006445209961384535, + 0.008672140538692474, + 0.011711466126143932, + 0.009424812160432339, + 0.007603996433317661, + 0.006510658655315638, + 0.01601223461329937, + 0.016930248588323593, + 0.00809154286980629, + 0.008809773251414299, + 0.01901569403707981, + 0.013624337501823902, + 0.0127123286947608, + 0.009472525678575039, + 0.014614228159189224, + 0.030658377334475517, + 0.009813467040657997, + 0.013418961316347122, + 0.009296619333326817, + 0.013266305439174175, + 0.0078087556175887585, + 0.007198878098279238, + 0.010912314988672733, + 0.021970948204398155, + 0.008454748429358006, + 0.008419614285230637, + 0.01269720122218132, + 0.01417633704841137, + 0.016533715650439262, + 0.010751977562904358, + 0.022309402003884315, + 0.01230785995721817, + 0.013732293620705605, + 0.01106006558984518, + 0.024328581988811493, + 0.027028556913137436, + 0.011870525777339935, + 0.01168864220380783, + 0.019112180918455124, + 0.01987648196518421, + 0.010094716213643551, + 0.015465343371033669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.005265416111797094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42894691228866577 + ], + "opposite_logprob_tokens": [ + -1.0539469718933105 + ], + "logprob_reward": -0.42894691228866577, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42894691228866577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06071823090314865, + 0.08826244622468948, + 0.04353346675634384, + 0.04638735204935074, + 0.08225804567337036, + 0.05052034556865692, + 0.2712458074092865, + 0.22098548710346222, + 0.043336909264326096, + 0.09275198727846146 + ], + "credit_map": [ + 0.006509327329695225, + 0.005685438401997089, + 0.006077087018638849, + 0.004571403842419386, + 0.01107164565473795, + 0.00909392535686493, + 0.009401503019034863, + 0.008307898417115211, + 0.008778496645390987, + 0.007955416105687618, + 0.013206045143306255, + 0.010640010237693787, + 0.009123960509896278, + 0.011645522899925709, + 0.01480377558618784, + 0.012109211646020412, + 0.00376914138905704, + 0.003856814233586192, + 0.004636512137949467, + 0.004198075737804174, + 0.006075060460716486, + 0.007209390867501497, + 0.008069178089499474, + 0.005719289183616638, + 0.004832519683986902, + 0.0054228650406003, + 0.005911789368838072, + 0.004868307616561651, + 0.007468494586646557, + 0.0074583012610673904, + 0.005095525644719601, + 0.005329540930688381, + 0.012041565962135792, + 0.009583369828760624, + 0.007948600687086582, + 0.006875954568386078, + 0.011700003407895565, + 0.018060147762298584, + 0.007005859632045031, + 0.009042536839842796, + 0.0071478961035609245, + 0.005251942202448845, + 0.004887194372713566, + 0.005077681969851255, + 0.006107451394200325, + 0.011351716704666615, + 0.005318352486938238, + 0.005378102883696556, + 0.04054544121026993, + 0.029932957142591476, + 0.03272399678826332, + 0.02413790300488472, + 0.031164392828941345, + 0.06300542503595352, + 0.021379336714744568, + 0.028356332331895828, + 0.023966863751411438, + 0.0350382998585701, + 0.020747587084770203, + 0.017670273780822754, + 0.0285651795566082, + 0.05419694632291794, + 0.019102422520518303, + 0.021697890013456345, + 0.00394095666706562, + 0.005697295535355806, + 0.00915718637406826, + 0.004741604905575514, + 0.0044452836737036705, + 0.004378022626042366, + 0.005577374715358019, + 0.005399183835834265, + 0.015811242163181305, + 0.018985021859407425, + 0.006948620080947876, + 0.005873526446521282, + 0.009185676462948322, + 0.02066214010119438, + 0.005248695611953735, + 0.010037058033049107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.42894691228866577 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898019254207611 + ], + "opposite_logprob_tokens": [ + -2.5789802074432373 + ], + "logprob_reward": -0.07898019254207611, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898019254207611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1342846155166626, + 0.18323445320129395, + 0.0680181160569191, + 0.21391242742538452, + 0.08678664267063141, + 0.13420908153057098, + 0.046790242195129395, + 0.04434356465935707, + 0.03022679127752781, + 0.058194056153297424 + ], + "credit_map": [ + 0.012659726664423943, + 0.017096316441893578, + 0.014002280309796333, + 0.011958086863160133, + 0.021637484431266785, + 0.01854047179222107, + 0.022102966904640198, + 0.01628728397190571, + 0.01738935150206089, + 0.021240048110485077, + 0.02753678895533085, + 0.019752323627471924, + 0.023804649710655212, + 0.027192432433366776, + 0.023896774277091026, + 0.022422077134251595, + 0.008941537700593472, + 0.007141795940697193, + 0.006233615800738335, + 0.006025980692356825, + 0.01139109767973423, + 0.011921491473913193, + 0.007962804287672043, + 0.008399792946875095, + 0.02140468917787075, + 0.026229487732052803, + 0.023262079805135727, + 0.022445116192102432, + 0.027661357074975967, + 0.03013266623020172, + 0.03152840957045555, + 0.031248629093170166, + 0.014020792208611965, + 0.009151443839073181, + 0.006381685379892588, + 0.0062767029739916325, + 0.018270445987582207, + 0.016487810760736465, + 0.00764871621504426, + 0.00854905229061842, + 0.020758332684636116, + 0.015111781656742096, + 0.012490146793425083, + 0.009107237681746483, + 0.020936615765094757, + 0.02953944355249405, + 0.010699128732085228, + 0.015566394664347172, + 0.006532055791467428, + 0.005197066348046064, + 0.0060650622472167015, + 0.004440366290509701, + 0.005612093955278397, + 0.009868491441011429, + 0.004321230109781027, + 0.004753873683512211, + 0.004363947082310915, + 0.00582008995115757, + 0.005559071898460388, + 0.0034619916696101427, + 0.005527013447135687, + 0.008117686957120895, + 0.005926183424890041, + 0.005567583721131086, + 0.003461367217823863, + 0.004276139661669731, + 0.003913953434675932, + 0.0030504968017339706, + 0.0044775307178497314, + 0.0037582414224743843, + 0.004077404737472534, + 0.003211657050997019, + 0.010912079364061356, + 0.013290279544889927, + 0.004949112888425589, + 0.00400859210640192, + 0.006867795251309872, + 0.00800118688493967, + 0.004182241391390562, + 0.005982772912830114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.07898019254207611 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle tracking shot glides smoothly above the water, moving parallel to the majestic, silent swan as it navigates the calm, glassy expanse. The swan's plumage is a flawless, brilliant white, its texture meticulously detailed, from a soft downy undercoat to a smooth, water-repellent outer coat. Its long, graceful neck forms a graceful S-curve, its head held with regal tranquility. Its bill, a vibrant orange with a distinctive black bottom knob, provides a single, sharp point of color in the otherwise muted scene. The swan glides forward effortlessly and deliberately, propelled by the powerful, invisible thumps of its webbed feet, leaving behind a subtle, expanding V-shaped wake that gently disturbs the mirror-like surface. The entire scene is shrouded in a thick, low-hanging morning mist, which softens all edges and partially obscures the distant outline of reeds on the shoreline. Soft, diffuse sunlight filters through the mist, casting a pearly sheen on the water and capturing the outline of the swan's form, imbuing the entire frame with a cool, tranquil, and ethereal atmosphere.", + "seed": 17102, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_1/index_6203_seed_17102_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": -0.038796144723892215, + "vlm_logprob": -0.038796144723892215, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.038796144723892215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05676332116127014 + ], + "opposite_logprob_tokens": [ + -3.8067634105682373 + ], + "logprob_reward": -0.05676332116127014, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.05676332116127014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21846981346607208, + 0.13292482495307922, + 0.056277282536029816, + 0.08952496200799942, + 0.06742196530103683, + 0.0931614339351654, + 0.20058126747608185, + 0.026683755218982697, + 0.08947719633579254, + 0.025477580726146698 + ], + "credit_map": [ + 0.017744142562150955, + 0.02039766125380993, + 0.029498964548110962, + 0.015696803107857704, + 0.06302940845489502, + 0.04300589859485626, + 0.015756292268633842, + 0.013340601697564125, + 0.010156303644180298, + 0.01738455891609192, + 0.016912205144762993, + 0.014102566055953503, + 0.019292959943413734, + 0.026066496968269348, + 0.014184107072651386, + 0.014825611375272274, + 0.005222362000495195, + 0.009924426674842834, + 0.007023293524980545, + 0.005765355192124844, + 0.006934192031621933, + 0.010153030976653099, + 0.004773661028593779, + 0.00648095365613699, + 0.00811068620532751, + 0.008558844216167927, + 0.01083326619118452, + 0.010475297458469868, + 0.016953548416495323, + 0.014195801690220833, + 0.01138349249958992, + 0.009014011360704899, + 0.006064287852495909, + 0.0065844119526445866, + 0.00911889597773552, + 0.004550780635327101, + 0.01250795554369688, + 0.015930024906992912, + 0.007237383630126715, + 0.00542821828275919, + 0.015398738905787468, + 0.010464615188539028, + 0.010452549904584885, + 0.005674123764038086, + 0.0166696198284626, + 0.019029127433896065, + 0.007447732612490654, + 0.008024916984140873, + 0.021406307816505432, + 0.016582215204834938, + 0.027599269524216652, + 0.013157831504940987, + 0.03925967961549759, + 0.05982602760195732, + 0.008355884812772274, + 0.014394037425518036, + 0.0022972740698605776, + 0.004870718810707331, + 0.003987496253103018, + 0.0030737228225916624, + 0.0036850159522145987, + 0.003697731764987111, + 0.002180539770051837, + 0.002891251351684332, + 0.008383659645915031, + 0.008339505642652512, + 0.012198188342154026, + 0.005129667930305004, + 0.02768840081989765, + 0.01571381650865078, + 0.005794718395918608, + 0.006229227874428034, + 0.002799572190269828, + 0.0049898517318069935, + 0.0027254936285316944, + 0.0027990643866360188, + 0.004039302933961153, + 0.002962245838716626, + 0.002245423151180148, + 0.0029166247695684433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.05676332116127014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868066683411598 + ], + "opposite_logprob_tokens": [ + -3.0486807823181152 + ], + "logprob_reward": -0.04868066683411598, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868066683411598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21195629239082336, + 0.04108019918203354, + 0.05080480873584747, + 0.04448302835226059, + 0.05069749429821968, + 0.10403917729854584, + 0.14904889464378357, + 0.2417377382516861, + 0.09382527321577072, + 0.012327092699706554 + ], + "credit_map": [ + 0.01442650891840458, + 0.020404431968927383, + 0.026073699817061424, + 0.01273957546800375, + 0.06702562421560287, + 0.04123053699731827, + 0.014361451379954815, + 0.015694469213485718, + 0.0031249569728970528, + 0.005671238526701927, + 0.0064109074883162975, + 0.0043616024777293205, + 0.007570510730147362, + 0.006532685365527868, + 0.0031721137929707766, + 0.004236182197928429, + 0.004320432431995869, + 0.0072044613771140575, + 0.007244105450809002, + 0.0046114008873701096, + 0.007880625315010548, + 0.010544183664023876, + 0.004116241354495287, + 0.00488335732370615, + 0.003981439862400293, + 0.004889186006039381, + 0.005304071586579084, + 0.004567023366689682, + 0.008123941719532013, + 0.006876879371702671, + 0.005969899240881205, + 0.004770591389387846, + 0.005346722900867462, + 0.003948627505451441, + 0.006102921906858683, + 0.004337289836257696, + 0.010035671293735504, + 0.010649901814758778, + 0.005238894838839769, + 0.005037468858063221, + 0.009609629400074482, + 0.01423027366399765, + 0.012202820740640163, + 0.01041555218398571, + 0.014831573702394962, + 0.019906416535377502, + 0.013144772499799728, + 0.009698149748146534, + 0.020787758752703667, + 0.007739651948213577, + 0.02107837423682213, + 0.0055174571461975574, + 0.031264789402484894, + 0.045843612402677536, + 0.005630210507661104, + 0.011187047697603703, + 0.01632797159254551, + 0.02516154572367668, + 0.03994118422269821, + 0.013481980189681053, + 0.059832993894815445, + 0.06095127761363983, + 0.009438151493668556, + 0.016602659597992897, + 0.008898412808775902, + 0.00845196470618248, + 0.012548963539302349, + 0.005612569395452738, + 0.02748832292854786, + 0.017187753692269325, + 0.006454771850258112, + 0.0071825142949819565, + 0.0012221521465107799, + 0.0026296922005712986, + 0.0014842480886727571, + 0.0011320038465783, + 0.0019152275053784251, + 0.0014801135985180736, + 0.0012849619379267097, + 0.0011786934919655323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.04868066683411598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032007675617933273 + ], + "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": [ + 0.1420988142490387, + 0.06309306621551514, + 0.058271072804927826, + 0.047555409371852875, + 0.1352732926607132, + 0.04689674824476242, + 0.2908805012702942, + 0.05775546282529831, + 0.12501226365566254, + 0.03316329047083855 + ], + "credit_map": [ + 0.011136142536997795, + 0.013480119407176971, + 0.014769214205443859, + 0.013136078603565693, + 0.03781077265739441, + 0.02598266676068306, + 0.013391715474426746, + 0.012392107397317886, + 0.004440002609044313, + 0.007362577132880688, + 0.01046293880790472, + 0.007181316614151001, + 0.014731744304299355, + 0.007005179300904274, + 0.006544872652739286, + 0.005364436190575361, + 0.003965053707361221, + 0.008855002000927925, + 0.008818420581519604, + 0.00579123105853796, + 0.008075721561908722, + 0.010564131662249565, + 0.005601244978606701, + 0.006600269582122564, + 0.003979697823524475, + 0.004777887370437384, + 0.005742871202528477, + 0.00453153345733881, + 0.009425520896911621, + 0.00698937987908721, + 0.00649499474093318, + 0.005613530054688454, + 0.014822473749518394, + 0.01223574299365282, + 0.016252320259809494, + 0.011649438180029392, + 0.023884451016783714, + 0.02996273711323738, + 0.013948744162917137, + 0.012517395429313183, + 0.005612646229565144, + 0.004442191682755947, + 0.0073036253452301025, + 0.004040130414068699, + 0.0070589142851531506, + 0.009605836123228073, + 0.004960195627063513, + 0.003873212495818734, + 0.03627748787403107, + 0.0182060394436121, + 0.04004368185997009, + 0.015122695825994015, + 0.06078822538256645, + 0.08804301917552948, + 0.012670484371483326, + 0.019728874787688255, + 0.004474966786801815, + 0.010321910493075848, + 0.006052015814930201, + 0.005216189194470644, + 0.012606087140738964, + 0.011065147817134857, + 0.004090733826160431, + 0.003928415477275848, + 0.012357933446764946, + 0.011263549327850342, + 0.016466109082102776, + 0.006313472520560026, + 0.038410235196352005, + 0.022423479706048965, + 0.008380824699997902, + 0.009396674111485481, + 0.003779627615585923, + 0.007335438393056393, + 0.0034681521356105804, + 0.003997626714408398, + 0.004804330877959728, + 0.0040799654088914394, + 0.002692622132599354, + 0.003005528822541237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0032007675617933273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204929202795029 + ], + "opposite_logprob_tokens": [ + -2.812049388885498 + ], + "logprob_reward": -0.06204929202795029, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204929202795029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10653261095285416, + 0.027991415932774544, + 0.18121777474880219, + 0.036972593516111374, + 0.2326119989156723, + 0.04680570587515831, + 0.1063438206911087, + 0.17535753548145294, + 0.054547689855098724, + 0.03161880373954773 + ], + "credit_map": [ + 0.009670406579971313, + 0.007453901693224907, + 0.01247837208211422, + 0.005839761812239885, + 0.0358673557639122, + 0.02216222509741783, + 0.0059472136199474335, + 0.0071133775636553764, + 0.0018703012028709054, + 0.003987133502960205, + 0.003975482191890478, + 0.002265217248350382, + 0.005933857522904873, + 0.004470502957701683, + 0.0020901730749756098, + 0.003398749278858304, + 0.015209676697850227, + 0.014149117283523083, + 0.026662180200219154, + 0.0179809108376503, + 0.03052675724029541, + 0.04227990284562111, + 0.018360156565904617, + 0.01604907400906086, + 0.004149375483393669, + 0.003119298256933689, + 0.004625742789357901, + 0.0032867146655917168, + 0.0064882622100412846, + 0.009164828807115555, + 0.003470591502264142, + 0.0026677832938730717, + 0.024748630821704865, + 0.018376465886831284, + 0.02830871008336544, + 0.016708113253116608, + 0.04269995540380478, + 0.06017129868268967, + 0.020700393244624138, + 0.020898470655083656, + 0.005676605273038149, + 0.0034750704653561115, + 0.007238647434860468, + 0.002173509681597352, + 0.009524068795144558, + 0.012645993381738663, + 0.0027722499798983335, + 0.0032995662186294794, + 0.015466603450477123, + 0.006031291093677282, + 0.014689542353153229, + 0.004201059229671955, + 0.021655486896634102, + 0.032342392951250076, + 0.00425511272624135, + 0.007702348288148642, + 0.011748554185032845, + 0.009057749062776566, + 0.02519494853913784, + 0.006906008347868919, + 0.05910952389240265, + 0.04769066721200943, + 0.0055001829750835896, + 0.01014990545809269, + 0.005318283569067717, + 0.0053649586625397205, + 0.007795235607773066, + 0.0031017265282571316, + 0.01527465134859085, + 0.010419579222798347, + 0.0035059961955994368, + 0.0037672624457627535, + 0.0035345647484064102, + 0.0025902893394231796, + 0.003449996467679739, + 0.0021044407039880753, + 0.007404820993542671, + 0.0076281349174678326, + 0.0020362643990665674, + 0.0028702937997877598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.06204929202795029 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328667603433132 + ], + "opposite_logprob_tokens": [ + -3.7732865810394287 + ], + "logprob_reward": -0.02328667603433132, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328667603433132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12994511425495148, + 0.139150008559227, + 0.03479587659239769, + 0.07440245896577835, + 0.1055586040019989, + 0.09211169183254242, + 0.20427413284778595, + 0.12727676331996918, + 0.07902578264474869, + 0.013459551148116589 + ], + "credit_map": [ + 0.01019700150936842, + 0.01180865429341793, + 0.014323703944683075, + 0.010724245570600033, + 0.033089522272348404, + 0.027379564940929413, + 0.012331577949225903, + 0.010090852156281471, + 0.009209082461893559, + 0.011182229034602642, + 0.019484637305140495, + 0.011781621724367142, + 0.034513067454099655, + 0.028747642412781715, + 0.01148797944188118, + 0.012743745930492878, + 0.003875563619658351, + 0.0046159978955984116, + 0.00432915985584259, + 0.0032502240501344204, + 0.004758401773869991, + 0.005609156563878059, + 0.0033946712501347065, + 0.004962699953466654, + 0.0065843891352415085, + 0.009962426498532295, + 0.008780212141573429, + 0.0071932473219931126, + 0.01255488209426403, + 0.010957743972539902, + 0.009992632083594799, + 0.00837691966444254, + 0.01035992056131363, + 0.013002712279558182, + 0.012438038364052773, + 0.00953580066561699, + 0.018847988918423653, + 0.021844828501343727, + 0.00956110842525959, + 0.009968197904527187, + 0.01246313750743866, + 0.01260362844914198, + 0.010243543423712254, + 0.00637650303542614, + 0.015620606020092964, + 0.02021985687315464, + 0.0065854876302182674, + 0.007998923771083355, + 0.025945192202925682, + 0.02350255660712719, + 0.026162510737776756, + 0.014286491088569164, + 0.041218847036361694, + 0.04703977331519127, + 0.011610159650444984, + 0.014508608728647232, + 0.008372948504984379, + 0.00754283182322979, + 0.01669931225478649, + 0.005105568561702967, + 0.04652034491300583, + 0.03341377526521683, + 0.0037832048255950212, + 0.0058387694880366325, + 0.007349273655563593, + 0.006826504599303007, + 0.010463938117027283, + 0.005275471601635218, + 0.022740881890058517, + 0.014957007020711899, + 0.005694486666470766, + 0.005718223284929991, + 0.0012728916481137276, + 0.0014580786228179932, + 0.0015322606777772307, + 0.0013945510145276785, + 0.0028959326446056366, + 0.002485399367287755, + 0.0010017255553975701, + 0.0014187112683430314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.02328667603433132 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle tracking shot glides smoothly above the water, moving parallel to the majestic, silent swan as it navigates the calm, glassy expanse. The swan's plumage is a flawless, brilliant white, its texture meticulously detailed, from a soft downy undercoat to a smooth, water-repellent outer coat. Its long, graceful neck forms a graceful S-curve, its head held with regal tranquility. Its bill, a vibrant orange with a distinctive black bottom knob, provides a single, sharp point of color in the otherwise muted scene. The swan glides forward effortlessly and deliberately, propelled by the powerful, invisible thumps of its webbed feet, leaving behind a subtle, expanding V-shaped wake that gently disturbs the mirror-like surface. The entire scene is shrouded in a thick, low-hanging morning mist, which softens all edges and partially obscures the distant outline of reeds on the shoreline. Soft, diffuse sunlight filters through the mist, casting a pearly sheen on the water and capturing the outline of the swan's form, imbuing the entire frame with a cool, tranquil, and ethereal atmosphere.", + "seed": 17103, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_1/index_6203_seed_17103_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": -0.07357094511389732, + "vlm_logprob": -0.07357094511389732, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.07357094511389732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07792400568723679 + ], + "opposite_logprob_tokens": [ + -3.5779240131378174 + ], + "logprob_reward": -0.07792400568723679, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.07792400568723679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08049602061510086, + 0.03782496601343155, + 0.09624101966619492, + 0.025249946862459183, + 0.09849435091018677, + 0.49848443269729614, + 0.07603276520967484, + 0.03128620609641075, + 0.03348283842206001, + 0.022407477721571922 + ], + "credit_map": [ + 0.017322296276688576, + 0.014239679090678692, + 0.007231972645968199, + 0.005101652815937996, + 0.005707926582545042, + 0.011698398739099503, + 0.006721422076225281, + 0.012472674250602722, + 0.007284005172550678, + 0.0086364122107625, + 0.003283238736912608, + 0.003288163570687175, + 0.004367851186543703, + 0.003327158046886325, + 0.002852175384759903, + 0.004785962402820587, + 0.012619460001587868, + 0.03339255601167679, + 0.004029415547847748, + 0.006066358648240566, + 0.008625587448477745, + 0.006665166933089495, + 0.0048029725439846516, + 0.020039500668644905, + 0.002496517263352871, + 0.006024376954883337, + 0.0020190952345728874, + 0.0016074491431936622, + 0.0019377837888896465, + 0.0023951074108481407, + 0.001928622485138476, + 0.006840994581580162, + 0.023938210681080818, + 0.011577039957046509, + 0.005860010161995888, + 0.003233483759686351, + 0.005415049381554127, + 0.006058234255760908, + 0.013403035700321198, + 0.029009288176894188, + 0.10631962865591049, + 0.21383967995643616, + 0.020105531439185143, + 0.015148158185184002, + 0.04356280714273453, + 0.020664164796471596, + 0.014938505366444588, + 0.06390595436096191, + 0.008586819283664227, + 0.01873672567307949, + 0.007524843327701092, + 0.007010065950453281, + 0.0061763557605445385, + 0.006271700840443373, + 0.006743394769728184, + 0.014982861466705799, + 0.004821394570171833, + 0.005638399161398411, + 0.0031024443451315165, + 0.0022479172330349684, + 0.0023757251910865307, + 0.002601515268906951, + 0.002683281200006604, + 0.00781552866101265, + 0.0026052973698824644, + 0.009706102311611176, + 0.001980325672775507, + 0.0013706013560295105, + 0.0026515766512602568, + 0.0026397723704576492, + 0.0029104389250278473, + 0.009618720971047878, + 0.003067917190492153, + 0.003674709238111973, + 0.0016690569464117289, + 0.0014460673555731773, + 0.005020324140787125, + 0.004467588383704424, + 0.001130756107158959, + 0.0019310583593323827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.07792400568723679 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1271049529314041 + ], + "opposite_logprob_tokens": [ + -2.1271049976348877 + ], + "logprob_reward": -0.1271049529314041, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1271049529314041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09061820060014725, + 0.087632916867733, + 0.3153426945209503, + 0.11539071798324585, + 0.1267571747303009, + 0.049193769693374634, + 0.0668143481016159, + 0.05697568878531456, + 0.03820106014609337, + 0.05307343974709511 + ], + "credit_map": [ + 0.012044116854667664, + 0.011362588033080101, + 0.011024423874914646, + 0.008645243011415005, + 0.01102527417242527, + 0.014414099045097828, + 0.012372483499348164, + 0.009729969315230846, + 0.012804358266294003, + 0.013259966857731342, + 0.0060661290772259235, + 0.005398017354309559, + 0.012166604399681091, + 0.023756081238389015, + 0.006260233465582132, + 0.00792152900248766, + 0.0398859865963459, + 0.10458292812108994, + 0.013699140399694443, + 0.01837804540991783, + 0.03248898684978485, + 0.030454939231276512, + 0.016363410279154778, + 0.05948927626013756, + 0.018917711451649666, + 0.019408561289310455, + 0.00826592929661274, + 0.009932918474078178, + 0.014332933351397514, + 0.023810407146811485, + 0.008284082636237144, + 0.012438173405826092, + 0.028122952207922935, + 0.012293006293475628, + 0.008058534003794193, + 0.005628061480820179, + 0.007339227478951216, + 0.007514196913689375, + 0.01809750683605671, + 0.03970368579030037, + 0.007180044427514076, + 0.00682323332875967, + 0.0048199729062616825, + 0.004304175730794668, + 0.004037334583699703, + 0.004588406067341566, + 0.004853408318012953, + 0.012587194330990314, + 0.008162825368344784, + 0.011074203066527843, + 0.005317689385265112, + 0.007672399748116732, + 0.008021811954677105, + 0.010811789892613888, + 0.008440694771707058, + 0.007312934845685959, + 0.006850427016615868, + 0.008798332884907722, + 0.004521395079791546, + 0.0029461425729095936, + 0.004250128287822008, + 0.006027445197105408, + 0.008008173666894436, + 0.015573643147945404, + 0.006265179254114628, + 0.011127997189760208, + 0.0025979795027524233, + 0.0018911369843408465, + 0.0044240341521799564, + 0.0035800370387732983, + 0.003604036057367921, + 0.004710658453404903, + 0.005655672866851091, + 0.008262259885668755, + 0.004585795570164919, + 0.00626766262575984, + 0.006471304688602686, + 0.010121997445821762, + 0.004362577572464943, + 0.007346172351390123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.1271049529314041 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595959834754467 + ], + "opposite_logprob_tokens": [ + -5.130959510803223 + ], + "logprob_reward": -0.00595959834754467, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595959834754467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09808894246816635, + 0.043899282813072205, + 0.14458522200584412, + 0.11557653546333313, + 0.179413303732872, + 0.14016053080558777, + 0.04083624109625816, + 0.127717986702919, + 0.06389045715332031, + 0.045831501483917236 + ], + "credit_map": [ + 0.012185758911073208, + 0.010312249884009361, + 0.011100389994680882, + 0.0083000548183918, + 0.010807656683027744, + 0.023331299424171448, + 0.011089503765106201, + 0.01096202339977026, + 0.00718449242413044, + 0.007263206876814365, + 0.004553606268018484, + 0.0029308225493878126, + 0.00461534271016717, + 0.00804055668413639, + 0.003894916269928217, + 0.005416336003690958, + 0.016826899722218513, + 0.044397249817848206, + 0.006636976730078459, + 0.009435364976525307, + 0.015725186094641685, + 0.017620015889406204, + 0.008119474165141582, + 0.025824055075645447, + 0.01804335229098797, + 0.03178056329488754, + 0.00950684119015932, + 0.008915126323699951, + 0.015064683742821217, + 0.014391953125596046, + 0.007237619254738092, + 0.010636389255523682, + 0.03384573012590408, + 0.01759026199579239, + 0.010680030100047588, + 0.0069555905647575855, + 0.00809562113136053, + 0.011746997945010662, + 0.02764938771724701, + 0.06284969300031662, + 0.028376247733831406, + 0.0519571527838707, + 0.00857418030500412, + 0.00853799283504486, + 0.013518559746444225, + 0.010175562463700771, + 0.005918917246162891, + 0.013101917691528797, + 0.008977466262876987, + 0.0065602934919297695, + 0.004461994860321283, + 0.004232880659401417, + 0.003008828032761812, + 0.005262034945189953, + 0.0034439414739608765, + 0.00488880043849349, + 0.015228891745209694, + 0.0201838631182909, + 0.010416749864816666, + 0.0065006366930902, + 0.010585824027657509, + 0.016919782385230064, + 0.01784384995698929, + 0.030038384720683098, + 0.004599864594638348, + 0.018275856971740723, + 0.003555823815986514, + 0.0030698205810040236, + 0.005438703112304211, + 0.003884869161993265, + 0.006307712756097317, + 0.018757812678813934, + 0.006748211570084095, + 0.007477922830730677, + 0.003974136430770159, + 0.0034511308185756207, + 0.008637286722660065, + 0.008694830350577831, + 0.0031415654812008142, + 0.003706415416672826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.00595959834754467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706726789474487 + ], + "opposite_logprob_tokens": [ + -2.1270673274993896 + ], + "logprob_reward": -0.12706726789474487, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706726789474487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12335751950740814, + 0.10801473259925842, + 0.1650533527135849, + 0.1936931312084198, + 0.12043023109436035, + 0.08507517725229263, + 0.03580452874302864, + 0.07886376231908798, + 0.052330225706100464, + 0.037377383559942245 + ], + "credit_map": [ + 0.026440605521202087, + 0.023481067270040512, + 0.012273238971829414, + 0.0074220336973667145, + 0.00983038917183876, + 0.013181968592107296, + 0.012784428894519806, + 0.017943773418664932, + 0.02422519028186798, + 0.0226152166724205, + 0.011460269801318645, + 0.00812278687953949, + 0.01216946542263031, + 0.010272249579429626, + 0.006997121963649988, + 0.012152429670095444, + 0.022610995918512344, + 0.04997485876083374, + 0.011125682853162289, + 0.009895454160869122, + 0.019635571166872978, + 0.02059324085712433, + 0.00878825318068266, + 0.022429287433624268, + 0.029562029987573624, + 0.05638023093342781, + 0.0158191230148077, + 0.013995973393321037, + 0.02296140044927597, + 0.022806020453572273, + 0.013039901852607727, + 0.01912844181060791, + 0.026449011638760567, + 0.01118775736540556, + 0.007471990771591663, + 0.004245734307914972, + 0.005463300738483667, + 0.006685999222099781, + 0.018190857023000717, + 0.04073556512594223, + 0.015001615509390831, + 0.03392551839351654, + 0.0037446944043040276, + 0.0037955502048134804, + 0.007391489110887051, + 0.0039689913392066956, + 0.0030305380932986736, + 0.014216773211956024, + 0.003608659841120243, + 0.007033067289739847, + 0.0039926450699567795, + 0.0033415635116398335, + 0.003989528864622116, + 0.0038995856884866953, + 0.003577065421268344, + 0.00636241165921092, + 0.01098957285284996, + 0.016376689076423645, + 0.006407835055142641, + 0.004738795571029186, + 0.005403963848948479, + 0.004310093820095062, + 0.009994832798838615, + 0.02064196951687336, + 0.008793809451162815, + 0.018976576626300812, + 0.0028266070876270533, + 0.0025960176717489958, + 0.007063809782266617, + 0.0036634430289268494, + 0.004171364940702915, + 0.004238596186041832, + 0.007007725536823273, + 0.005832716356962919, + 0.003492562798783183, + 0.00229746219702065, + 0.007578550837934017, + 0.0059111942537128925, + 0.0022239815443754196, + 0.0030331830494105816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.12706726789474487 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 and tracking smoothly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a swan with flawless brilliant white plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swan have a long neck forming a graceful S-curve with a vibrant orange bill and black bottom knob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in thick low-hanging morning mist that obscures distant reeds on the shoreline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the swan leave a subtle expanding V-shaped wake on the mirror-like water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798900708556175 + ], + "opposite_logprob_tokens": [ + -3.529798984527588 + ], + "logprob_reward": -0.029798900708556175, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798900708556175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019530439749360085, + 0.024540791288018227, + 0.02522972971200943, + 0.05080664902925491, + 0.180751770734787, + 0.574273943901062, + 0.05490950122475624, + 0.021582169458270073, + 0.03854009136557579, + 0.009834897704422474 + ], + "credit_map": [ + 0.0026465714909136295, + 0.002686477266252041, + 0.0019063364015892148, + 0.00176201481372118, + 0.0018416064558550715, + 0.004433013964444399, + 0.0019587783608585596, + 0.00229563913308084, + 0.0029268295038491488, + 0.003392556682229042, + 0.004478709306567907, + 0.001766555942595005, + 0.0028743522707372904, + 0.004787005949765444, + 0.001744008855894208, + 0.00257076951675117, + 0.0024555933196097612, + 0.006357968784868717, + 0.0031951081473380327, + 0.0020249374210834503, + 0.002132209949195385, + 0.0026596374809741974, + 0.0016595171764492989, + 0.004744755569845438, + 0.007210548501461744, + 0.01795322820544243, + 0.002621519612148404, + 0.0026222518645226955, + 0.0038254756946116686, + 0.005413525737822056, + 0.0023719847667962313, + 0.008788114413619041, + 0.042228735983371735, + 0.019369442015886307, + 0.010858852416276932, + 0.006085833068937063, + 0.009055422618985176, + 0.010323431342840195, + 0.026214029639959335, + 0.056616008281707764, + 0.1195603683590889, + 0.2456323504447937, + 0.02261822670698166, + 0.01722867786884308, + 0.05045802891254425, + 0.024700867012143135, + 0.016130607575178146, + 0.07794475555419922, + 0.0065086581744253635, + 0.009090829640626907, + 0.006385136861354113, + 0.003223589388653636, + 0.0035109964665025473, + 0.0029230096843093634, + 0.008342377841472626, + 0.014924895949661732, + 0.003053907072171569, + 0.005030108615756035, + 0.002561053028330207, + 0.0016163690015673637, + 0.001957300351932645, + 0.0023972440976649523, + 0.0016453360440209508, + 0.0033208495005965233, + 0.00637782970443368, + 0.01181208435446024, + 0.002478984883055091, + 0.0015443128068000078, + 0.004493032582104206, + 0.002316725440323353, + 0.0038938685320317745, + 0.005623249802738428, + 0.0024468710180372, + 0.00155785889364779, + 0.0011315243318676949, + 0.0006978530436754227, + 0.0013439254835247993, + 0.0011359861819073558, + 0.0006181741482578218, + 0.0009027040214277804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.029798900708556175 + } + ], + "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/0000028/group_10/index_tensor([9784, 9784])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_10/index_tensor([9784, 9784])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e58703f57eab89d80dfe8f1e1b785af644683a1f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_10/index_tensor([9784, 9784])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins in medium shot, slightly tilted upward, then slowly and steadily pulls back. The camera initially focuses on an elderly farmer, straining to hold a massive, textured, deep orange pumpkin. He is a dark, weathered man, his brow furrowed with exertion beneath his graying hair, his teeth clenched with clenched teeth, and beads of sweat beading on his forehead. He wears a faded blue plaid flannel shirt, sleeves rolled up to the elbows, revealing his muscular, veiny forearms. He wears dark brown canvas trousers stained with dirt. Bent over, his body trembling slightly from the weight of the pumpkin, he cautiously advances across the harvested land with slow, heavy steps. As the camera steadily pulls back, the image gradually widens, placing the farmer's figure within a larger context: a vast expanse of bountiful fields, shimmering in the golden evening sun. Only neatly harvested wheat stubble remained in the fields, long shadows stretching across the ground. On the distant horizon, a sparse row of trees and the silhouette of a small farmhouse could be seen. A soft, warm light shone a gentle glow over the scene, highlighting the tiny dust particles dancing in the air and creating an atmosphere of simplicity, tranquility, and contentment, silently celebrating hard work and the joy of a bountiful harvest.", + "seed": 4857, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_10/index_9784_seed_4857_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": -0.09114392814226449, + "vlm_logprob": -0.09114392814226449, + "vlm_yes_no_margin": 3.624999988079071, + "vlm_token_credit": -0.09114392814226449, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31878000497817993 + ], + "opposite_logprob_tokens": [ + -1.3187799453735352 + ], + "logprob_reward": -0.31878000497817993, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31878000497817993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3109452724456787, + 0.2606836259365082, + 0.04992857575416565, + 0.09464713931083679, + 0.04848989471793175, + 0.013024533167481422, + 0.019909001886844635, + 0.020909108221530914, + 0.1058824360370636, + 0.0755804255604744 + ], + "credit_map": [ + 0.03996073454618454, + 0.03237110748887062, + 0.06487121433019638, + 0.02916709892451763, + 0.034997597336769104, + 0.047684308141469955, + 0.029187314212322235, + 0.0327058807015419, + 0.02312086522579193, + 0.012114403769373894, + 0.10341235250234604, + 0.019094809889793396, + 0.016524089500308037, + 0.021273206919431686, + 0.042192209511995316, + 0.02295166812837124, + 0.006412511691451073, + 0.0036951026413589716, + 0.014669504016637802, + 0.005556188523769379, + 0.007346709258854389, + 0.0035813297145068645, + 0.004388764034956694, + 0.004278464242815971, + 0.010908774100244045, + 0.006096123717725277, + 0.02865687385201454, + 0.01715857908129692, + 0.00552982185035944, + 0.005435374565422535, + 0.009064074605703354, + 0.011797518469393253, + 0.009967193938791752, + 0.0030852421186864376, + 0.014186433516442776, + 0.005540322978049517, + 0.003979184199124575, + 0.003919502720236778, + 0.0030177836306393147, + 0.00479423301294446, + 0.0021025578025728464, + 0.0012984129134565592, + 0.0017784035298973322, + 0.001774039352312684, + 0.0019570793956518173, + 0.0013155135093256831, + 0.00150857784319669, + 0.0012899487046524882, + 0.002291243290528655, + 0.0019611206371337175, + 0.0033870916813611984, + 0.002421771641820669, + 0.002797831781208515, + 0.0022579210344702005, + 0.0021893878001719713, + 0.0026026354171335697, + 0.0018169861286878586, + 0.0024538321886211634, + 0.0044920542277395725, + 0.002759041730314493, + 0.0026470464654266834, + 0.0021159499883651733, + 0.0025661883410066366, + 0.002058009384199977, + 0.013474751263856888, + 0.012988144531846046, + 0.019996142014861107, + 0.008564785122871399, + 0.012530441395938396, + 0.013576505705714226, + 0.010678253136575222, + 0.014073416590690613, + 0.007588955108076334, + 0.009622778743505478, + 0.019383205100893974, + 0.009360979311168194, + 0.008668809197843075, + 0.007005586754530668, + 0.005518873687833548, + 0.008431239984929562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.31878000497817993 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028564624954015017 + ], + "opposite_logprob_tokens": [ + -5.877856254577637 + ], + "logprob_reward": -0.0028564624954015017, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028564624954015017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3020520806312561, + 0.39196452498435974, + 0.06160389631986618, + 0.0717957392334938, + 0.02311054803431034, + 0.022106293588876724, + 0.017273608595132828, + 0.023166373372077942, + 0.050262439996004105, + 0.03666455298662186 + ], + "credit_map": [ + 0.026301594451069832, + 0.04496040195226669, + 0.049015261232852936, + 0.03148692846298218, + 0.04746932536363602, + 0.033679183572530746, + 0.02982170507311821, + 0.039317671209573746, + 0.02410816214978695, + 0.017325863242149353, + 0.14678455889225006, + 0.03350767120718956, + 0.020251967012882233, + 0.0726211741566658, + 0.05125361308455467, + 0.026111504063010216, + 0.007291701156646013, + 0.00432258564978838, + 0.018103493377566338, + 0.007156162057071924, + 0.009065642952919006, + 0.004611968994140625, + 0.005362341646105051, + 0.005690000951290131, + 0.011659511364996433, + 0.003647345583885908, + 0.02673366479575634, + 0.007214791607111692, + 0.005161055363714695, + 0.00473019341006875, + 0.0038677994161844254, + 0.008781377226114273, + 0.002513939281925559, + 0.002308677649125457, + 0.003146989271044731, + 0.003067977726459503, + 0.00455946521833539, + 0.0023802125360816717, + 0.002440832555294037, + 0.0026924533303827047, + 0.003340374445542693, + 0.0027970976661890745, + 0.0027122776955366135, + 0.003641069633886218, + 0.0027969791553914547, + 0.0017608779016882181, + 0.0029437404591590166, + 0.002113878261297941, + 0.0019450717372819781, + 0.0018972106045112014, + 0.0028458903543651104, + 0.0021407746244221926, + 0.002436051843687892, + 0.001625220524147153, + 0.0022567545529454947, + 0.0021266357507556677, + 0.0027377924416214228, + 0.0021655878517776728, + 0.004680379759520292, + 0.0030236185993999243, + 0.0021457518450915813, + 0.0027184344362467527, + 0.0028884410858154297, + 0.002806365955621004, + 0.006906383670866489, + 0.00552828423678875, + 0.009682090021669865, + 0.004289617296308279, + 0.005316512193530798, + 0.006322498898953199, + 0.005279728677123785, + 0.006937328260391951, + 0.0059386612847447395, + 0.00449042534455657, + 0.007124802563339472, + 0.003954730462282896, + 0.003016130765900016, + 0.004435764625668526, + 0.0030196441803127527, + 0.004684391897171736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0028564624954015017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864111352711916 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.00864111352711916, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864111352711916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30341991782188416, + 0.23695741593837738, + 0.17264266312122345, + 0.09241709113121033, + 0.019066043198108673, + 0.02203834056854248, + 0.020748408511281013, + 0.03825509175658226, + 0.05293996259570122, + 0.041515033692121506 + ], + "credit_map": [ + 0.03814948350191116, + 0.02835548110306263, + 0.0558626651763916, + 0.046741921454668045, + 0.02616972289979458, + 0.04552309215068817, + 0.03190712258219719, + 0.030710412189364433, + 0.01615997776389122, + 0.014513561502099037, + 0.08930090814828873, + 0.02005847729742527, + 0.01744355633854866, + 0.03449541702866554, + 0.024685174226760864, + 0.020300351083278656, + 0.02577802911400795, + 0.013784484006464481, + 0.04418659582734108, + 0.021070795133709908, + 0.025574609637260437, + 0.010992889292538166, + 0.01708024926483631, + 0.01417502574622631, + 0.009626764804124832, + 0.011296269483864307, + 0.02039107121527195, + 0.00930870696902275, + 0.015399318188428879, + 0.005723325069993734, + 0.00863496121019125, + 0.012036675587296486, + 0.0025312011130154133, + 0.0019133458845317364, + 0.004826304502785206, + 0.0018046221230179071, + 0.0032531945034861565, + 0.0012784613063558936, + 0.001633065170608461, + 0.001825849525630474, + 0.002755258698016405, + 0.0021344677079468966, + 0.004357465542852879, + 0.002464679069817066, + 0.003791908733546734, + 0.0017609942005947232, + 0.002763518365100026, + 0.002010048134252429, + 0.003452532459050417, + 0.0015696281334385276, + 0.006229161750525236, + 0.0023166907485574484, + 0.0018562017939984798, + 0.0014558507828041911, + 0.0020777364261448383, + 0.0017906056018546224, + 0.003511143149808049, + 0.003984435461461544, + 0.006785181351006031, + 0.005581436213105917, + 0.005548637360334396, + 0.004539300221949816, + 0.004848124925047159, + 0.0034568337723612785, + 0.006576747167855501, + 0.00581215787678957, + 0.009531987830996513, + 0.00473113963380456, + 0.006381249520927668, + 0.006817175541073084, + 0.005850398447364569, + 0.007239105645567179, + 0.00543801998719573, + 0.0047667683102190495, + 0.010463494807481766, + 0.0047764647752046585, + 0.004152119159698486, + 0.003799019381403923, + 0.0029306551441550255, + 0.005188490264117718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00864111352711916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535588815808296 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535588815808296, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535588815808296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3939746618270874, + 0.1223456934094429, + 0.06321923434734344, + 0.09034248441457748, + 0.07197793573141098, + 0.09603901207447052, + 0.019344566389918327, + 0.022189101204276085, + 0.009488223120570183, + 0.11107909679412842 + ], + "credit_map": [ + 0.03894801065325737, + 0.042896077036857605, + 0.08005324006080627, + 0.04199640452861786, + 0.04299696907401085, + 0.05782583728432655, + 0.03906607627868652, + 0.05019203945994377, + 0.008257118985056877, + 0.004071775358170271, + 0.05126268044114113, + 0.009128626435995102, + 0.0050635673105716705, + 0.01956750638782978, + 0.015203687362372875, + 0.009790726937353611, + 0.011334440670907497, + 0.0030663509387522936, + 0.017494739964604378, + 0.008994617499411106, + 0.006540031637996435, + 0.003977625630795956, + 0.0062364148907363415, + 0.005575012881308794, + 0.012797311879694462, + 0.006676602177321911, + 0.020938856527209282, + 0.01939389482140541, + 0.005565925966948271, + 0.005091474391520023, + 0.011046928353607655, + 0.008831481449306011, + 0.014243761077523232, + 0.004449297208338976, + 0.02459152415394783, + 0.005396328400820494, + 0.006059465929865837, + 0.0059899818152189255, + 0.004832861479371786, + 0.0064147156663239, + 0.01972532831132412, + 0.008925197646021843, + 0.013963484205305576, + 0.010723215527832508, + 0.009848949499428272, + 0.01019508857280016, + 0.007717830128967762, + 0.014939919114112854, + 0.00266195903532207, + 0.0018686124822124839, + 0.0037120296619832516, + 0.0027951891534030437, + 0.0022176303900778294, + 0.0017283209599554539, + 0.0022378272842615843, + 0.0021229987032711506, + 0.003940627444535494, + 0.0024433820508420467, + 0.0029043410904705524, + 0.003047939855605364, + 0.0018014031229540706, + 0.0028431820683181286, + 0.0030157826840877533, + 0.0021924441680312157, + 0.0009875857504084706, + 0.0013721463037654757, + 0.0011471034958958626, + 0.0015247648116201162, + 0.0011616781121119857, + 0.0007090282742865384, + 0.0013847934314981103, + 0.0012011234648525715, + 0.020145956426858902, + 0.009301649406552315, + 0.025575604289770126, + 0.008352789096534252, + 0.009245093911886215, + 0.016587845981121063, + 0.007029298227280378, + 0.014840853400528431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012535588815808296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290647089481354 + ], + "opposite_logprob_tokens": [ + -2.2379064559936523 + ], + "logprob_reward": -0.11290647089481354, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290647089481354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24352885782718658, + 0.23021802306175232, + 0.09956993907690048, + 0.06919943541288376, + 0.02627718448638916, + 0.02642802707850933, + 0.10557005554437637, + 0.026538345962762833, + 0.09514295309782028, + 0.07752720266580582 + ], + "credit_map": [ + 0.02368845045566559, + 0.023168062791228294, + 0.049897514283657074, + 0.03514625504612923, + 0.019641326740384102, + 0.04132802411913872, + 0.0249626487493515, + 0.02569657377898693, + 0.014922327362000942, + 0.007471405901014805, + 0.09281978756189346, + 0.01856556348502636, + 0.009606361389160156, + 0.04280155897140503, + 0.025775596499443054, + 0.01825542375445366, + 0.01658773608505726, + 0.00576378358528018, + 0.023257935419678688, + 0.015520391054451466, + 0.008892093785107136, + 0.0077703106217086315, + 0.01259857602417469, + 0.009179114364087582, + 0.0054048821330070496, + 0.00812022015452385, + 0.013472474180161953, + 0.010712028481066227, + 0.010725978761911392, + 0.0054382383823394775, + 0.008067273534834385, + 0.0072583360597491264, + 0.0034453882835805416, + 0.00245141563937068, + 0.005848359782248735, + 0.003715291852131486, + 0.0036275258753448725, + 0.001566073508001864, + 0.002746542915701866, + 0.002876585815101862, + 0.0032914550974965096, + 0.00257247150875628, + 0.004432064481079578, + 0.0026837007608264685, + 0.004971878603100777, + 0.002407871186733246, + 0.00335019757039845, + 0.0027183853089809418, + 0.01487381849437952, + 0.01113115344196558, + 0.030237114056944847, + 0.013499420136213303, + 0.010814477689564228, + 0.006574470549821854, + 0.009569377638399601, + 0.008870227262377739, + 0.0035063924733549356, + 0.0026397521141916513, + 0.0036179013550281525, + 0.003323829034343362, + 0.0027449659537523985, + 0.003880744567140937, + 0.0034894037526100874, + 0.0033353567123413086, + 0.011492010205984116, + 0.010859917849302292, + 0.020222827792167664, + 0.00974252074956894, + 0.008915858343243599, + 0.011963428929448128, + 0.01129222847521305, + 0.010654150508344173, + 0.010906137526035309, + 0.012364196591079235, + 0.009270472452044487, + 0.01023984793573618, + 0.009863643907010555, + 0.008926414884626865, + 0.005902361590415239, + 0.010054129175841808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.11290647089481354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in medium shot, slightly tilted upward, then slowly and steadily pulls back. The camera initially focuses on an elderly farmer, straining to hold a massive, textured, deep orange pumpkin. He is a dark, weathered man, his brow furrowed with exertion beneath his graying hair, his teeth clenched with clenched teeth, and beads of sweat beading on his forehead. He wears a faded blue plaid flannel shirt, sleeves rolled up to the elbows, revealing his muscular, veiny forearms. He wears dark brown canvas trousers stained with dirt. Bent over, his body trembling slightly from the weight of the pumpkin, he cautiously advances across the harvested land with slow, heavy steps. As the camera steadily pulls back, the image gradually widens, placing the farmer's figure within a larger context: a vast expanse of bountiful fields, shimmering in the golden evening sun. Only neatly harvested wheat stubble remained in the fields, long shadows stretching across the ground. On the distant horizon, a sparse row of trees and the silhouette of a small farmhouse could be seen. A soft, warm light shone a gentle glow over the scene, highlighting the tiny dust particles dancing in the air and creating an atmosphere of simplicity, tranquility, and contentment, silently celebrating hard work and the joy of a bountiful harvest.", + "seed": 4858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_10/index_9784_seed_4858_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": -0.020112136006355287, + "vlm_logprob": -0.020112136006355287, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.020112136006355287, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02308656834065914 + ], + "opposite_logprob_tokens": [ + -3.8980865478515625 + ], + "logprob_reward": -0.02308656834065914, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02308656834065914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04346740245819092, + 0.030311020091176033, + 0.08984555304050446, + 0.030286937952041626, + 0.08605760335922241, + 0.05644450709223747, + 0.03395073488354683, + 0.05806344375014305, + 0.5129321217536926, + 0.05864070728421211 + ], + "credit_map": [ + 0.003927216865122318, + 0.005963724106550217, + 0.003743038047105074, + 0.004446165170520544, + 0.007979796268045902, + 0.0036888145841658115, + 0.009719613939523697, + 0.003999034874141216, + 0.0024749820586293936, + 0.006219718139618635, + 0.003475070931017399, + 0.003373106475919485, + 0.004180944059044123, + 0.0019240279216319323, + 0.004891352728009224, + 0.003771816845983267, + 0.0032055105548352003, + 0.003356202272698283, + 0.024142060428857803, + 0.033065713942050934, + 0.004952526651322842, + 0.002715463051572442, + 0.012416206300258636, + 0.005991862155497074, + 0.002923747757449746, + 0.0028357738628983498, + 0.007186463102698326, + 0.00496518611907959, + 0.002664007479324937, + 0.0022000581957399845, + 0.005033598281443119, + 0.0024781024549156427, + 0.015368446707725525, + 0.003861980512738228, + 0.03585095331072807, + 0.009128721430897713, + 0.002789620775729418, + 0.0025315519887953997, + 0.011708620004355907, + 0.004817701876163483, + 0.006435886491090059, + 0.0048454985953867435, + 0.009888357482850552, + 0.004653492011129856, + 0.005458880215883255, + 0.011005121283233166, + 0.007427644915878773, + 0.006729628425091505, + 0.0026035236660391092, + 0.004336708690971136, + 0.009864398278295994, + 0.004603301640599966, + 0.002633689669892192, + 0.0016384879127144814, + 0.004294989164918661, + 0.003975636325776577, + 0.006308193318545818, + 0.01057291217148304, + 0.01594807393848896, + 0.008798211812973022, + 0.004170572850853205, + 0.0034219620283693075, + 0.004666722379624844, + 0.004176795948296785, + 0.1527715027332306, + 0.051803089678287506, + 0.1205693706870079, + 0.04187701642513275, + 0.030521726235747337, + 0.05141260847449303, + 0.018313730135560036, + 0.045663077384233475, + 0.005959033500403166, + 0.00880357250571251, + 0.013043920509517193, + 0.01353063527494669, + 0.005705788265913725, + 0.0033385350834578276, + 0.004253337159752846, + 0.004005884286016226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.02308656834065914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007652491331100464 + ], + "opposite_logprob_tokens": [ + -4.882652282714844 + ], + "logprob_reward": -0.007652491331100464, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007652491331100464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045008499175310135, + 0.05316108092665672, + 0.49553075432777405, + 0.15717999637126923, + 0.045500777661800385, + 0.05809818208217621, + 0.06386730819940567, + 0.018724776804447174, + 0.04334056377410889, + 0.019588034600019455 + ], + "credit_map": [ + 0.004552213475108147, + 0.0064374422654509544, + 0.005554853938519955, + 0.00465595256537199, + 0.006461107637733221, + 0.0036524285096675158, + 0.008749335072934628, + 0.004945165012031794, + 0.003980879206210375, + 0.013899363577365875, + 0.004225002136081457, + 0.005749288946390152, + 0.004437266383320093, + 0.002633757656440139, + 0.010098830796778202, + 0.008136695250868797, + 0.015877550467848778, + 0.02012077160179615, + 0.10974643379449844, + 0.1865559071302414, + 0.03246992081403732, + 0.012870932929217815, + 0.08994928747415543, + 0.027939971536397934, + 0.010762969963252544, + 0.014134594239294529, + 0.01942797563970089, + 0.03337913751602173, + 0.020065849646925926, + 0.014418300241231918, + 0.03022482991218567, + 0.014766339212656021, + 0.004697915632277727, + 0.0038742523174732924, + 0.01207074336707592, + 0.004502133000642061, + 0.006295355036854744, + 0.003800776321440935, + 0.005631826817989349, + 0.004627775400876999, + 0.0064290058799088, + 0.003669646568596363, + 0.01585913449525833, + 0.0035087254364043474, + 0.003279075725004077, + 0.017039595171809196, + 0.004382879473268986, + 0.003930121660232544, + 0.004880909807980061, + 0.006829415913671255, + 0.017077093943953514, + 0.013390292413532734, + 0.003596151014789939, + 0.0039324285462498665, + 0.00717000151053071, + 0.0069910138845443726, + 0.0019952445290982723, + 0.0026602395810186863, + 0.004083656240254641, + 0.002994868205860257, + 0.002177770482376218, + 0.0015669767744839191, + 0.0017033576732501388, + 0.0015426623867824674, + 0.013426010496914387, + 0.004284904338419437, + 0.00956022273749113, + 0.0037213207688182592, + 0.0030320773366838694, + 0.004164643120020628, + 0.0015187329845502973, + 0.0036326537374407053, + 0.002787449164316058, + 0.0024586438667029142, + 0.004080842714756727, + 0.004106765612959862, + 0.0016208578599616885, + 0.0014853193424642086, + 0.0013384876074269414, + 0.0017096676165238023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.007652491331100464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08544985949993134, + 0.10183477401733398, + 0.06629425287246704, + 0.06283210217952728, + 0.1970599889755249, + 0.19711542129516602, + 0.09403930604457855, + 0.08704070746898651, + 0.04164261370897293, + 0.06669099628925323 + ], + "credit_map": [ + 0.010297009721398354, + 0.010099965147674084, + 0.010517876595258713, + 0.009618125855922699, + 0.011170019395649433, + 0.007841896265745163, + 0.015948455780744553, + 0.009956501424312592, + 0.009636754170060158, + 0.013371774926781654, + 0.011079722084105015, + 0.011977484449744225, + 0.020615272223949432, + 0.00862782634794712, + 0.011048964224755764, + 0.015476960688829422, + 0.00268690031953156, + 0.003234616480767727, + 0.008633692748844624, + 0.02109021507203579, + 0.0063888258300721645, + 0.0029864811804145575, + 0.014674165286123753, + 0.0065993499010801315, + 0.005959338508546352, + 0.00479832012206316, + 0.0158843956887722, + 0.007969982922077179, + 0.008057087659835815, + 0.0047506229020655155, + 0.010313233360648155, + 0.005099120084196329, + 0.02045903168618679, + 0.014239240437746048, + 0.054936401546001434, + 0.015459141694009304, + 0.03626885637640953, + 0.017094364389777184, + 0.02027287892997265, + 0.01833004504442215, + 0.016357317566871643, + 0.023733271285891533, + 0.03161291033029556, + 0.013019612990319729, + 0.04208612069487572, + 0.019562482833862305, + 0.02826762944459915, + 0.022476056590676308, + 0.006913603749126196, + 0.009503181092441082, + 0.02039123699069023, + 0.01862870715558529, + 0.010827599093317986, + 0.0066880653612315655, + 0.012753089889883995, + 0.008333808742463589, + 0.007119921501725912, + 0.018432913348078728, + 0.009895307943224907, + 0.009117155335843563, + 0.015720752999186516, + 0.00762407248839736, + 0.011299734003841877, + 0.007830843329429626, + 0.010572710074484348, + 0.0046361880376935005, + 0.00485944515094161, + 0.004333986435085535, + 0.004597852937877178, + 0.004830397665500641, + 0.0031945405062288046, + 0.0046174898743629456, + 0.007626026403158903, + 0.005488844122737646, + 0.0161688644438982, + 0.017785029485821724, + 0.004374115262180567, + 0.005726035218685865, + 0.003257234813645482, + 0.006264836527407169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.009784122928977013 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634631097316742 + ], + "opposite_logprob_tokens": [ + -3.651346206665039 + ], + "logprob_reward": -0.02634631097316742, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634631097316742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10108115524053574, + 0.03036961704492569, + 0.13479427993297577, + 0.03240488842129707, + 0.0710722878575325, + 0.109803207218647, + 0.052838049829006195, + 0.08811365067958832, + 0.12453894317150116, + 0.2549838721752167 + ], + "credit_map": [ + 0.01086382381618023, + 0.010914619080722332, + 0.008744963444769382, + 0.015013575553894043, + 0.011356031522154808, + 0.011773483827710152, + 0.01937788724899292, + 0.013036779128015041, + 0.004522341303527355, + 0.003474630182608962, + 0.004285493399947882, + 0.0042795455083251, + 0.002758381189778447, + 0.0036890844348818064, + 0.0033950405195355415, + 0.003965100739151239, + 0.005840906873345375, + 0.008339042775332928, + 0.03972227871417999, + 0.04069099575281143, + 0.010906847193837166, + 0.0052570924162864685, + 0.014724981971085072, + 0.009312146343290806, + 0.0030072845984250307, + 0.003021799959242344, + 0.005305530037730932, + 0.0062539102509617805, + 0.003007348161190748, + 0.0042890156619250774, + 0.003946893848478794, + 0.0035731091629713774, + 0.011048433370888233, + 0.004630527924746275, + 0.02300107665359974, + 0.006795183755457401, + 0.0077241393737494946, + 0.003977079410105944, + 0.009214479476213455, + 0.004681367427110672, + 0.0114771518856287, + 0.008052382618188858, + 0.027101699262857437, + 0.008507605642080307, + 0.007342843804508448, + 0.023945767432451248, + 0.010841012932360172, + 0.012534739449620247, + 0.004489601124078035, + 0.00770772248506546, + 0.014801193960011005, + 0.007156141102313995, + 0.004114082548767328, + 0.002741351956501603, + 0.006074086297303438, + 0.005753872450441122, + 0.009048664011061192, + 0.010054024867713451, + 0.030353205278515816, + 0.01646248809993267, + 0.00632542185485363, + 0.004081397317349911, + 0.006219177972525358, + 0.005569273605942726, + 0.029104333370923996, + 0.01351213175803423, + 0.01614394225180149, + 0.009567885659635067, + 0.010682972148060799, + 0.022502345964312553, + 0.0073568010702729225, + 0.01566852256655693, + 0.022565478459000587, + 0.019310012459754944, + 0.06360961496829987, + 0.09280268847942352, + 0.01462281122803688, + 0.010601208545267582, + 0.015063892118632793, + 0.016408180817961693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.02634631097316742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.05904532968997955, + 0.04311954602599144, + 0.11584988981485367, + 0.041620880365371704, + 0.15135575830936432, + 0.10810072720050812, + 0.07775760442018509, + 0.08536580950021744, + 0.30031073093414307, + 0.017473772168159485 + ], + "credit_map": [ + 0.006157995201647282, + 0.007171047851443291, + 0.00601202342659235, + 0.0070137702859938145, + 0.006992529612034559, + 0.006492940243333578, + 0.012280329130589962, + 0.006924688816070557, + 0.004586519207805395, + 0.00496470183134079, + 0.0040130773559212685, + 0.004807286895811558, + 0.007914246991276741, + 0.005086581688374281, + 0.004919028840959072, + 0.006828104145824909, + 0.0063906507566571236, + 0.007285646162927151, + 0.030390284955501556, + 0.024199552834033966, + 0.011314206756651402, + 0.006208231206983328, + 0.018374839797616005, + 0.011686477810144424, + 0.003636116161942482, + 0.003912987187504768, + 0.00492944847792387, + 0.010064402595162392, + 0.004886865150183439, + 0.003834743984043598, + 0.006630107760429382, + 0.003726203925907612, + 0.019115444272756577, + 0.008583116345107555, + 0.05295102298259735, + 0.014864804223179817, + 0.016184750944375992, + 0.007951115258038044, + 0.01689525507390499, + 0.014810241758823395, + 0.01674458757042885, + 0.008692263625562191, + 0.011901593767106533, + 0.007375254761427641, + 0.009498830884695053, + 0.024657657369971275, + 0.015184511430561543, + 0.01404601987451315, + 0.00717128487303853, + 0.007422972470521927, + 0.023601964116096497, + 0.01371084339916706, + 0.004518479108810425, + 0.003986130002886057, + 0.008245945908129215, + 0.009099985472857952, + 0.007088059093803167, + 0.015005738474428654, + 0.022056162357330322, + 0.014251312240958214, + 0.008790381252765656, + 0.004622654989361763, + 0.008287444710731506, + 0.0052640559151768684, + 0.07198896259069443, + 0.023962179198861122, + 0.08420085906982422, + 0.03125625103712082, + 0.014671777375042439, + 0.026518968865275383, + 0.012462213635444641, + 0.0352494977414608, + 0.0018951250240206718, + 0.0019145194673910737, + 0.003914601635187864, + 0.004228363279253244, + 0.0016060809139162302, + 0.0012974472483620048, + 0.0009967652149498463, + 0.0016208691522479057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.03369118645787239 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in medium shot, slightly tilted upward, then slowly and steadily pulls back. The camera initially focuses on an elderly farmer, straining to hold a massive, textured, deep orange pumpkin. He is a dark, weathered man, his brow furrowed with exertion beneath his graying hair, his teeth clenched with clenched teeth, and beads of sweat beading on his forehead. He wears a faded blue plaid flannel shirt, sleeves rolled up to the elbows, revealing his muscular, veiny forearms. He wears dark brown canvas trousers stained with dirt. Bent over, his body trembling slightly from the weight of the pumpkin, he cautiously advances across the harvested land with slow, heavy steps. As the camera steadily pulls back, the image gradually widens, placing the farmer's figure within a larger context: a vast expanse of bountiful fields, shimmering in the golden evening sun. Only neatly harvested wheat stubble remained in the fields, long shadows stretching across the ground. On the distant horizon, a sparse row of trees and the silhouette of a small farmhouse could be seen. A soft, warm light shone a gentle glow over the scene, highlighting the tiny dust particles dancing in the air and creating an atmosphere of simplicity, tranquility, and contentment, silently celebrating hard work and the joy of a bountiful harvest.", + "seed": 4859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_10/index_9784_seed_4859_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": -0.051106905471533536, + "vlm_logprob": -0.051106905471533536, + "vlm_yes_no_margin": 4.0749999523162845, + "vlm_token_credit": -0.051106905471533536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2095409482717514 + ], + "opposite_logprob_tokens": [ + -1.7095409631729126 + ], + "logprob_reward": -0.2095409482717514, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2095409482717514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12996220588684082, + 0.1595137119293213, + 0.09708128869533539, + 0.18858210742473602, + 0.0691746175289154, + 0.03803739696741104, + 0.01797083206474781, + 0.07887828350067139, + 0.11912699788808823, + 0.10167255252599716 + ], + "credit_map": [ + 0.013316217809915543, + 0.011637032963335514, + 0.011732790619134903, + 0.021182693541049957, + 0.007480383850634098, + 0.024163728579878807, + 0.022028598934412003, + 0.018420755863189697, + 0.01106681302189827, + 0.014106165617704391, + 0.04571390897035599, + 0.01651414856314659, + 0.013654560782015324, + 0.012163580395281315, + 0.03516200929880142, + 0.011132560670375824, + 0.010304044932126999, + 0.01879918947815895, + 0.012151281349360943, + 0.012942830100655556, + 0.012263094075024128, + 0.007304222323000431, + 0.01274596992880106, + 0.01057067047804594, + 0.03323229402303696, + 0.028356501832604408, + 0.015103909187018871, + 0.01250735018402338, + 0.022314203903079033, + 0.03731239214539528, + 0.02038545534014702, + 0.01937001384794712, + 0.01450612023472786, + 0.015549813397228718, + 0.006306840572506189, + 0.007204838562756777, + 0.007711580488830805, + 0.005182089749723673, + 0.006004590541124344, + 0.006708748172968626, + 0.0034360717982053757, + 0.005790907423943281, + 0.005178921390324831, + 0.004093489609658718, + 0.004927733447402716, + 0.002254429506137967, + 0.005115487612783909, + 0.0072403582744300365, + 0.0021267759148031473, + 0.00187761674169451, + 0.0025787348859012127, + 0.0022472094278782606, + 0.0025113774463534355, + 0.0015385713195428252, + 0.0023054517805576324, + 0.002785096177831292, + 0.010721812956035137, + 0.005136802792549133, + 0.023967351764440536, + 0.0060616848058998585, + 0.005817498546093702, + 0.005418028216809034, + 0.004767707549035549, + 0.016987405717372894, + 0.00976716447621584, + 0.013109938241541386, + 0.039984919130802155, + 0.007258472964167595, + 0.008983328938484192, + 0.01121096033602953, + 0.009554014541208744, + 0.01925821043550968, + 0.012039950117468834, + 0.010181514546275139, + 0.013224313035607338, + 0.010936693288385868, + 0.013799676671624184, + 0.010312017053365707, + 0.017171338200569153, + 0.014007055200636387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.2095409482717514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028708456084132195 + ], + "opposite_logprob_tokens": [ + -5.877871036529541 + ], + "logprob_reward": -0.0028708456084132195, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028708456084132195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26973992586135864, + 0.16988898813724518, + 0.13943159580230713, + 0.08036056905984879, + 0.13108937442302704, + 0.03843947872519493, + 0.01730506867170334, + 0.046581026166677475, + 0.05996904522180557, + 0.04719493165612221 + ], + "credit_map": [ + 0.0267796628177166, + 0.024464454501867294, + 0.026524115353822708, + 0.039699405431747437, + 0.019477808848023415, + 0.06015099212527275, + 0.03408786654472351, + 0.038555584847927094, + 0.01291780173778534, + 0.03698867931962013, + 0.04053662344813347, + 0.0202795322984457, + 0.014142612926661968, + 0.00914961751550436, + 0.02687123417854309, + 0.009002875536680222, + 0.015500862151384354, + 0.019520241767168045, + 0.014101293869316578, + 0.015161930583417416, + 0.016161544248461723, + 0.014632821083068848, + 0.020305072888731956, + 0.024047819897532463, + 0.01938755251467228, + 0.008472356013953686, + 0.00988527201116085, + 0.005215608049184084, + 0.006433194037526846, + 0.016522901132702827, + 0.0073122261092066765, + 0.007131457794457674, + 0.028953468427062035, + 0.029220644384622574, + 0.014135909266769886, + 0.011673013679683208, + 0.014982477761805058, + 0.010574747808277607, + 0.010075054131448269, + 0.011474049650132656, + 0.0036354558542370796, + 0.004736988339573145, + 0.004481780342757702, + 0.0038318003062158823, + 0.003975406289100647, + 0.0020544584840536118, + 0.006532063242048025, + 0.009191521443426609, + 0.002125135390087962, + 0.0023215720430016518, + 0.0019188934238627553, + 0.0028841011226177216, + 0.002001108368858695, + 0.001566417166031897, + 0.002293235855177045, + 0.0021946027409285307, + 0.007666280027478933, + 0.004002868197858334, + 0.011982155032455921, + 0.004064666107296944, + 0.005074420943856239, + 0.003356623463332653, + 0.003569812048226595, + 0.006864197552204132, + 0.004291545134037733, + 0.00721305888146162, + 0.016021208837628365, + 0.0041737062856554985, + 0.005434147082269192, + 0.0042240675538778305, + 0.004661958664655685, + 0.013949349522590637, + 0.004520036280155182, + 0.006473977118730545, + 0.007965578697621822, + 0.005492562428116798, + 0.004701490513980389, + 0.005656900815665722, + 0.005024369806051254, + 0.007360011339187622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0028708456084132195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746016908437014 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006746016908437014, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746016908437014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09963512420654297, + 0.26765695214271545, + 0.09707404673099518, + 0.1334238201379776, + 0.05819430202245712, + 0.1559891402721405, + 0.02225535921752453, + 0.05358865112066269, + 0.058790408074855804, + 0.05339215323328972 + ], + "credit_map": [ + 0.008384336717426777, + 0.007571616675704718, + 0.007100273855030537, + 0.011535793542861938, + 0.0056981309317052364, + 0.028779208660125732, + 0.012476454488933086, + 0.018089313060045242, + 0.021579299122095108, + 0.09063944965600967, + 0.02068326622247696, + 0.03567425534129143, + 0.0370083749294281, + 0.015446973964571953, + 0.03531790152192116, + 0.011307453736662865, + 0.010052632540464401, + 0.0169829148799181, + 0.011512497439980507, + 0.0100975027307868, + 0.012113276869058609, + 0.008894059807062149, + 0.011757186613976955, + 0.015663981437683105, + 0.014037634246051311, + 0.020008830353617668, + 0.011852514930069447, + 0.015416259877383709, + 0.01738397404551506, + 0.023612432181835175, + 0.01599237322807312, + 0.015119797550141811, + 0.011989250779151917, + 0.012183086015284061, + 0.007246552035212517, + 0.004256794229149818, + 0.0061797890812158585, + 0.004530509002506733, + 0.00499818054959178, + 0.006810144986957312, + 0.01917141303420067, + 0.019316095858812332, + 0.01755836419761181, + 0.009981268085539341, + 0.014730608090758324, + 0.007738138549029827, + 0.018614472821354866, + 0.048878785222768784, + 0.0020056013017892838, + 0.002011594595387578, + 0.0024365540593862534, + 0.0034598142374306917, + 0.003182823769748211, + 0.0022953995503485203, + 0.0034497797023504972, + 0.0034137938637286425, + 0.009518220089375973, + 0.004540645983070135, + 0.010331712663173676, + 0.004911686759442091, + 0.007063833065330982, + 0.004773581400513649, + 0.00488272262737155, + 0.007566254120320082, + 0.005361340939998627, + 0.008126549422740936, + 0.0161991436034441, + 0.004527652636170387, + 0.005506465677171946, + 0.005488887894898653, + 0.006086073815822601, + 0.007494297344237566, + 0.005686379037797451, + 0.006092366762459278, + 0.0077600800432264805, + 0.006163755431771278, + 0.007297874893993139, + 0.0047835009172558784, + 0.008383273147046566, + 0.0072249239310622215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.006746016908437014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190113827586174 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018190113827586174, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018190113827586174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21070916950702667, + 0.056122198700904846, + 0.07278718799352646, + 0.13009296357631683, + 0.05160422623157501, + 0.044677767902612686, + 0.030889034271240234, + 0.10622493177652359, + 0.08482742309570312, + 0.21206513047218323 + ], + "credit_map": [ + 0.02393030747771263, + 0.02001398801803589, + 0.032357633113861084, + 0.05061738193035126, + 0.012803972698748112, + 0.018652642145752907, + 0.03749227523803711, + 0.014840963296592236, + 0.004636922851204872, + 0.009532565250992775, + 0.009455897845327854, + 0.007235226221382618, + 0.007344667334109545, + 0.005833564326167107, + 0.007756980136036873, + 0.004326375667005777, + 0.008026033639907837, + 0.01344479899853468, + 0.009519732557237148, + 0.007824289612472057, + 0.008545701391994953, + 0.006239823531359434, + 0.011233790777623653, + 0.007953021675348282, + 0.023145992308855057, + 0.018889710307121277, + 0.02054382860660553, + 0.014548267237842083, + 0.016561346128582954, + 0.011472664773464203, + 0.01359797827899456, + 0.011333171278238297, + 0.008106851018965244, + 0.006026502698659897, + 0.009580858051776886, + 0.003941410221159458, + 0.00406687380746007, + 0.005441843066364527, + 0.006533647887408733, + 0.007906241342425346, + 0.00511636259034276, + 0.0048021841794252396, + 0.007802348583936691, + 0.0035477911587804556, + 0.003513074479997158, + 0.003779712598770857, + 0.00588627252727747, + 0.01023002341389656, + 0.0028963680379092693, + 0.002504114294424653, + 0.004040861967951059, + 0.003771636402234435, + 0.0031705540604889393, + 0.0028947971295565367, + 0.004339741077274084, + 0.007270961068570614, + 0.015833161771297455, + 0.006804090924561024, + 0.035458117723464966, + 0.008240004070103168, + 0.006943853106349707, + 0.0058688269928097725, + 0.0078693563118577, + 0.01920752227306366, + 0.0072574918158352375, + 0.014343957416713238, + 0.02541286312043667, + 0.004137752577662468, + 0.004706989508122206, + 0.0035019181668758392, + 0.004625582601875067, + 0.020840875804424286, + 0.0333220548927784, + 0.04879869893193245, + 0.02980135567486286, + 0.01791587844491005, + 0.015099186450242996, + 0.014506389386951923, + 0.01828645169734955, + 0.034335117787122726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.018190113827586174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186602741479874 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186602741479874, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186602741479874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2115333378314972, + 0.12532950937747955, + 0.06904282420873642, + 0.09865696728229523, + 0.03160202503204346, + 0.03134583681821823, + 0.038029514253139496, + 0.1540382206439972, + 0.17806175351142883, + 0.06236007437109947 + ], + "credit_map": [ + 0.017079226672649384, + 0.00981482956558466, + 0.023817656561732292, + 0.015490858815610409, + 0.01057993434369564, + 0.06415551155805588, + 0.028530890122056007, + 0.04206442832946777, + 0.008858616463840008, + 0.012132475152611732, + 0.03514469787478447, + 0.01494365744292736, + 0.010280981659889221, + 0.009171709418296814, + 0.027633845806121826, + 0.007163518574088812, + 0.006696399301290512, + 0.014064662158489227, + 0.009625399485230446, + 0.009730491787195206, + 0.008603973314166069, + 0.006500024348497391, + 0.007133197039365768, + 0.006688680965453386, + 0.01963198371231556, + 0.012626467272639275, + 0.019468985497951508, + 0.010315565392374992, + 0.009616110473871231, + 0.007040772121399641, + 0.010587720200419426, + 0.00936936680227518, + 0.003131809178739786, + 0.0043527488596737385, + 0.0050878981128335, + 0.003536554053425789, + 0.0029028994031250477, + 0.002987435320392251, + 0.0035361561458557844, + 0.006066522095352411, + 0.003611629595980048, + 0.003380609443411231, + 0.004480150993913412, + 0.003674606326967478, + 0.0041173724457621574, + 0.0024712197482585907, + 0.003964401315897703, + 0.005645847879350185, + 0.0044301594607532024, + 0.0038080497179180384, + 0.005303534213453531, + 0.004456598777323961, + 0.00563128013163805, + 0.0024552263785153627, + 0.005777737591415644, + 0.006166926585137844, + 0.019536025822162628, + 0.011836175806820393, + 0.04926111921668053, + 0.011125344783067703, + 0.008284009993076324, + 0.007956777699291706, + 0.010948793031275272, + 0.03508998453617096, + 0.015878818929195404, + 0.03116619773209095, + 0.047775764018297195, + 0.008787439204752445, + 0.010417437180876732, + 0.008947692811489105, + 0.010661418549716473, + 0.04442697763442993, + 0.008236086927354336, + 0.010469402186572552, + 0.00960394274443388, + 0.006495102308690548, + 0.0061620017513632774, + 0.004600767511874437, + 0.006913323421031237, + 0.009879452176392078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.018186602741479874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in medium shot, slightly tilted upward, then slowly and steadily pulls back. The camera initially focuses on an elderly farmer, straining to hold a massive, textured, deep orange pumpkin. He is a dark, weathered man, his brow furrowed with exertion beneath his graying hair, his teeth clenched with clenched teeth, and beads of sweat beading on his forehead. He wears a faded blue plaid flannel shirt, sleeves rolled up to the elbows, revealing his muscular, veiny forearms. He wears dark brown canvas trousers stained with dirt. Bent over, his body trembling slightly from the weight of the pumpkin, he cautiously advances across the harvested land with slow, heavy steps. As the camera steadily pulls back, the image gradually widens, placing the farmer's figure within a larger context: a vast expanse of bountiful fields, shimmering in the golden evening sun. Only neatly harvested wheat stubble remained in the fields, long shadows stretching across the ground. On the distant horizon, a sparse row of trees and the silhouette of a small farmhouse could be seen. A soft, warm light shone a gentle glow over the scene, highlighting the tiny dust particles dancing in the air and creating an atmosphere of simplicity, tranquility, and contentment, silently celebrating hard work and the joy of a bountiful harvest.", + "seed": 4860, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_10/index_9784_seed_4860_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": -0.08439125763252378, + "vlm_logprob": -0.08439125763252378, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.08439125763252378, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23190876841545105 + ], + "opposite_logprob_tokens": [ + -1.6069087982177734 + ], + "logprob_reward": -0.23190876841545105, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23190876841545105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.272640585899353, + 0.026918409392237663, + 0.033903371542692184, + 0.028312047943472862, + 0.27946099638938904, + 0.03421652317047119, + 0.1795923411846161, + 0.06703143566846848, + 0.02407952956855297, + 0.05384476110339165 + ], + "credit_map": [ + 0.018836043775081635, + 0.05327432602643967, + 0.053896255791187286, + 0.03827189654111862, + 0.031152285635471344, + 0.020539993420243263, + 0.0303074698895216, + 0.026362307369709015, + 0.0020166542381048203, + 0.002414829796180129, + 0.0040151807479560375, + 0.0038030347786843777, + 0.002597160404548049, + 0.0030164450872689486, + 0.005922279320657253, + 0.003132828278467059, + 0.0032169607002288103, + 0.002604142064228654, + 0.005766961723566055, + 0.006246537435799837, + 0.0032406384125351906, + 0.003679809393361211, + 0.005279979202896357, + 0.0038683447055518627, + 0.0023539462126791477, + 0.0018222053768113256, + 0.004411894362419844, + 0.0031320988200604916, + 0.0047639054246246815, + 0.002505203941836953, + 0.006599122658371925, + 0.002723670331761241, + 0.021215932443737984, + 0.013985449448227882, + 0.0596061572432518, + 0.03613671287894249, + 0.07981640845537186, + 0.03177197650074959, + 0.018206659704446793, + 0.018721695989370346, + 0.0029170059133321047, + 0.002733440836891532, + 0.007921485230326653, + 0.005094499327242374, + 0.004192850086838007, + 0.0034004116896539927, + 0.00375471287406981, + 0.0042021190747618675, + 0.024769209325313568, + 0.01993861421942711, + 0.045340534299612045, + 0.03138487786054611, + 0.013451038859784603, + 0.007121250964701176, + 0.025632351636886597, + 0.0119544742628932, + 0.006561371497809887, + 0.004300746601074934, + 0.021838359534740448, + 0.01035919040441513, + 0.0042583150789141655, + 0.005572219844907522, + 0.009421021677553654, + 0.0047202082350850105, + 0.0020753261633217335, + 0.0017100548138841987, + 0.006214487832039595, + 0.0037891478277742863, + 0.002617295365780592, + 0.0015585210639983416, + 0.003911206964403391, + 0.0022034908179193735, + 0.0054764472879469395, + 0.004614224191755056, + 0.016427626833319664, + 0.010493157431483269, + 0.002976636867970228, + 0.002864888170734048, + 0.007337353657931089, + 0.0036544264294207096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.23190876841545105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043113868683576584 + ], + "opposite_logprob_tokens": [ + -3.168113946914673 + ], + "logprob_reward": -0.043113868683576584, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043113868683576584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0631558820605278, + 0.11054917424917221, + 0.05975956469774246, + 0.02764858677983284, + 0.5837118625640869, + 0.044876568019390106, + 0.03176255524158478, + 0.04805601015686989, + 0.0077210660092532635, + 0.022758737206459045 + ], + "credit_map": [ + 0.0036727541591972113, + 0.013914281502366066, + 0.010066073387861252, + 0.010249461978673935, + 0.00805798452347517, + 0.004438874777406454, + 0.007967225275933743, + 0.004789226688444614, + 0.005734790116548538, + 0.016065137460827827, + 0.011135206557810307, + 0.014569423161447048, + 0.01737549901008606, + 0.015517085790634155, + 0.018524574115872383, + 0.011627455241978168, + 0.004454629961401224, + 0.004742167424410582, + 0.010058719664812088, + 0.014572812244296074, + 0.005174939054995775, + 0.00511128781363368, + 0.009418709203600883, + 0.00622630026191473, + 0.001905181328766048, + 0.0026934375055134296, + 0.002363592153415084, + 0.0027077151462435722, + 0.006351619027554989, + 0.004307151772081852, + 0.004420167300850153, + 0.0028997217305004597, + 0.045744266360998154, + 0.029261058196425438, + 0.11790543049573898, + 0.0719422772526741, + 0.1708962321281433, + 0.06799596548080444, + 0.03927944228053093, + 0.0406871996819973, + 0.0042433287017047405, + 0.0037769381888210773, + 0.00944670382887125, + 0.006873851176351309, + 0.004075157456099987, + 0.002929368056356907, + 0.007679147180169821, + 0.005852075293660164, + 0.004152223467826843, + 0.003444741712883115, + 0.0058454410172998905, + 0.00471543800085783, + 0.0028585041873157024, + 0.002026643604040146, + 0.006273080129176378, + 0.0024464819580316544, + 0.004886562470346689, + 0.0037214611656963825, + 0.01608501560986042, + 0.007249940652400255, + 0.0031949919648468494, + 0.0025948805268853903, + 0.007206806447356939, + 0.003116352017968893, + 0.0007958753267303109, + 0.0006802026182413101, + 0.0013160125818103552, + 0.001212458941154182, + 0.0010509603889659047, + 0.0006646453402936459, + 0.0012916798004880548, + 0.0007092305459082127, + 0.002205910161137581, + 0.001668386859819293, + 0.0040864828042685986, + 0.003934813663363457, + 0.0020974313374608755, + 0.001598241738975048, + 0.004570258781313896, + 0.0025972104631364346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.043113868683576584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643500808626413 + ], + "opposite_logprob_tokens": [ + -4.882643699645996 + ], + "logprob_reward": -0.007643500808626413, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643500808626413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09461574256420135, + 0.02270369417965412, + 0.018791934475302696, + 0.07388120144605637, + 0.6441135406494141, + 0.019694168120622635, + 0.0794665664434433, + 0.024043478071689606, + 0.011125980876386166, + 0.01156365405768156 + ], + "credit_map": [ + 0.004877138417214155, + 0.021073175594210625, + 0.017405515536665916, + 0.012748001143336296, + 0.01120354700833559, + 0.007794386241585016, + 0.009190998040139675, + 0.01032298244535923, + 0.0014236546121537685, + 0.0026965830475091934, + 0.0030568644870072603, + 0.002411276800557971, + 0.003295374335721135, + 0.0037742697168141603, + 0.0034301155246794224, + 0.002615556353703141, + 0.0015587704256176949, + 0.0013071930734440684, + 0.00326678529381752, + 0.004010698292404413, + 0.0016322655137628317, + 0.0020140039268881083, + 0.0026725027710199356, + 0.0023297155275940895, + 0.003665725700557232, + 0.006647863425314426, + 0.006295974366366863, + 0.006026261020451784, + 0.019560178741812706, + 0.013712623156607151, + 0.010803639888763428, + 0.007168934680521488, + 0.05665008723735809, + 0.035474665462970734, + 0.09861680120229721, + 0.06497344374656677, + 0.20420782268047333, + 0.08156042546033859, + 0.0536462627351284, + 0.04898407310247421, + 0.0026375195011496544, + 0.0020726097282022238, + 0.0026091784238815308, + 0.00223480723798275, + 0.002176237292587757, + 0.0018538840813562274, + 0.002628432121127844, + 0.0034815012477338314, + 0.010810964740812778, + 0.009280717931687832, + 0.019311439245939255, + 0.013787646777927876, + 0.006135525181889534, + 0.003134624334052205, + 0.012005720287561417, + 0.004999939817935228, + 0.0023817538749426603, + 0.0020307544618844986, + 0.007332790642976761, + 0.0035380651243031025, + 0.0018029428320005536, + 0.0014058802044019103, + 0.003680251305922866, + 0.001871041371487081, + 0.0010875735897570848, + 0.0007351457607001066, + 0.002557758241891861, + 0.0015447448240593076, + 0.001974667189642787, + 0.0011758452747017145, + 0.0011521709384396672, + 0.0008980759303085506, + 0.001357261324301362, + 0.0012656071921810508, + 0.0025314046069979668, + 0.001977169420570135, + 0.001221307786181569, + 0.0007694382802583277, + 0.0017543508438393474, + 0.0006871164659969509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.007643500808626413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263565294444561 + ], + "opposite_logprob_tokens": [ + -3.6513564586639404 + ], + "logprob_reward": -0.0263565294444561, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263565294444561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07741310447454453, + 0.06146063283085823, + 0.03688555583357811, + 0.01941508986055851, + 0.4418032169342041, + 0.032545119524002075, + 0.24156567454338074, + 0.05553211644291878, + 0.005799693986773491, + 0.027579840272665024 + ], + "credit_map": [ + 0.005079869646579027, + 0.015145284123718739, + 0.014821439981460571, + 0.011231858283281326, + 0.008157030679285526, + 0.005450860597193241, + 0.011156270280480385, + 0.006370494142174721, + 0.0037306577432900667, + 0.009758676402270794, + 0.007432121783494949, + 0.010118305683135986, + 0.007986984215676785, + 0.007531542330980301, + 0.009153624065220356, + 0.005748721770942211, + 0.004178699571639299, + 0.001961574424058199, + 0.006237383931875229, + 0.005525920540094376, + 0.0028825344052165747, + 0.002361712045967579, + 0.010466177947819233, + 0.0032715534325689077, + 0.0013337895506992936, + 0.001669756369665265, + 0.004091999493539333, + 0.0036793414037674665, + 0.0020429303403943777, + 0.001333457650616765, + 0.002597159007564187, + 0.002666654996573925, + 0.03660018369555473, + 0.02356070838868618, + 0.07895244657993317, + 0.05044395476579666, + 0.1338907778263092, + 0.05356366187334061, + 0.03268229961395264, + 0.03210917487740517, + 0.004045041743665934, + 0.0019317887490615249, + 0.006650072522461414, + 0.00447193393483758, + 0.0029656870756298304, + 0.0024475071113556623, + 0.00489523122087121, + 0.005137855652719736, + 0.03235520422458649, + 0.028138775378465652, + 0.057088058441877365, + 0.04478522762656212, + 0.01828989014029503, + 0.010727900080382824, + 0.035699512809515, + 0.014481114223599434, + 0.005792805925011635, + 0.0028331861831247807, + 0.02242453396320343, + 0.00920418556779623, + 0.002725444734096527, + 0.002592088421806693, + 0.006892526056617498, + 0.003067346289753914, + 0.0005891616456210613, + 0.000288261566311121, + 0.0017586685717105865, + 0.0008324839291162789, + 0.0007208494353108108, + 0.00046114542055875063, + 0.0007135561318136752, + 0.0004355673445388675, + 0.0027115473058074713, + 0.0017505197320133448, + 0.008883719332516193, + 0.005892072804272175, + 0.0016641590045765042, + 0.00153442588634789, + 0.0036330053117126226, + 0.0015103910118341446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.0263565294444561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 farmer with graying hair and a weathered appearance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer holding a massive, textured, deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the farmer wearing a faded blue plaid flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a vast expanse of harvested fields with wheat stubble and long shadows under a golden evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot tilted upward and slowly pull back to reveal the wider landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293362081050873 + ], + "opposite_logprob_tokens": [ + -2.23793363571167 + ], + "logprob_reward": -0.11293362081050873, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293362081050873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0270115714520216, + 0.021723056212067604, + 0.009269612841308117, + 0.02533814124763012, + 0.765689492225647, + 0.026118526235222816, + 0.08310007303953171, + 0.025362253189086914, + 0.01094460766762495, + 0.0054426719434559345 + ], + "credit_map": [ + 0.0012123194755986333, + 0.00707315793260932, + 0.0037478741724044085, + 0.003433497156947851, + 0.003146691247820854, + 0.002440732903778553, + 0.0037703043781220913, + 0.0021869963966310024, + 0.0012840486597269773, + 0.0022311287466436625, + 0.003951410297304392, + 0.002184404293075204, + 0.0029431511647999287, + 0.003306053578853607, + 0.003130655037239194, + 0.0026922058314085007, + 0.0006785622681491077, + 0.0008026331197470427, + 0.0015053421957418323, + 0.0018232358852401376, + 0.0009467259515076876, + 0.0010170565219596028, + 0.0015352633781731129, + 0.0009607935789972544, + 0.0016822466859593987, + 0.0018122178735211492, + 0.003329339437186718, + 0.003107229946181178, + 0.0045158034190535545, + 0.002196872839704156, + 0.006352806463837624, + 0.002341624815016985, + 0.06623610109090805, + 0.04176487401127815, + 0.12075455486774445, + 0.07881966233253479, + 0.24186357855796814, + 0.09596405178308487, + 0.0627593994140625, + 0.05752728134393692, + 0.0035135624930262566, + 0.0017179951537400484, + 0.004621998872607946, + 0.0033424198627471924, + 0.00217974535189569, + 0.0024644595105201006, + 0.004104190040379763, + 0.0041741556487977505, + 0.011690754443407059, + 0.009522697888314724, + 0.01997125893831253, + 0.014501865021884441, + 0.006518653593957424, + 0.003982733469456434, + 0.011884195730090141, + 0.005027914885431528, + 0.0021538909059017897, + 0.001956277061253786, + 0.007739682216197252, + 0.0030710166320204735, + 0.0024766046553850174, + 0.001954885432496667, + 0.003391691716387868, + 0.002618207596242428, + 0.001078875968232751, + 0.000775630644056946, + 0.003361355746164918, + 0.0024397182278335094, + 0.000621416256763041, + 0.0003816258686129004, + 0.0014807370025664568, + 0.0008052492048591375, + 0.0005196743877604604, + 0.00040224302210845053, + 0.0014862752286717296, + 0.0009754587081260979, + 0.0005675131105817854, + 0.00033566210186108947, + 0.0007089592982083559, + 0.0004468863771762699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.11293362081050873 + } + ], + "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/0000028/group_11/index_tensor([5667, 5667])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_11/index_tensor([5667, 5667])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e291fe5f1a1eaf2a468ecd8559deb7848c616a59 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_11/index_tensor([5667, 5667])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, captured from a low, upward angle using a telephoto lens, freezes the epic confrontation atop the Yellow Crane Tower. In the frame, two warriors clad in full bronze armor face off atop the iconic yellow glazed tile roof. The telephoto lens compresses the space, making the turbulent sky behind them seem incredibly close and oppressive. The warrior on the left slumps slightly, his legs bent in a lunge, his center of gravity low. Beneath his beast-faced helmet, his piercing eyes fixate on his opponent, his right hand gripping the hilt of the longsword at his waist, veins visible. His bronze armor is intricately engraved with cloud and thunder patterns, and the slight scratches along its edges, worn by years of battle, reflect a cold, hard gleam in the gloomy sky. The warrior opposite him stands erect, his figure even more imposing, the long red tassel on his helmet whipping the air fiercely in the wind. He didn't place his hands on his weapons, but held his arms slightly outstretched. His posture appeared relaxed, but his set jaw and unyielding gaze betrayed a deep sense of vigilance. The glazed tiles beneath their feet shone with a damp sheen in the humid air. Against the backdrop was a vast, boundless sky, thick, dark clouds rolling and converging at a visible speed. Occasionally, a gap in the clouds would reveal a brief, dramatic burst of light, imbuing this fateful confrontation with a solemn and awe-inspiring atmosphere. The dominant tones of the scene, the gray-blue of the sky, the bright yellow of the roofs, and the dark copper of the armor, imbued the scene with a chilling, cinematic quality.", + "seed": 23634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_11/index_5667_seed_23634_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": -0.1600138608366251, + "vlm_logprob": -0.1600138608366251, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.1600138608366251, + "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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7002094388008118 + ], + "opposite_logprob_tokens": [ + -0.7002094388008118 + ], + "logprob_reward": -0.7002094388008118, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7002094388008118, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23852957785129547, + 0.06104356795549393, + 0.028093775734305382, + 0.039596471935510635, + 0.04469939321279526, + 0.394360214471817, + 0.12540097534656525, + 0.0361834391951561, + 0.009649018757045269, + 0.02244354970753193 + ], + "credit_map": [ + 0.043011374771595, + 0.0577477291226387, + 0.020659083500504494, + 0.020320754498243332, + 0.03378919139504433, + 0.024223975837230682, + 0.011681852862238884, + 0.0270956140011549, + 0.013937787152826786, + 0.009495939128100872, + 0.00522957555949688, + 0.0041679497808218, + 0.0063029928132891655, + 0.009361883625388145, + 0.005188222508877516, + 0.007359216921031475, + 0.0069985780864953995, + 0.0033917815890163183, + 0.0029270027298480272, + 0.0018053405219689012, + 0.0026161179412156343, + 0.004606103058904409, + 0.0020165853202342987, + 0.0037322677671909332, + 0.012285836972296238, + 0.005882224068045616, + 0.0022182539105415344, + 0.0017789591802284122, + 0.003333330387249589, + 0.01047444436699152, + 0.0018217804608866572, + 0.0018016400281339884, + 0.013455219566822052, + 0.0051401033997535706, + 0.0031497052405029535, + 0.00320049119181931, + 0.002931159920990467, + 0.010759043507277966, + 0.002218693494796753, + 0.0038449757266789675, + 0.17233604192733765, + 0.01410752348601818, + 0.01361799705773592, + 0.009297608397901058, + 0.015320953913033009, + 0.15574099123477936, + 0.005650021135807037, + 0.008289073593914509, + 0.034668534994125366, + 0.012747722677886486, + 0.012496759183704853, + 0.005894997622817755, + 0.011174307204782963, + 0.030425867065787315, + 0.006359956692904234, + 0.011632824316620827, + 0.005969775840640068, + 0.004339074715971947, + 0.003813566407188773, + 0.004169052001088858, + 0.004113435745239258, + 0.00408926559612155, + 0.003925233613699675, + 0.005764031317085028, + 0.001266066450625658, + 0.0027494526002556086, + 0.0008474802598357201, + 0.000700541480910033, + 0.0014655017293989658, + 0.0008324887603521347, + 0.0006444490281865001, + 0.0011430387385189533, + 0.003510585054755211, + 0.005437837913632393, + 0.002316561993211508, + 0.0016128397546708584, + 0.0035002294462174177, + 0.0020452053286135197, + 0.0015729747246950865, + 0.002447315491735935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.7002094388008118 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029832107946276665 + ], + "opposite_logprob_tokens": [ + -3.529832124710083 + ], + "logprob_reward": -0.029832107946276665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029832107946276665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20462319254875183, + 0.09603453427553177, + 0.11278341710567474, + 0.042755816131830215, + 0.03673364594578743, + 0.2559274137020111, + 0.06788423657417297, + 0.03576182946562767, + 0.11557343602180481, + 0.031922440975904465 + ], + "credit_map": [ + 0.04238809272646904, + 0.05069289356470108, + 0.020622173324227333, + 0.012174923904240131, + 0.018916118890047073, + 0.02164073847234249, + 0.01511580403894186, + 0.023072445765137672, + 0.018325578421354294, + 0.012764865532517433, + 0.01045140065252781, + 0.009061336517333984, + 0.007575163617730141, + 0.01548792514950037, + 0.009212478064000607, + 0.013155779801309109, + 0.032576803117990494, + 0.01041158102452755, + 0.007419025059789419, + 0.0069404831156134605, + 0.007515144068747759, + 0.03190368786454201, + 0.00595102971419692, + 0.01006566546857357, + 0.014653624035418034, + 0.004008335527032614, + 0.0027993666008114815, + 0.002511759987100959, + 0.0025633694604039192, + 0.012249057181179523, + 0.0017715846188366413, + 0.002198722679167986, + 0.01116219349205494, + 0.0031904284842312336, + 0.0031201671808958054, + 0.0021229407284408808, + 0.0034149789717048407, + 0.00859880167990923, + 0.0018866673344746232, + 0.0032374688889831305, + 0.10835853219032288, + 0.011203297413885593, + 0.009259150363504887, + 0.007848558947443962, + 0.010195893235504627, + 0.09822127223014832, + 0.005570791196078062, + 0.005269924178719521, + 0.02573527581989765, + 0.0032519581727683544, + 0.0038433445151895285, + 0.0020948676392436028, + 0.002832921454682946, + 0.025603188201785088, + 0.0019136755727231503, + 0.002609006594866514, + 0.006198690738528967, + 0.004320292267948389, + 0.0033836811780929565, + 0.004122294019907713, + 0.003992539830505848, + 0.0035384094808250666, + 0.005393550265580416, + 0.004812370520085096, + 0.018120065331459045, + 0.03865282982587814, + 0.009166017174720764, + 0.007185237016528845, + 0.016361873596906662, + 0.011160897091031075, + 0.006024582777172327, + 0.008901940658688545, + 0.00487238634377718, + 0.007524221669882536, + 0.0037793407682329416, + 0.0022965283133089542, + 0.004781805910170078, + 0.0033581703901290894, + 0.002133703790605068, + 0.003176284022629261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.029832107946276665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293547332286835 + ], + "opposite_logprob_tokens": [ + -3.7732934951782227 + ], + "logprob_reward": -0.023293547332286835, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293547332286835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24483288824558258, + 0.12057550996541977, + 0.08312932401895523, + 0.04057801514863968, + 0.07826612889766693, + 0.10887376219034195, + 0.14481137692928314, + 0.027079910039901733, + 0.12948176264762878, + 0.022371292114257812 + ], + "credit_map": [ + 0.04884671792387962, + 0.055915214121341705, + 0.02142578549683094, + 0.017944026738405228, + 0.02687627635896206, + 0.028351392596960068, + 0.013543838635087013, + 0.031929630786180496, + 0.01874830201268196, + 0.019920315593481064, + 0.010924717411398888, + 0.012362252920866013, + 0.0130003672093153, + 0.016705479472875595, + 0.012242604047060013, + 0.016671478748321533, + 0.01572412997484207, + 0.014045028015971184, + 0.008236661553382874, + 0.007025816477835178, + 0.010416156612336636, + 0.012909485958516598, + 0.005300586577504873, + 0.009471460245549679, + 0.009724173694849014, + 0.006594580132514238, + 0.003488639136776328, + 0.003804184962064028, + 0.0031405072659254074, + 0.00646206783130765, + 0.004087316803634167, + 0.0032765453215688467, + 0.023207804188132286, + 0.010515900328755379, + 0.006221660412847996, + 0.004956550896167755, + 0.004088542889803648, + 0.017853762954473495, + 0.0037056910805404186, + 0.007716224528849125, + 0.037391144782304764, + 0.00708438316360116, + 0.006973168347030878, + 0.004414573777467012, + 0.008302374742925167, + 0.034878216683864594, + 0.004915692377835512, + 0.0049142115749418736, + 0.049563948065042496, + 0.008812818676233292, + 0.011079869233071804, + 0.005060259252786636, + 0.007259346544742584, + 0.048112232238054276, + 0.006541614420711994, + 0.008381291292607784, + 0.004792111925780773, + 0.002540148328989744, + 0.002491831546649337, + 0.002932209288701415, + 0.0025932190474122763, + 0.002916760975494981, + 0.004422089084982872, + 0.004391540307551622, + 0.01842660643160343, + 0.04583146795630455, + 0.00973863061517477, + 0.008708043023943901, + 0.01929112896323204, + 0.012756324373185635, + 0.006465256214141846, + 0.008264300413429737, + 0.003708226839080453, + 0.0046580880880355835, + 0.002175744157284498, + 0.0018619103357195854, + 0.0026163419242948294, + 0.002487026620656252, + 0.002039917977526784, + 0.00282403570599854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.023293547332286835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863473117351532 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.00863473117351532, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863473117351532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19072091579437256, + 0.07768331468105316, + 0.1199122816324234, + 0.08075664192438126, + 0.05279317498207092, + 0.2806531488895416, + 0.10917636007070541, + 0.011687297374010086, + 0.05422044172883034, + 0.022396445274353027 + ], + "credit_map": [ + 0.032951563596725464, + 0.046523187309503555, + 0.016917044296860695, + 0.015764202922582626, + 0.026147466152906418, + 0.02074345387518406, + 0.009363495744764805, + 0.022310487926006317, + 0.01309142354875803, + 0.011858382262289524, + 0.007353358902037144, + 0.008226387202739716, + 0.009751394391059875, + 0.008554606698453426, + 0.00816366821527481, + 0.010684086009860039, + 0.02973663993179798, + 0.013043118640780449, + 0.011099777184426785, + 0.008116370067000389, + 0.009929072111845016, + 0.02747948467731476, + 0.007285439874976873, + 0.01322236843407154, + 0.02288910746574402, + 0.0112299844622612, + 0.005429097451269627, + 0.004550965968519449, + 0.009651890955865383, + 0.01956326514482498, + 0.003551560454070568, + 0.0038907702546566725, + 0.009865937754511833, + 0.01087872963398695, + 0.004099707584828138, + 0.005036240443587303, + 0.006767353508621454, + 0.007792254444211721, + 0.003008843632414937, + 0.005344105884432793, + 0.12224533408880234, + 0.010534150525927544, + 0.009823543019592762, + 0.006626589689403772, + 0.011331259272992611, + 0.10993680357933044, + 0.004483124241232872, + 0.005672310944646597, + 0.036816105246543884, + 0.0069662644527852535, + 0.008118808269500732, + 0.0038957016076892614, + 0.005968910176306963, + 0.03581881523132324, + 0.005257471930235624, + 0.006334279663860798, + 0.0024507613852620125, + 0.001708506839349866, + 0.0011295215226709843, + 0.0009507158538326621, + 0.0012066575000062585, + 0.0012929014628753066, + 0.0014552322681993246, + 0.0014929984463378787, + 0.009533625096082687, + 0.009787891060113907, + 0.005473271012306213, + 0.0036493935622274876, + 0.004076478537172079, + 0.005349447019398212, + 0.004851321689784527, + 0.0114990109577775, + 0.004107821732759476, + 0.005595221649855375, + 0.0019162535900250077, + 0.001388736767694354, + 0.003401577938348055, + 0.0022215673234313726, + 0.001496414770372212, + 0.0022688498720526695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.00863473117351532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809947893023491 + ], + "opposite_logprob_tokens": [ + -3.288099527359009 + ], + "logprob_reward": -0.03809947893023491, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809947893023491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04440166428685188, + 0.029829606413841248, + 0.02972804382443428, + 0.13295741379261017, + 0.10574803501367569, + 0.0688139945268631, + 0.4789508581161499, + 0.011482834815979004, + 0.05843169614672661, + 0.03965586796402931 + ], + "credit_map": [ + 0.00947253592312336, + 0.009606244042515755, + 0.0053508225828409195, + 0.0030223510693758726, + 0.003590415231883526, + 0.004660556558519602, + 0.003894857596606016, + 0.004803878255188465, + 0.005524338688701391, + 0.0056314547546207905, + 0.002822832902893424, + 0.00223755044862628, + 0.005701842252165079, + 0.00330081838183105, + 0.0022907874081283808, + 0.002319981809705496, + 0.008161223493516445, + 0.002546339761465788, + 0.002839654916897416, + 0.0018869423074647784, + 0.002547929296270013, + 0.007326622027903795, + 0.002189216436818242, + 0.0022301168646663427, + 0.03862718492746353, + 0.020962899550795555, + 0.006309137679636478, + 0.006651273462921381, + 0.014035110361874104, + 0.03482479974627495, + 0.003988029435276985, + 0.007558976300060749, + 0.021688634529709816, + 0.02209220454096794, + 0.009399600327014923, + 0.008956789039075375, + 0.010602282360196114, + 0.014251729473471642, + 0.006442360579967499, + 0.01231443602591753, + 0.02786477841436863, + 0.003071004059165716, + 0.002989378059282899, + 0.0022337441332638264, + 0.003548999084159732, + 0.025173936039209366, + 0.001966486219316721, + 0.001965669449418783, + 0.16758960485458374, + 0.03169051557779312, + 0.03019365295767784, + 0.015609822236001492, + 0.02599375881254673, + 0.16396035254001617, + 0.016813108697533607, + 0.02710004150867462, + 0.002091284841299057, + 0.0016662618145346642, + 0.0011344444938004017, + 0.001402605907060206, + 0.001195250661112368, + 0.0010905265808105469, + 0.0014993540244176984, + 0.0014031061436980963, + 0.00965131912380457, + 0.017181744799017906, + 0.005139509215950966, + 0.0033364538103342056, + 0.006856821943074465, + 0.005660173483192921, + 0.003561647143214941, + 0.0070440201088786125, + 0.005770822986960411, + 0.010829955339431763, + 0.003737410996109247, + 0.002611227100715041, + 0.008536075241863728, + 0.004110637586563826, + 0.00223446823656559, + 0.0018252693116664886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.03809947893023491 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, captured from a low, upward angle using a telephoto lens, freezes the epic confrontation atop the Yellow Crane Tower. In the frame, two warriors clad in full bronze armor face off atop the iconic yellow glazed tile roof. The telephoto lens compresses the space, making the turbulent sky behind them seem incredibly close and oppressive. The warrior on the left slumps slightly, his legs bent in a lunge, his center of gravity low. Beneath his beast-faced helmet, his piercing eyes fixate on his opponent, his right hand gripping the hilt of the longsword at his waist, veins visible. His bronze armor is intricately engraved with cloud and thunder patterns, and the slight scratches along its edges, worn by years of battle, reflect a cold, hard gleam in the gloomy sky. The warrior opposite him stands erect, his figure even more imposing, the long red tassel on his helmet whipping the air fiercely in the wind. He didn't place his hands on his weapons, but held his arms slightly outstretched. His posture appeared relaxed, but his set jaw and unyielding gaze betrayed a deep sense of vigilance. The glazed tiles beneath their feet shone with a damp sheen in the humid air. Against the backdrop was a vast, boundless sky, thick, dark clouds rolling and converging at a visible speed. Occasionally, a gap in the clouds would reveal a brief, dramatic burst of light, imbuing this fateful confrontation with a solemn and awe-inspiring atmosphere. The dominant tones of the scene, the gray-blue of the sky, the bright yellow of the roofs, and the dark copper of the armor, imbued the scene with a chilling, cinematic quality.", + "seed": 23635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_11/index_5667_seed_23635_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": -0.0842504463158548, + "vlm_logprob": -0.0842504463158548, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.0842504463158548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3552058935165405 + ], + "opposite_logprob_tokens": [ + -1.2302058935165405 + ], + "logprob_reward": -0.3552058935165405, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3552058935165405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27511948347091675, + 0.1052570566534996, + 0.06051096320152283, + 0.025965770706534386, + 0.033690229058265686, + 0.10339829325675964, + 0.0727178081870079, + 0.1022258847951889, + 0.10132791846990585, + 0.1197865828871727 + ], + "credit_map": [ + 0.05091347545385361, + 0.02792392112314701, + 0.044457707554101944, + 0.02181483618915081, + 0.026140756905078888, + 0.04403220862150192, + 0.015413805842399597, + 0.044422782957553864, + 0.020693404600024223, + 0.013623145408928394, + 0.01700695976614952, + 0.008857611566781998, + 0.008004057221114635, + 0.012107252143323421, + 0.008207354694604874, + 0.01675727218389511, + 0.01545433048158884, + 0.01000554021447897, + 0.005012114066630602, + 0.0035655880346894264, + 0.004564564675092697, + 0.009796703234314919, + 0.002992101013660431, + 0.009120023809373379, + 0.0033781167585402727, + 0.0027580696623772383, + 0.005542109254747629, + 0.0024377391673624516, + 0.002596180187538266, + 0.002696077572181821, + 0.0020603754092007875, + 0.004497102927416563, + 0.006744721904397011, + 0.0031196107156574726, + 0.004966708365827799, + 0.0034809778444468975, + 0.0023892014287412167, + 0.005360451061278582, + 0.0023647143971174955, + 0.005263843107968569, + 0.015512755140662193, + 0.01021028496325016, + 0.018882915377616882, + 0.007860453799366951, + 0.009886834770441055, + 0.010997580364346504, + 0.00729857524856925, + 0.022748885676264763, + 0.008343392983078957, + 0.008180602453649044, + 0.009658566676080227, + 0.01011593546718359, + 0.007811113726347685, + 0.010156083852052689, + 0.0070184580981731415, + 0.01143365167081356, + 0.010538498871028423, + 0.0076175471767783165, + 0.016050178557634354, + 0.015906736254692078, + 0.010585085488855839, + 0.01112212985754013, + 0.01370910182595253, + 0.016696609556674957, + 0.01506979763507843, + 0.02478010021150112, + 0.008284301497042179, + 0.014915434643626213, + 0.008858505636453629, + 0.01336780283600092, + 0.007265607360750437, + 0.008786378428339958, + 0.011430558748543262, + 0.01096663810312748, + 0.01498466357588768, + 0.019183337688446045, + 0.020897412672638893, + 0.012142466381192207, + 0.015351632609963417, + 0.014829868450760841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.3552058935165405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563959695398808 + ], + "opposite_logprob_tokens": [ + -4.387564182281494 + ], + "logprob_reward": -0.012563959695398808, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563959695398808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2714609205722809, + 0.13079331815242767, + 0.08541704714298248, + 0.06790120899677277, + 0.02888193354010582, + 0.06887847185134888, + 0.05194831266999245, + 0.07073681056499481, + 0.19366003572940826, + 0.030321946367621422 + ], + "credit_map": [ + 0.05370617285370827, + 0.03735683858394623, + 0.032251596450805664, + 0.025665804743766785, + 0.032988041639328, + 0.0409979447722435, + 0.020060699433088303, + 0.028433814644813538, + 0.034962158650159836, + 0.021318921819329262, + 0.00852672103792429, + 0.009774507954716682, + 0.014127240516245365, + 0.023512767627835274, + 0.008378393948078156, + 0.01019260287284851, + 0.02161562629044056, + 0.011877842247486115, + 0.004632375668734312, + 0.007742058020085096, + 0.00821218453347683, + 0.016720958054065704, + 0.007767699658870697, + 0.0068482947535812855, + 0.014946158044040203, + 0.008077992126345634, + 0.005604809150099754, + 0.006461931858211756, + 0.006415470968931913, + 0.012537575326859951, + 0.00789686106145382, + 0.0059604039415717125, + 0.003702371846884489, + 0.002784671960398555, + 0.003952020779252052, + 0.003900994313880801, + 0.003616234054788947, + 0.002838073531165719, + 0.003931232262402773, + 0.004156333860009909, + 0.012398301623761654, + 0.007394534070044756, + 0.011298221535980701, + 0.0076704444363713264, + 0.004561005160212517, + 0.010215479880571365, + 0.005326890852302313, + 0.010013588704168797, + 0.0059822299517691135, + 0.006504612509161234, + 0.007864472456276417, + 0.007362889125943184, + 0.005022767931222916, + 0.006061483174562454, + 0.004411107394844294, + 0.008738748729228973, + 0.009938783943653107, + 0.013411073945462704, + 0.008758297190070152, + 0.00928870402276516, + 0.007473886478692293, + 0.0072296131402254105, + 0.006309235468506813, + 0.008327214047312737, + 0.021967768669128418, + 0.029078219085931778, + 0.01718290150165558, + 0.03534027934074402, + 0.025505445897579193, + 0.018981683999300003, + 0.024694619700312614, + 0.020909098908305168, + 0.003630907740443945, + 0.0029642947483807802, + 0.004190121311694384, + 0.004444942343980074, + 0.00404253089800477, + 0.003213166492059827, + 0.004008523188531399, + 0.003827458480373025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.012563959695398808 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065435484051704 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065435484051704, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065435484051704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17366793751716614, + 0.07822097837924957, + 0.14291232824325562, + 0.09838530421257019, + 0.04408746957778931, + 0.1066630557179451, + 0.042149126529693604, + 0.09763125330209732, + 0.1668352484703064, + 0.04944728687405586 + ], + "credit_map": [ + 0.021488003432750702, + 0.0187701229006052, + 0.02363213337957859, + 0.02320200204849243, + 0.02034202590584755, + 0.01847757026553154, + 0.021819472312927246, + 0.025936607271432877, + 0.020882578566670418, + 0.011164261028170586, + 0.005653900094330311, + 0.0068095591850578785, + 0.007937069050967693, + 0.013633308932185173, + 0.00625556381419301, + 0.005884737707674503, + 0.03915839269757271, + 0.025211775675415993, + 0.009917165152728558, + 0.01163068599998951, + 0.010203249752521515, + 0.02608095109462738, + 0.009351993910968304, + 0.011358109302818775, + 0.01859666407108307, + 0.011440797708928585, + 0.008601486682891846, + 0.009641805663704872, + 0.011554939672350883, + 0.014786136336624622, + 0.013087183237075806, + 0.010676284320652485, + 0.00809532217681408, + 0.0052389963530004025, + 0.00563089782372117, + 0.005892511457204819, + 0.004036820027977228, + 0.005758350715041161, + 0.00426908815279603, + 0.005165481008589268, + 0.02550707571208477, + 0.011112343519926071, + 0.013998846523463726, + 0.008048669435083866, + 0.011000278405845165, + 0.022142961621284485, + 0.007025504019111395, + 0.00782738160341978, + 0.005251544527709484, + 0.005519599188119173, + 0.005305537488311529, + 0.005235901568084955, + 0.004596490412950516, + 0.006740896962583065, + 0.004832037258893251, + 0.00466711912304163, + 0.01115292590111494, + 0.011201689951121807, + 0.012822058983147144, + 0.014694929122924805, + 0.010922598652541637, + 0.011637951247394085, + 0.012353338301181793, + 0.012845756486058235, + 0.0210341177880764, + 0.031270917505025864, + 0.01384118665009737, + 0.028165193274617195, + 0.021068599075078964, + 0.016964364796876907, + 0.01800553686916828, + 0.016485342755913734, + 0.005118579603731632, + 0.006077657453715801, + 0.006224240176379681, + 0.006521592382341623, + 0.007335392292588949, + 0.006435273680835962, + 0.0058105564676225185, + 0.005923996213823557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.016065435484051704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622324395924807 + ], + "opposite_logprob_tokens": [ + -4.882622241973877 + ], + "logprob_reward": -0.007622324395924807, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622324395924807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1888197809457779, + 0.07281526178121567, + 0.1061115711927414, + 0.08392594009637833, + 0.04208318516612053, + 0.12569384276866913, + 0.07927785068750381, + 0.1628500521183014, + 0.08596178144216537, + 0.052460771054029465 + ], + "credit_map": [ + 0.03889038413763046, + 0.027728727087378502, + 0.020290862768888474, + 0.01668184995651245, + 0.017302334308624268, + 0.0255426038056612, + 0.017462367191910744, + 0.02492062933743, + 0.01304417010396719, + 0.006472792010754347, + 0.009178852662444115, + 0.006740395911037922, + 0.010517115704715252, + 0.009428715333342552, + 0.006275910884141922, + 0.01115730032324791, + 0.02858848311007023, + 0.017563220113515854, + 0.005959989503026009, + 0.009598398581147194, + 0.011378013528883457, + 0.01849205791950226, + 0.008228396065533161, + 0.006303011439740658, + 0.02065364085137844, + 0.011223547160625458, + 0.006150661502033472, + 0.007336643990129232, + 0.00918840803205967, + 0.015338137745857239, + 0.008393927477300167, + 0.005640973336994648, + 0.006170848850160837, + 0.004412379115819931, + 0.005432347301393747, + 0.006374018732458353, + 0.0057560959830880165, + 0.004808950237929821, + 0.004824618808925152, + 0.00430392287671566, + 0.031747035682201385, + 0.01342698372900486, + 0.01390068605542183, + 0.009293198585510254, + 0.01236528716981411, + 0.02613171935081482, + 0.007936326786875725, + 0.010892600752413273, + 0.009381751529872417, + 0.010922018438577652, + 0.01081069465726614, + 0.009275388903915882, + 0.007724633906036615, + 0.0144214928150177, + 0.007568910252302885, + 0.009172957390546799, + 0.016435351222753525, + 0.012438977137207985, + 0.026017192751169205, + 0.026775188744068146, + 0.019073843955993652, + 0.01728641428053379, + 0.023122191429138184, + 0.02170088142156601, + 0.010332422330975533, + 0.012216540053486824, + 0.009230232797563076, + 0.013211652636528015, + 0.011956636793911457, + 0.008876764215528965, + 0.008894897066056728, + 0.011242630891501904, + 0.005193001590669155, + 0.005545435938984156, + 0.006872384808957577, + 0.008046627044677734, + 0.008915122598409653, + 0.0057584261521697044, + 0.006017254199832678, + 0.006112514063715935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.007622324395924807 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794618487358093 + ], + "opposite_logprob_tokens": [ + -3.529794692993164 + ], + "logprob_reward": -0.029794618487358093, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794618487358093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09860222786664963, + 0.08307082951068878, + 0.06894287467002869, + 0.07213273644447327, + 0.03214116394519806, + 0.2337968349456787, + 0.04665077105164528, + 0.17411503195762634, + 0.1397315263748169, + 0.050816066563129425 + ], + "credit_map": [ + 0.01684287004172802, + 0.010088136419653893, + 0.01222627516835928, + 0.007960143499076366, + 0.012809633277356625, + 0.013989835977554321, + 0.008241874165832996, + 0.01644345186650753, + 0.020887503400444984, + 0.01182478666305542, + 0.006098579615354538, + 0.004822183400392532, + 0.008076244965195656, + 0.01738888956606388, + 0.0050211152993142605, + 0.008951517753303051, + 0.01129467785358429, + 0.006884985603392124, + 0.007677234709262848, + 0.007800854742527008, + 0.008081278763711452, + 0.008150001987814903, + 0.007998489774763584, + 0.011055344715714455, + 0.01294851116836071, + 0.009698554873466492, + 0.007748451083898544, + 0.008902842178940773, + 0.00727126793935895, + 0.008876972831785679, + 0.010904297232627869, + 0.005781835410743952, + 0.0051860930398106575, + 0.00452428450807929, + 0.0029986912850290537, + 0.0039380621165037155, + 0.005722982343286276, + 0.0031647353898733854, + 0.0030961011070758104, + 0.003510214388370514, + 0.07640201598405838, + 0.029980933293700218, + 0.012755991891026497, + 0.01059738453477621, + 0.020911453291773796, + 0.06506036967039108, + 0.009042170830070972, + 0.009046519175171852, + 0.005775545258074999, + 0.006658180151134729, + 0.0048313806764781475, + 0.006289737299084663, + 0.006320025306195021, + 0.007124265190213919, + 0.0034136392641812563, + 0.00623799255117774, + 0.01729297824203968, + 0.013564333319664001, + 0.025827297940850258, + 0.028032097965478897, + 0.018884597346186638, + 0.020710082724690437, + 0.024973968043923378, + 0.02482966147363186, + 0.01892980933189392, + 0.028238099068403244, + 0.012928652577102184, + 0.021755481138825417, + 0.01656607910990715, + 0.015052319504320621, + 0.012771660462021828, + 0.013489414937794209, + 0.005427461117506027, + 0.006901279091835022, + 0.006244793999940157, + 0.006775247398763895, + 0.008240330964326859, + 0.0057467930018901825, + 0.0057200053706765175, + 0.00576015142723918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.029794618487358093 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, captured from a low, upward angle using a telephoto lens, freezes the epic confrontation atop the Yellow Crane Tower. In the frame, two warriors clad in full bronze armor face off atop the iconic yellow glazed tile roof. The telephoto lens compresses the space, making the turbulent sky behind them seem incredibly close and oppressive. The warrior on the left slumps slightly, his legs bent in a lunge, his center of gravity low. Beneath his beast-faced helmet, his piercing eyes fixate on his opponent, his right hand gripping the hilt of the longsword at his waist, veins visible. His bronze armor is intricately engraved with cloud and thunder patterns, and the slight scratches along its edges, worn by years of battle, reflect a cold, hard gleam in the gloomy sky. The warrior opposite him stands erect, his figure even more imposing, the long red tassel on his helmet whipping the air fiercely in the wind. He didn't place his hands on his weapons, but held his arms slightly outstretched. His posture appeared relaxed, but his set jaw and unyielding gaze betrayed a deep sense of vigilance. The glazed tiles beneath their feet shone with a damp sheen in the humid air. Against the backdrop was a vast, boundless sky, thick, dark clouds rolling and converging at a visible speed. Occasionally, a gap in the clouds would reveal a brief, dramatic burst of light, imbuing this fateful confrontation with a solemn and awe-inspiring atmosphere. The dominant tones of the scene, the gray-blue of the sky, the bright yellow of the roofs, and the dark copper of the armor, imbued the scene with a chilling, cinematic quality.", + "seed": 23636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_11/index_5667_seed_23636_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": -0.0986558503471315, + "vlm_logprob": -0.0986558503471315, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.0986558503471315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43769022822380066 + ], + "opposite_logprob_tokens": [ + -1.062690258026123 + ], + "logprob_reward": -0.43769022822380066, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43769022822380066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4344814717769623, + 0.08251632004976273, + 0.0832914412021637, + 0.07829160988330841, + 0.04175727814435959, + 0.05005725473165512, + 0.05757111683487892, + 0.09113321453332901, + 0.030945148319005966, + 0.04995518550276756 + ], + "credit_map": [ + 0.064877450466156, + 0.03219016268849373, + 0.07034589350223541, + 0.0362948440015316, + 0.04046126455068588, + 0.048460014164447784, + 0.06286470592021942, + 0.07898713648319244, + 0.01148782204836607, + 0.005782111082226038, + 0.015910733491182327, + 0.0067955562844872475, + 0.010490433312952518, + 0.008791911415755749, + 0.00784511212259531, + 0.015412644483149052, + 0.010380995459854603, + 0.005697698798030615, + 0.018607329577207565, + 0.006105558015406132, + 0.012150748632848263, + 0.009059824980795383, + 0.005050331354141235, + 0.016238952055573463, + 0.010691055096685886, + 0.006913573015481234, + 0.011200610548257828, + 0.0066491663455963135, + 0.011871608905494213, + 0.009481986984610558, + 0.00705085601657629, + 0.014432751573622227, + 0.008093084208667278, + 0.0035387121606618166, + 0.008250000886619091, + 0.0032896501943469048, + 0.006509657949209213, + 0.004723280668258667, + 0.0026266155764460564, + 0.004726276732981205, + 0.007985723204910755, + 0.005287265870720148, + 0.0053556994535028934, + 0.0033009846229106188, + 0.01137937419116497, + 0.006609409581869841, + 0.005521547049283981, + 0.004617250058799982, + 0.007536235265433788, + 0.004237777553498745, + 0.008260641247034073, + 0.0032817504834383726, + 0.005785549059510231, + 0.0063666668720543385, + 0.0037670491728931665, + 0.01833544671535492, + 0.012237321585416794, + 0.007647736929357052, + 0.00691062631085515, + 0.009671666659414768, + 0.023126551881432533, + 0.011485866270959377, + 0.009916862472891808, + 0.010136580094695091, + 0.0039947680197656155, + 0.0027981132734566927, + 0.004285638220608234, + 0.006341991480439901, + 0.0036665932275354862, + 0.002967695239931345, + 0.003847329178825021, + 0.0030430196784436703, + 0.0072505539283156395, + 0.0034404892940074205, + 0.005480998195707798, + 0.008693956770002842, + 0.008952014148235321, + 0.005210586357861757, + 0.0074986061081290245, + 0.003427976742386818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.43769022822380066 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009804898872971535 + ], + "opposite_logprob_tokens": [ + -4.634804725646973 + ], + "logprob_reward": -0.009804898872971535, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009804898872971535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3168230652809143, + 0.08018084615468979, + 0.06157678738236427, + 0.06341560930013657, + 0.028198041021823883, + 0.10084687918424606, + 0.0974377915263176, + 0.0781092569231987, + 0.09798803925514221, + 0.07542368769645691 + ], + "credit_map": [ + 0.05800837278366089, + 0.030544092878699303, + 0.03457765281200409, + 0.02953840233385563, + 0.039957065135240555, + 0.040487825870513916, + 0.04056679457426071, + 0.04314284399151802, + 0.013532204553484917, + 0.007349330000579357, + 0.00975141953676939, + 0.006917995400726795, + 0.01367912907153368, + 0.010200745426118374, + 0.009566478431224823, + 0.009183543734252453, + 0.008021004498004913, + 0.0037383108865469694, + 0.01091835554689169, + 0.004820031579583883, + 0.008234985172748566, + 0.007055980619043112, + 0.006610160693526268, + 0.01217795442789793, + 0.007955714128911495, + 0.006255981978029013, + 0.009931200183928013, + 0.005392369348555803, + 0.011983383446931839, + 0.006351948250085115, + 0.005077150650322437, + 0.010467860847711563, + 0.004123975522816181, + 0.0025060572661459446, + 0.004676245152950287, + 0.0028417855501174927, + 0.0044829207472503185, + 0.0032747816294431686, + 0.0024872436188161373, + 0.0038050322327762842, + 0.013898072764277458, + 0.006058747414499521, + 0.017614349722862244, + 0.006648464128375053, + 0.013734453357756138, + 0.012697997502982616, + 0.007570328190922737, + 0.022624466568231583, + 0.017595527693629265, + 0.009548588655889034, + 0.013828206807374954, + 0.010055785067379475, + 0.01595931686460972, + 0.010863197967410088, + 0.006799498572945595, + 0.012787669897079468, + 0.01145558338612318, + 0.006755499169230461, + 0.007979202084243298, + 0.008050947450101376, + 0.01538178976625204, + 0.009071026928722858, + 0.008073432371020317, + 0.011341772973537445, + 0.012631815858185291, + 0.008602071553468704, + 0.009935911744832993, + 0.007566533982753754, + 0.024073440581560135, + 0.01232956349849701, + 0.011199363507330418, + 0.011649337597191334, + 0.009823181666433811, + 0.007499313913285732, + 0.009806928224861622, + 0.007472863886505365, + 0.01249518059194088, + 0.008214037865400314, + 0.012105913832783699, + 0.008006270974874496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.009804898872971535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195065014064312 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195065014064312, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195065014064312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17094232141971588, + 0.06081287935376167, + 0.05369884893298149, + 0.045526642352342606, + 0.058166176080703735, + 0.11384261399507523, + 0.09167402982711792, + 0.10378599166870117, + 0.07575560361146927, + 0.22579488158226013 + ], + "credit_map": [ + 0.023535577580332756, + 0.017841113731265068, + 0.018999069929122925, + 0.022459017112851143, + 0.03119942545890808, + 0.018166834488511086, + 0.015604527667164803, + 0.023136746138334274, + 0.007669541984796524, + 0.005237920209765434, + 0.005132385063916445, + 0.007887503132224083, + 0.013608194887638092, + 0.0074140895158052444, + 0.007036199327558279, + 0.006827045697718859, + 0.008276714943349361, + 0.003927732817828655, + 0.00571616692468524, + 0.005813213065266609, + 0.010116481222212315, + 0.006854485720396042, + 0.006599599961191416, + 0.0063944547437131405, + 0.0072198547422885895, + 0.005180106498301029, + 0.002989714965224266, + 0.003252946538850665, + 0.011395513080060482, + 0.00701493863016367, + 0.0059998491778969765, + 0.0024737149942666292, + 0.007362720090895891, + 0.005272847134619951, + 0.00796066876500845, + 0.004572777077555656, + 0.01370907574892044, + 0.006481755059212446, + 0.006295718718320131, + 0.006510612554848194, + 0.015423587523400784, + 0.007347264792770147, + 0.019213203340768814, + 0.0076725296676158905, + 0.012510636821389198, + 0.013528699986636639, + 0.008501345291733742, + 0.029645344242453575, + 0.011411095969378948, + 0.007276464253664017, + 0.013671617023646832, + 0.006698648910969496, + 0.007923535071313381, + 0.011234154924750328, + 0.007812971249222755, + 0.025645539164543152, + 0.015036072582006454, + 0.008639455772936344, + 0.009724130854010582, + 0.01117274072021246, + 0.01628115586936474, + 0.013224947266280651, + 0.013170584104955196, + 0.016536904498934746, + 0.01678427681326866, + 0.005536287557333708, + 0.008668792434036732, + 0.009788393042981625, + 0.008086828514933586, + 0.008121076971292496, + 0.008237537927925587, + 0.010532410815358162, + 0.02854001335799694, + 0.020364342257380486, + 0.032786983996629715, + 0.02773795835673809, + 0.03409138321876526, + 0.020258646458387375, + 0.03982460871338844, + 0.02219095453619957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.014195065014064312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052475095726549625 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052475095726549625, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052475095726549625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2606363594532013, + 0.08545982837677002, + 0.06553524732589722, + 0.06033998727798462, + 0.04458990320563316, + 0.04258478805422783, + 0.10095708817243576, + 0.06132259592413902, + 0.12576277554035187, + 0.15281149744987488 + ], + "credit_map": [ + 0.040183816105127335, + 0.021124251186847687, + 0.033244743943214417, + 0.026188040152192116, + 0.04216679185628891, + 0.027683021500706673, + 0.02676970697939396, + 0.043275970965623856, + 0.01155385747551918, + 0.007274334784597158, + 0.006834331899881363, + 0.008973941206932068, + 0.019959403201937675, + 0.010547648184001446, + 0.009534595534205437, + 0.01078170258551836, + 0.008827948942780495, + 0.0038787624798715115, + 0.011136746034026146, + 0.006355029996484518, + 0.009921944700181484, + 0.007709704339504242, + 0.006554445251822472, + 0.011150662787258625, + 0.00806928426027298, + 0.005912034306675196, + 0.006910811644047499, + 0.0047895112074911594, + 0.012999300844967365, + 0.007633273955434561, + 0.00686459057033062, + 0.007161179557442665, + 0.005846328102052212, + 0.003643693868070841, + 0.004695637617260218, + 0.0063035390339791775, + 0.006258787587285042, + 0.004444978199899197, + 0.005389834754168987, + 0.008007100783288479, + 0.006591560319066048, + 0.005048103630542755, + 0.004281745757907629, + 0.0045245043002069, + 0.008917367085814476, + 0.004608437418937683, + 0.0042071216739714146, + 0.004405944608151913, + 0.013822641223669052, + 0.010706804692745209, + 0.01632700301706791, + 0.011544558219611645, + 0.009803129360079765, + 0.011340539902448654, + 0.01058534998446703, + 0.01682705245912075, + 0.008469915948808193, + 0.006171173416078091, + 0.0065291342325508595, + 0.010763227939605713, + 0.008948253467679024, + 0.006797037087380886, + 0.00739519065245986, + 0.006248661782592535, + 0.01885664276778698, + 0.01216980628669262, + 0.011965575627982616, + 0.008917001076042652, + 0.03161980211734772, + 0.015216621570289135, + 0.013858451507985592, + 0.013158876448869705, + 0.018941817805171013, + 0.011128068901598454, + 0.020879795774817467, + 0.015045477077364922, + 0.02411593124270439, + 0.01623573526740074, + 0.026167377829551697, + 0.020297275856137276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0052475095726549625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341550052165985 + ], + "opposite_logprob_tokens": [ + -3.651341438293457 + ], + "logprob_reward": -0.026341550052165985, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341550052165985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09107032418251038, + 0.07890504598617554, + 0.060277730226516724, + 0.0717213824391365, + 0.09933574497699738, + 0.23669889569282532, + 0.20892518758773804, + 0.038744401186704636, + 0.038934655487537384, + 0.07538668066263199 + ], + "credit_map": [ + 0.014631740748882294, + 0.006671834271401167, + 0.011146330274641514, + 0.010931244120001793, + 0.010907778516411781, + 0.010740289464592934, + 0.011543412692844868, + 0.014497700147330761, + 0.013068526051938534, + 0.006982297170907259, + 0.01089002937078476, + 0.006687738932669163, + 0.007151747588068247, + 0.009560024365782738, + 0.008191529661417007, + 0.016373150050640106, + 0.013507753610610962, + 0.006218201946467161, + 0.007072793319821358, + 0.004838867112994194, + 0.009968044236302376, + 0.00812899973243475, + 0.0040917471051216125, + 0.0064513166435062885, + 0.011292639188468456, + 0.00623612804338336, + 0.01089122798293829, + 0.004785408265888691, + 0.011361614800989628, + 0.008526676334440708, + 0.0068963742814958096, + 0.01173130702227354, + 0.014388746581971645, + 0.007691750768572092, + 0.012160825543105602, + 0.0068243639543652534, + 0.02426336333155632, + 0.01225361693650484, + 0.009364882484078407, + 0.012388184666633606, + 0.03471264988183975, + 0.012401520274579525, + 0.042180705815553665, + 0.013240386731922626, + 0.024891166016459465, + 0.02935771457850933, + 0.016819730401039124, + 0.06309499591588974, + 0.031746264547109604, + 0.014876136556267738, + 0.030068401247262955, + 0.028130102902650833, + 0.022610152140259743, + 0.0224971491843462, + 0.018790770322084427, + 0.040206216275691986, + 0.006340468302369118, + 0.003482903353869915, + 0.003941188100725412, + 0.004542226437479258, + 0.006149000488221645, + 0.004807687364518642, + 0.0033725968096405268, + 0.006108326371759176, + 0.006212632171809673, + 0.0034684815909713507, + 0.0045844740234315395, + 0.006017428822815418, + 0.006476699374616146, + 0.003742561209946871, + 0.0047638900578022, + 0.003668486839160323, + 0.010159268043935299, + 0.007003652863204479, + 0.008145827800035477, + 0.010795949026942253, + 0.01381225697696209, + 0.008536883629858494, + 0.011098469607532024, + 0.005834363866597414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.026341550052165985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, captured from a low, upward angle using a telephoto lens, freezes the epic confrontation atop the Yellow Crane Tower. In the frame, two warriors clad in full bronze armor face off atop the iconic yellow glazed tile roof. The telephoto lens compresses the space, making the turbulent sky behind them seem incredibly close and oppressive. The warrior on the left slumps slightly, his legs bent in a lunge, his center of gravity low. Beneath his beast-faced helmet, his piercing eyes fixate on his opponent, his right hand gripping the hilt of the longsword at his waist, veins visible. His bronze armor is intricately engraved with cloud and thunder patterns, and the slight scratches along its edges, worn by years of battle, reflect a cold, hard gleam in the gloomy sky. The warrior opposite him stands erect, his figure even more imposing, the long red tassel on his helmet whipping the air fiercely in the wind. He didn't place his hands on his weapons, but held his arms slightly outstretched. His posture appeared relaxed, but his set jaw and unyielding gaze betrayed a deep sense of vigilance. The glazed tiles beneath their feet shone with a damp sheen in the humid air. Against the backdrop was a vast, boundless sky, thick, dark clouds rolling and converging at a visible speed. Occasionally, a gap in the clouds would reveal a brief, dramatic burst of light, imbuing this fateful confrontation with a solemn and awe-inspiring atmosphere. The dominant tones of the scene, the gray-blue of the sky, the bright yellow of the roofs, and the dark copper of the armor, imbued the scene with a chilling, cinematic quality.", + "seed": 23637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_11/index_5667_seed_23637_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": -0.10074036791920662, + "vlm_logprob": -0.10074036791920662, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.10074036791920662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3973524570465088 + ], + "opposite_logprob_tokens": [ + -1.1473524570465088 + ], + "logprob_reward": -0.3973524570465088, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3973524570465088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14784859120845795, + 0.093601293861866, + 0.04741639643907547, + 0.12113956362009048, + 0.2643206715583801, + 0.07900498807430267, + 0.041293900460004807, + 0.10259643197059631, + 0.06554564833641052, + 0.03723248094320297 + ], + "credit_map": [ + 0.012428024783730507, + 0.01706656627357006, + 0.018661556765437126, + 0.02547546662390232, + 0.020504716783761978, + 0.02158203348517418, + 0.015232373960316181, + 0.01689785346388817, + 0.007480224594473839, + 0.011104844510555267, + 0.012274079956114292, + 0.010327452793717384, + 0.020258091390132904, + 0.01414969190955162, + 0.007974384352564812, + 0.0100325345993042, + 0.008638176135718822, + 0.004095281939953566, + 0.006872424390166998, + 0.006748993881046772, + 0.005000611767172813, + 0.006898493506014347, + 0.005230296868830919, + 0.003932121675461531, + 0.015992777422070503, + 0.01554641593247652, + 0.013047299347817898, + 0.013558505102992058, + 0.025770993903279305, + 0.015122692100703716, + 0.010675046592950821, + 0.011425844393670559, + 0.03205282241106033, + 0.03493884205818176, + 0.02387365512549877, + 0.022141993045806885, + 0.06411947309970856, + 0.038201745599508286, + 0.022360796108841896, + 0.02663136087357998, + 0.007294975686818361, + 0.013968391343951225, + 0.007268944289535284, + 0.010561933740973473, + 0.009947765618562698, + 0.01029222086071968, + 0.007781709544360638, + 0.01188905630260706, + 0.004195668734610081, + 0.005412235856056213, + 0.0037509368266910315, + 0.004905392415821552, + 0.006914572324603796, + 0.005874263122677803, + 0.004382546525448561, + 0.005858289077877998, + 0.021160081028938293, + 0.013117155991494656, + 0.015529557131230831, + 0.009483798407018185, + 0.009386821649968624, + 0.013769884593784809, + 0.010537794791162014, + 0.00961134023964405, + 0.0074477726593613625, + 0.01686755381524563, + 0.005804153624922037, + 0.008840378373861313, + 0.008074408397078514, + 0.008722017519176006, + 0.005982029251754284, + 0.0038073367904871702, + 0.006812260486185551, + 0.006887440104037523, + 0.004478457383811474, + 0.004264605697244406, + 0.004633863922208548, + 0.004209184553474188, + 0.0033865126315504313, + 0.0025601570960134268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.3973524570465088 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609968952834606 + ], + "opposite_logprob_tokens": [ + -4.141099452972412 + ], + "logprob_reward": -0.01609968952834606, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609968952834606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20483873784542084, + 0.12262043356895447, + 0.10431598871946335, + 0.06756341457366943, + 0.06329941749572754, + 0.1563616842031479, + 0.03291837126016617, + 0.11384031921625137, + 0.07813575118780136, + 0.05610585957765579 + ], + "credit_map": [ + 0.016745537519454956, + 0.01879778504371643, + 0.022196097299456596, + 0.030307581648230553, + 0.04841306433081627, + 0.032409630715847015, + 0.020300723612308502, + 0.015668310225009918, + 0.012488001026213169, + 0.02044820599257946, + 0.010245955549180508, + 0.01152585819363594, + 0.02588539570569992, + 0.016884036362171173, + 0.014873750507831573, + 0.01026923954486847, + 0.009362378157675266, + 0.00796371791511774, + 0.011204179376363754, + 0.011214871890842915, + 0.030947018414735794, + 0.01495009008795023, + 0.011285785585641861, + 0.00738794868811965, + 0.009368706494569778, + 0.007429535035043955, + 0.006765585858374834, + 0.006067859940230846, + 0.014931551180779934, + 0.009671257808804512, + 0.0053721764124929905, + 0.007956739515066147, + 0.006737968418747187, + 0.008650063537061214, + 0.007825614884495735, + 0.007474791258573532, + 0.010012456215918064, + 0.006884698290377855, + 0.006312008015811443, + 0.009401816874742508, + 0.014620809815824032, + 0.04382199048995972, + 0.011564948596060276, + 0.01791277900338173, + 0.027083856984972954, + 0.019683105871081352, + 0.012153569608926773, + 0.009520627558231354, + 0.003294755006209016, + 0.005777332931756973, + 0.0034372282680124044, + 0.003232754999771714, + 0.004998273681849241, + 0.0041021062061190605, + 0.004016060847789049, + 0.0040598586201667786, + 0.01768907904624939, + 0.01831994019448757, + 0.013514836318790913, + 0.01169671956449747, + 0.01445608027279377, + 0.014089914038777351, + 0.013621843419969082, + 0.010451913811266422, + 0.009633995592594147, + 0.01853930950164795, + 0.007040924858301878, + 0.009177060797810555, + 0.010719680227339268, + 0.009524007327854633, + 0.008273382671177387, + 0.005227391142398119, + 0.007874825969338417, + 0.012134110555052757, + 0.005311809945851564, + 0.006650341674685478, + 0.00825163722038269, + 0.0052869911305606365, + 0.0063707358203828335, + 0.004225414711982012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01609968952834606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582424476742744 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582424476742744, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582424476742744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1984037309885025, + 0.13165581226348877, + 0.06961371749639511, + 0.07498028874397278, + 0.06829994916915894, + 0.15844911336898804, + 0.05210117995738983, + 0.1263086348772049, + 0.08035673946142197, + 0.039830803871154785 + ], + "credit_map": [ + 0.013733035884797573, + 0.018010880798101425, + 0.02245085872709751, + 0.02700764685869217, + 0.04773665592074394, + 0.031003955751657486, + 0.021559379994869232, + 0.016901323571801186, + 0.012412557378411293, + 0.025611819699406624, + 0.012152315117418766, + 0.019236024469137192, + 0.020505391061306, + 0.01848307065665722, + 0.014100409112870693, + 0.009154221974313259, + 0.010493495501577854, + 0.006415656767785549, + 0.008119801059365273, + 0.00805137399584055, + 0.012314284220337868, + 0.01034498866647482, + 0.006037041544914246, + 0.007837073877453804, + 0.011942722834646702, + 0.010014166124165058, + 0.006908091250807047, + 0.007807515561580658, + 0.015011685900390148, + 0.01115342602133751, + 0.005566794890910387, + 0.006575890816748142, + 0.006203933153301477, + 0.013194214552640915, + 0.004477871116250753, + 0.005441227927803993, + 0.01790684275329113, + 0.00983393657952547, + 0.007488340139389038, + 0.0037535829469561577, + 0.01426768209785223, + 0.03551008924841881, + 0.015312114730477333, + 0.016699286177754402, + 0.019928814843297005, + 0.020820297300815582, + 0.01171522494405508, + 0.024195604026317596, + 0.004540999419987202, + 0.009029646404087543, + 0.0054460191167891026, + 0.00955358799546957, + 0.007223288528621197, + 0.006487064529210329, + 0.004438210278749466, + 0.005382361821830273, + 0.020901743322610855, + 0.016797127202153206, + 0.015279630199074745, + 0.013782186433672905, + 0.016036316752433777, + 0.01579168066382408, + 0.016200648620724678, + 0.011519298888742924, + 0.010208520106971264, + 0.02033144421875477, + 0.0071363006718456745, + 0.010996328666806221, + 0.010693696327507496, + 0.009934857487678528, + 0.007006977219134569, + 0.004048612434417009, + 0.004661067388951778, + 0.009620122611522675, + 0.004247364588081837, + 0.005313330329954624, + 0.0051316628232598305, + 0.004867215175181627, + 0.0033937287516891956, + 0.0025963119696825743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.020582424476742744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624808698892593 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624808698892593, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624808698892593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13899578154087067, + 0.0966857522726059, + 0.24580694735050201, + 0.06348208338022232, + 0.06499391794204712, + 0.10352084040641785, + 0.06352268904447556, + 0.0842675194144249, + 0.07377991825342178, + 0.06494452059268951 + ], + "credit_map": [ + 0.013301166705787182, + 0.011481450870633125, + 0.01770579256117344, + 0.028521720319986343, + 0.016070285812020302, + 0.01631522737443447, + 0.016809366643428802, + 0.018790777772665024, + 0.007471215445548296, + 0.01199880801141262, + 0.0076331826858222485, + 0.009056509472429752, + 0.028957877308130264, + 0.016407178714871407, + 0.009218447841703892, + 0.005942537449300289, + 0.025458795949816704, + 0.015366657637059689, + 0.02058977261185646, + 0.028537938371300697, + 0.07316910475492477, + 0.03771485015749931, + 0.02688068524003029, + 0.01808914728462696, + 0.006460700184106827, + 0.006841781083494425, + 0.006735637318342924, + 0.006713838316500187, + 0.014460868202149868, + 0.009012392722070217, + 0.006613707635551691, + 0.006643161177635193, + 0.011423482559621334, + 0.011544021777808666, + 0.00547441653907299, + 0.005341075360774994, + 0.0116517199203372, + 0.010051165707409382, + 0.004087270237505436, + 0.005420778412371874, + 0.009896533563733101, + 0.031118664890527725, + 0.00822216086089611, + 0.010660162195563316, + 0.012481887824833393, + 0.013972167856991291, + 0.006712331902235746, + 0.010456939227879047, + 0.006107691675424576, + 0.010280857793986797, + 0.004739874973893166, + 0.007166137453168631, + 0.012796479277312756, + 0.0088426498696208, + 0.006138751748949289, + 0.007450249511748552, + 0.015002928674221039, + 0.01057033147662878, + 0.011831364594399929, + 0.00825938768684864, + 0.009705069474875927, + 0.011045179329812527, + 0.010326824150979519, + 0.007526439148932695, + 0.01064206101000309, + 0.018447337672114372, + 0.006732191890478134, + 0.008604150265455246, + 0.010798400267958641, + 0.0081449244171381, + 0.005620226263999939, + 0.004790632985532284, + 0.011177371256053448, + 0.014408488757908344, + 0.007199015934020281, + 0.00801975466310978, + 0.008031210862100124, + 0.006402626633644104, + 0.005168772768229246, + 0.0045372857712209225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.007624808698892593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 roof of the Yellow Crane Tower with yellow glazed tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two warriors visible wearing full bronze armor with cloud and thunder patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior on the left in a slumped lunge position with his right hand gripping a longsword hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior on the right standing erect with arms outstretched and a long red tassel on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a low upward angle using a telephoto lens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204245984554291 + ], + "opposite_logprob_tokens": [ + -2.812042474746704 + ], + "logprob_reward": -0.06204245984554291, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204245984554291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14997421205043793, + 0.10497893393039703, + 0.053054068237543106, + 0.09251970052719116, + 0.09471821039915085, + 0.19230833649635315, + 0.17510423064231873, + 0.05273550748825073, + 0.05377672612667084, + 0.030830085277557373 + ], + "credit_map": [ + 0.014813638292253017, + 0.015030005015432835, + 0.018420282751321793, + 0.020103925839066505, + 0.031373970210552216, + 0.021903229877352715, + 0.016143646091222763, + 0.012185534462332726, + 0.01168771367520094, + 0.016273990273475647, + 0.008745911531150341, + 0.009667277336120605, + 0.022841157391667366, + 0.01374609861522913, + 0.01235789991915226, + 0.009658888913691044, + 0.007112933788448572, + 0.007057262118905783, + 0.004761704709380865, + 0.008046831004321575, + 0.010017413645982742, + 0.007524312473833561, + 0.004769493360072374, + 0.0037641222588717937, + 0.009965975768864155, + 0.01412281859666109, + 0.00751490006223321, + 0.01377540361136198, + 0.018164189532399178, + 0.011580469086766243, + 0.008962385356426239, + 0.008433552458882332, + 0.013605926185846329, + 0.012740887701511383, + 0.009930318221449852, + 0.010221302509307861, + 0.018561432138085365, + 0.013785756193101406, + 0.0064919982105493546, + 0.009380588307976723, + 0.019220920279622078, + 0.02454335242509842, + 0.022534750401973724, + 0.025486325845122337, + 0.015895821154117584, + 0.02631717547774315, + 0.0177652258425951, + 0.0405447743833065, + 0.01987767592072487, + 0.02464655600488186, + 0.013623352162539959, + 0.016854485496878624, + 0.03475015610456467, + 0.02569103240966797, + 0.01763623021543026, + 0.02202475070953369, + 0.007403591647744179, + 0.007659723982214928, + 0.006780699826776981, + 0.004796018823981285, + 0.009243919514119625, + 0.007400372996926308, + 0.0054189283400774, + 0.004032256547361612, + 0.004880337044596672, + 0.009771659970283508, + 0.0067251576110720634, + 0.008812788873910904, + 0.005274193361401558, + 0.005984196905046701, + 0.007692582439631224, + 0.004635811783373356, + 0.003245993284508586, + 0.0051998840644955635, + 0.003879668889567256, + 0.004277571104466915, + 0.0038250626530498266, + 0.003207449335604906, + 0.0043740528635680676, + 0.0028204049449414015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.06204245984554291 + } + ], + "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/0000028/group_12/index_tensor([9885, 9885])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_12/index_tensor([9885, 9885])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..baeb815f44e6c24ce804ea85a4ca24a025e47e2f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_12/index_tensor([9885, 9885])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping wide shot begins at the center of a swirling energy vortex and zooms out at explosive speed, revealing the magnificent spectacle of the dawn of the universe: Amidst the boundless blackness of the void, countless shards of energy, composed of fiery red and deep blue, violently scatter in Van Gogh's signature whirlpools. These energies are not smooth gases, but entities with the texture of thick oil paintings, each twisting stream composed of a three-dimensional brushstroke. In the exquisite detail of 8K ultra-high-definition resolution, viewers can clearly see the microscopic ridges and gullies formed by the accumulation of pigments. The fiery red energy burns at the edges, colliding and merging with the deep blue, stirring up violent waves of violet and magenta. The entire image is filled with dynamic tension, as the vast vortex of energy expands outward with unstoppable force, tearing apart the tranquil background and creating an atmosphere of primordial, chaotic, and unparalleled grandeur.", + "seed": 14885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_12/index_9885_seed_14885_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": -0.5962710654363036, + "vlm_logprob": -0.5962710654363036, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.5962710654363036, + "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\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028870249167084694 + ], + "opposite_logprob_tokens": [ + -3.6538703441619873 + ], + "logprob_reward": -0.028870249167084694, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028870249167084694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14282912015914917, + 0.1347665637731552, + 0.041715845465660095, + 0.13605980575084686, + 0.07417688518762589, + 0.10487926751375198, + 0.17151743173599243, + 0.11401737481355667, + 0.035119809210300446, + 0.04491789638996124 + ], + "credit_map": [ + 0.02207830548286438, + 0.015925461426377296, + 0.019659100100398064, + 0.012549660168588161, + 0.016814248636364937, + 0.012615847401320934, + 0.01829746924340725, + 0.0248890183866024, + 0.01668049953877926, + 0.017275573685765266, + 0.01756170764565468, + 0.010379877872765064, + 0.012738746590912342, + 0.011315293610095978, + 0.024122746661305428, + 0.02469211257994175, + 0.005184502340853214, + 0.005670770071446896, + 0.004303422756493092, + 0.0029988784808665514, + 0.007901771925389767, + 0.0041014389134943485, + 0.004560731817036867, + 0.006994327995926142, + 0.014039123430848122, + 0.011069302447140217, + 0.010732856579124928, + 0.012937774881720543, + 0.019613856449723244, + 0.0290916096419096, + 0.020024374127388, + 0.018550900742411613, + 0.005669306498020887, + 0.004815060179680586, + 0.020639602094888687, + 0.005159363150596619, + 0.0065604038536548615, + 0.017437860369682312, + 0.0059531331062316895, + 0.007942156866192818, + 0.004948548506945372, + 0.008006214164197445, + 0.010222949087619781, + 0.00856047309935093, + 0.03371569141745567, + 0.01009205263108015, + 0.011316203512251377, + 0.018017131835222244, + 0.023632464930415154, + 0.014622152782976627, + 0.008525271899998188, + 0.010228528641164303, + 0.06834785640239716, + 0.016400141641497612, + 0.015632426366209984, + 0.014128589071333408, + 0.01278302539139986, + 0.0064820218831300735, + 0.009753103367984295, + 0.007865101099014282, + 0.007946512661874294, + 0.03460579365491867, + 0.008286166936159134, + 0.026295656338334084, + 0.003174164332449436, + 0.004949751775711775, + 0.0031078755855560303, + 0.0033538867719471455, + 0.0039496286772191525, + 0.004938748199492693, + 0.008439273573458195, + 0.0032064819242805243, + 0.0024930608924478292, + 0.003613766049966216, + 0.00387117313221097, + 0.0033041522838175297, + 0.0034559874329715967, + 0.0070435116067528725, + 0.005975978448987007, + 0.015160267241299152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.028870249167084694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.6950230598449707 + ], + "opposite_logprob_tokens": [ + -0.07002314925193787 + ], + "logprob_reward": -2.6950230598449707, + "yes_no_margin_reward": -2.625, + "token_credit_reward": -2.6950230598449707, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0732177197933197, + 0.17457856237888336, + 0.035661730915308, + 0.06465645134449005, + 0.2043176293373108, + 0.11446389555931091, + 0.10172813385725021, + 0.08660277724266052, + 0.11818691343069077, + 0.026586124673485756 + ], + "credit_map": [ + 0.00801097322255373, + 0.009617394767701626, + 0.0074199712835252285, + 0.00616077845916152, + 0.009575209580361843, + 0.009582211263477802, + 0.014412066899240017, + 0.00843912735581398, + 0.058635272085666656, + 0.01763932965695858, + 0.017319384962320328, + 0.006567543372511864, + 0.007653479464352131, + 0.01265561394393444, + 0.012063691392540932, + 0.042044274508953094, + 0.0044232327491045, + 0.0033537812996655703, + 0.002823603805154562, + 0.0039890180341899395, + 0.005910318344831467, + 0.003188603324815631, + 0.006194423884153366, + 0.005778755992650986, + 0.008417059667408466, + 0.007447457872331142, + 0.007590648718178272, + 0.009379825554788113, + 0.007617375813424587, + 0.008724107407033443, + 0.008810472674667835, + 0.0066695138812065125, + 0.014268219470977783, + 0.008393673226237297, + 0.06823206692934036, + 0.010583995841443539, + 0.014913108199834824, + 0.05805875360965729, + 0.009916970506310463, + 0.0199508685618639, + 0.00490358704701066, + 0.00876936037093401, + 0.014525388367474079, + 0.00613402621820569, + 0.03737177327275276, + 0.01370516698807478, + 0.005592242814600468, + 0.023462360724806786, + 0.016695015132427216, + 0.008080189116299152, + 0.006153583060950041, + 0.006830841768532991, + 0.036537040024995804, + 0.009901887737214565, + 0.009955721907317638, + 0.007573868613690138, + 0.02338813990354538, + 0.007006336469203234, + 0.00703709339722991, + 0.00752929225564003, + 0.007389626000076532, + 0.006801188923418522, + 0.010836522094905376, + 0.01661459356546402, + 0.012571842409670353, + 0.014314609579741955, + 0.014433825388550758, + 0.01408623717725277, + 0.013200276531279087, + 0.0177830271422863, + 0.017343804240226746, + 0.014453304000198841, + 0.0018742418615147471, + 0.0024662886280566454, + 0.002074457937851548, + 0.003300443524494767, + 0.0021594162099063396, + 0.003530611051246524, + 0.004074086435139179, + 0.007106582168489695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -2.6950230598449707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020652256906032562 + ], + "opposite_logprob_tokens": [ + -3.8956522941589355 + ], + "logprob_reward": -0.020652256906032562, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020652256906032562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04949074611067772, + 0.12441853433847427, + 0.04486884921789169, + 0.03721603751182556, + 0.2251366823911667, + 0.30758410692214966, + 0.04197348654270172, + 0.08610028028488159, + 0.058940574526786804, + 0.024270616471767426 + ], + "credit_map": [ + 0.006260406691581011, + 0.006620438769459724, + 0.005489352159202099, + 0.0038123808335512877, + 0.0090770423412323, + 0.0060178255662322044, + 0.00546944560483098, + 0.006743851117789745, + 0.019111379981040955, + 0.01212993822991848, + 0.01343337818980217, + 0.009618210606276989, + 0.011877900920808315, + 0.012823895551264286, + 0.023909058421850204, + 0.0215147752314806, + 0.00427264254540205, + 0.005179522093385458, + 0.0028180433437228203, + 0.003399481298401952, + 0.006799033377319574, + 0.00753282755613327, + 0.005901036784052849, + 0.008966263383626938, + 0.003119463101029396, + 0.006366042885929346, + 0.002220689319074154, + 0.0034464106429368258, + 0.00924082100391388, + 0.004094849806278944, + 0.0046744029968976974, + 0.004053360782563686, + 0.016197483986616135, + 0.008176603354513645, + 0.07337623834609985, + 0.011528480798006058, + 0.012658911757171154, + 0.06780730187892914, + 0.011070726439356804, + 0.024320941418409348, + 0.011749436147511005, + 0.022298982366919518, + 0.03308512642979622, + 0.014838832430541515, + 0.10696148127317429, + 0.035608772188425064, + 0.01579422503709793, + 0.06724725663661957, + 0.005162963178008795, + 0.0033949674107134342, + 0.0021120987366884947, + 0.0022767249029129744, + 0.015389945358037949, + 0.004545745439827442, + 0.00485317874699831, + 0.004237862769514322, + 0.013157997280359268, + 0.005475135054439306, + 0.006477137096226215, + 0.004965505097061396, + 0.005904105957597494, + 0.02255832590162754, + 0.007707173936069012, + 0.019854901358485222, + 0.0036515516694635153, + 0.006530751008540392, + 0.0055809603072702885, + 0.006505190394818783, + 0.007025481667369604, + 0.00791824609041214, + 0.015648147091269493, + 0.00608025211840868, + 0.0019367436179891229, + 0.0018387590534985065, + 0.0018073082901537418, + 0.0027489797212183475, + 0.0018453019438311458, + 0.00476572522893548, + 0.0029361573979258537, + 0.0063916416838765144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.020652256906032562 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2257213443517685 + ], + "opposite_logprob_tokens": [ + -1.6007213592529297 + ], + "logprob_reward": -0.2257213443517685, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2257213443517685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08306191116571426, + 0.07236351072788239, + 0.038636673241853714, + 0.07508783042430878, + 0.05509012192487717, + 0.22829154133796692, + 0.2720930278301239, + 0.08633911609649658, + 0.07588885724544525, + 0.013147439807653427 + ], + "credit_map": [ + 0.008442782796919346, + 0.013173621147871017, + 0.008419538848102093, + 0.007486326619982719, + 0.011821137741208076, + 0.011739770881831646, + 0.009624721482396126, + 0.01235400140285492, + 0.01808592863380909, + 0.00595029816031456, + 0.006222751457244158, + 0.004469839856028557, + 0.00663432851433754, + 0.007053556386381388, + 0.008927704766392708, + 0.015019095502793789, + 0.0045427256263792515, + 0.005159636028110981, + 0.0057640159502625465, + 0.006304611451923847, + 0.004540854133665562, + 0.0031863923650234938, + 0.003548799082636833, + 0.005589635577052832, + 0.00688181584700942, + 0.011467115953564644, + 0.005754769779741764, + 0.0086277537047863, + 0.015405148267745972, + 0.010398787446320057, + 0.009353391826152802, + 0.0071990350261330605, + 0.004737447015941143, + 0.00582311674952507, + 0.008001687936484814, + 0.0051760198548436165, + 0.005621136631816626, + 0.01233516912907362, + 0.006359877996146679, + 0.007035661023110151, + 0.009667053818702698, + 0.020443657413125038, + 0.019738398492336273, + 0.013735676184296608, + 0.0682574138045311, + 0.030816828832030296, + 0.014970909804105759, + 0.05066157132387161, + 0.04050947725772858, + 0.02206544205546379, + 0.011570111848413944, + 0.01572258584201336, + 0.10966730862855911, + 0.025292595848441124, + 0.02594096027314663, + 0.021324530243873596, + 0.008281180635094643, + 0.006736578419804573, + 0.0067911576479673386, + 0.004569451790302992, + 0.006042934022843838, + 0.02781335450708866, + 0.005184744019061327, + 0.02091970667243004, + 0.005286781582981348, + 0.009875801391899586, + 0.006536677945405245, + 0.007935154251754284, + 0.008878549560904503, + 0.009143080562353134, + 0.019886581227183342, + 0.00834622047841549, + 0.0015105654019862413, + 0.001287330174818635, + 0.0011866348795592785, + 0.0018612731946632266, + 0.0010360756423324347, + 0.0014435984194278717, + 0.002020461717620492, + 0.002801500027999282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.2257213443517685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088416911661625 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088416911661625, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088416911661625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08797813951969147, + 0.18450215458869934, + 0.03811042383313179, + 0.0379808135330677, + 0.08418677747249603, + 0.17784051597118378, + 0.1554199755191803, + 0.08827055990695953, + 0.11309982091188431, + 0.03261078894138336 + ], + "credit_map": [ + 0.009595057927072048, + 0.012456376105546951, + 0.011261280626058578, + 0.00740279583260417, + 0.015135352499783039, + 0.010303066112101078, + 0.010705334134399891, + 0.011118876747786999, + 0.059617504477500916, + 0.02035428024828434, + 0.015812896192073822, + 0.007894815877079964, + 0.012812401168048382, + 0.012587104924023151, + 0.013920975849032402, + 0.04150219261646271, + 0.005181418731808662, + 0.003950311802327633, + 0.004904937464743853, + 0.006811060942709446, + 0.004287265241146088, + 0.0035637293476611376, + 0.005665169097483158, + 0.0037465309724211693, + 0.004620036575943232, + 0.0035150395706295967, + 0.004242449998855591, + 0.004399208817631006, + 0.006408967077732086, + 0.00608460558578372, + 0.0045028431341052055, + 0.004207662306725979, + 0.006147258915007114, + 0.005851458292454481, + 0.02076428383588791, + 0.005950568709522486, + 0.008130684494972229, + 0.020677845925092697, + 0.005814963020384312, + 0.010849718004465103, + 0.006406731903553009, + 0.012800437398254871, + 0.023227915167808533, + 0.00865039974451065, + 0.06458202004432678, + 0.018716812133789062, + 0.00856033805757761, + 0.0348958894610405, + 0.013046267442405224, + 0.013413084670901299, + 0.0073353806510567665, + 0.009662231430411339, + 0.059833262115716934, + 0.01504878792911768, + 0.015154631808400154, + 0.02192634344100952, + 0.017852358520030975, + 0.0062608602456748486, + 0.006213207729160786, + 0.0056826528161764145, + 0.007010541390627623, + 0.015953870490193367, + 0.009241416119039059, + 0.020055659115314484, + 0.00965391006320715, + 0.007583616301417351, + 0.011906740255653858, + 0.010068033821880817, + 0.00984242931008339, + 0.030196160078048706, + 0.01979213021695614, + 0.014056805521249771, + 0.0030447698663920164, + 0.003007109509781003, + 0.0025980749633163214, + 0.004783856216818094, + 0.002467257436364889, + 0.0035478186327964067, + 0.005348951555788517, + 0.007812953554093838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.011088416911661625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins at the center of a swirling energy vortex and zooms out at explosive speed, revealing the magnificent spectacle of the dawn of the universe: Amidst the boundless blackness of the void, countless shards of energy, composed of fiery red and deep blue, violently scatter in Van Gogh's signature whirlpools. These energies are not smooth gases, but entities with the texture of thick oil paintings, each twisting stream composed of a three-dimensional brushstroke. In the exquisite detail of 8K ultra-high-definition resolution, viewers can clearly see the microscopic ridges and gullies formed by the accumulation of pigments. The fiery red energy burns at the edges, colliding and merging with the deep blue, stirring up violent waves of violet and magenta. The entire image is filled with dynamic tension, as the vast vortex of energy expands outward with unstoppable force, tearing apart the tranquil background and creating an atmosphere of primordial, chaotic, and unparalleled grandeur.", + "seed": 14886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_12/index_9885_seed_14886_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": -0.3862283671274781, + "vlm_logprob": -0.3862283671274781, + "vlm_yes_no_margin": 2.425000023841858, + "vlm_token_credit": -0.3862283671274781, + "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\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05746353045105934 + ], + "opposite_logprob_tokens": [ + -2.9324636459350586 + ], + "logprob_reward": -0.05746353045105934, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05746353045105934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20486968755722046, + 0.20722897350788116, + 0.03814031183719635, + 0.15435734391212463, + 0.1760721206665039, + 0.060526058077812195, + 0.023650581017136574, + 0.05246537923812866, + 0.034475333988666534, + 0.04821418970823288 + ], + "credit_map": [ + 0.01723380945622921, + 0.021225634962320328, + 0.020537806674838066, + 0.03450151905417442, + 0.020114967599511147, + 0.025837497785687447, + 0.03146085888147354, + 0.03395760804414749, + 0.00862967874854803, + 0.009144569747149944, + 0.08034662902355194, + 0.06622502207756042, + 0.00851877685636282, + 0.012098376639187336, + 0.013086586259305477, + 0.009179322980344296, + 0.0034721186384558678, + 0.003055272623896599, + 0.009435664862394333, + 0.0029145493172109127, + 0.0036352539900690317, + 0.007352478802204132, + 0.002961567370221019, + 0.005313407629728317, + 0.017577530816197395, + 0.02005097083747387, + 0.035296279937028885, + 0.02063838019967079, + 0.011721006594598293, + 0.014538802206516266, + 0.019597817212343216, + 0.014936555176973343, + 0.030407464131712914, + 0.015758249908685684, + 0.029924292117357254, + 0.01669158600270748, + 0.01049137208610773, + 0.018131621181964874, + 0.0252760648727417, + 0.02939148060977459, + 0.005017826799303293, + 0.005600322037935257, + 0.009826559573411942, + 0.006347657181322575, + 0.005452979821711779, + 0.012913334183394909, + 0.007683685049414635, + 0.007683693431317806, + 0.003552740439772606, + 0.0020126854069530964, + 0.0021951156668365, + 0.003458628663793206, + 0.0021113441325724125, + 0.004696140065789223, + 0.002569003961980343, + 0.003054920816794038, + 0.005209576338529587, + 0.006025277078151703, + 0.01249216590076685, + 0.007006691303104162, + 0.007992122322320938, + 0.002778919879347086, + 0.006496906280517578, + 0.004463716875761747, + 0.004003507550805807, + 0.0031782379373908043, + 0.004153053276240826, + 0.0054262131452560425, + 0.0027188165113329887, + 0.0030720399226993322, + 0.005470767617225647, + 0.006452697794884443, + 0.0063632032833993435, + 0.005419210996478796, + 0.00411875918507576, + 0.00627120491117239, + 0.005429266486316919, + 0.004988167434930801, + 0.005050776991993189, + 0.010573601350188255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.05746353045105934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.7015564441680908 + ], + "opposite_logprob_tokens": [ + -0.2015565037727356 + ], + "logprob_reward": -1.7015564441680908, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015564441680908, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20148442685604095, + 0.39727044105529785, + 0.023083869367837906, + 0.0281393863260746, + 0.059530872851610184, + 0.09237868338823318, + 0.04936655983328819, + 0.08731679618358612, + 0.03407770395278931, + 0.027351323515176773 + ], + "credit_map": [ + 0.015177569352090359, + 0.022845152765512466, + 0.022222064435482025, + 0.03157489746809006, + 0.022790011018514633, + 0.02153303660452366, + 0.030153078958392143, + 0.03518861532211304, + 0.013531992211937904, + 0.01631828583776951, + 0.15630733966827393, + 0.1376805603504181, + 0.01443630363792181, + 0.019211914390325546, + 0.023727023974061012, + 0.016056997701525688, + 0.002093520713970065, + 0.0020002240780740976, + 0.005465032998472452, + 0.002306065522134304, + 0.002823607763275504, + 0.0028810061048716307, + 0.0025020309258252382, + 0.0030123821925371885, + 0.0022490539122372866, + 0.004723218735307455, + 0.0039813825860619545, + 0.005511228460818529, + 0.003249090863391757, + 0.0021708221174776554, + 0.003872934030368924, + 0.002381655154749751, + 0.005718831904232502, + 0.006896620616316795, + 0.008831353858113289, + 0.01212325133383274, + 0.005020948592573404, + 0.0036606478970497847, + 0.013650735840201378, + 0.003628484206274152, + 0.00899145845323801, + 0.008269491605460644, + 0.012435082346200943, + 0.010453557595610619, + 0.007717772852629423, + 0.023266101256012917, + 0.011243913322687149, + 0.010001306422054768, + 0.010754186660051346, + 0.004649740643799305, + 0.005600604694336653, + 0.005554722622036934, + 0.004303451161831617, + 0.006098578684031963, + 0.005743635818362236, + 0.006661637220531702, + 0.009395278058946133, + 0.006474783644080162, + 0.031060578301548958, + 0.012373557314276695, + 0.003284614998847246, + 0.004382110200822353, + 0.007450112607330084, + 0.012895761989057064, + 0.004028047434985638, + 0.002867364091798663, + 0.004736815579235554, + 0.006525791250169277, + 0.002041526837274432, + 0.002841058885678649, + 0.005366326775401831, + 0.005670773331075907, + 0.003136842278763652, + 0.002762551885098219, + 0.002015190664678812, + 0.004322170279920101, + 0.0022601501550525427, + 0.0031184342224150896, + 0.002636641263961792, + 0.007099342066794634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -1.7015564441680908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023318586871027946 + ], + "opposite_logprob_tokens": [ + -3.7733185291290283 + ], + "logprob_reward": -0.023318586871027946, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023318586871027946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0494406521320343, + 0.3458850681781769, + 0.06599204987287521, + 0.10609106719493866, + 0.17219360172748566, + 0.057894445955753326, + 0.05736742913722992, + 0.07244076579809189, + 0.0429423525929451, + 0.02975250408053398 + ], + "credit_map": [ + 0.004061183426529169, + 0.005686139687895775, + 0.011101805604994297, + 0.009473149664700031, + 0.006076206918805838, + 0.003401403548195958, + 0.0052483174949884415, + 0.004392446018755436, + 0.013643424957990646, + 0.01453192625194788, + 0.1360323280096054, + 0.11698469519615173, + 0.012746626511216164, + 0.01755121909081936, + 0.02006141096353531, + 0.014333453960716724, + 0.007870855741202831, + 0.0047883084043860435, + 0.009794002398848534, + 0.00520847225561738, + 0.0072713978588581085, + 0.015574002638459206, + 0.006325544323772192, + 0.009159472770988941, + 0.00924033671617508, + 0.022410010918974876, + 0.03081471286714077, + 0.010681123472750187, + 0.008006603457033634, + 0.0066389720886945724, + 0.00904749147593975, + 0.009251824580132961, + 0.029243294149637222, + 0.016408639028668404, + 0.026681210845708847, + 0.020076802000403404, + 0.01137872040271759, + 0.01598040573298931, + 0.02512594871222973, + 0.02729858085513115, + 0.006058600265532732, + 0.006084791850298643, + 0.007242449093610048, + 0.005644155200570822, + 0.006152155809104443, + 0.012941395863890648, + 0.008138641715049744, + 0.0056322552263736725, + 0.008027422241866589, + 0.005362992640584707, + 0.005882877390831709, + 0.008077835664153099, + 0.008493127301335335, + 0.005804754327982664, + 0.010510070249438286, + 0.005208352580666542, + 0.00695766368880868, + 0.008223896846175194, + 0.01620784029364586, + 0.010130600072443485, + 0.010243488475680351, + 0.004194587469100952, + 0.007959806360304356, + 0.008522889576852322, + 0.00440182676538825, + 0.003273620968684554, + 0.005031507927924395, + 0.006385582499206066, + 0.0035258650314062834, + 0.003949521575123072, + 0.007384766358882189, + 0.008989663794636726, + 0.0029246944468468428, + 0.0027941109146922827, + 0.004256156738847494, + 0.004254844971001148, + 0.00433740857988596, + 0.003141621593385935, + 0.0028398761060088873, + 0.005203790031373501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.023318586871027946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285315573215485 + ], + "opposite_logprob_tokens": [ + -2.017853260040283 + ], + "logprob_reward": -0.14285315573215485, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14285315573215485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03606182709336281, + 0.7417720556259155, + 0.05218268558382988, + 0.06642176955938339, + 0.03197336569428444, + 0.02270418033003807, + 0.021456683054566383, + 0.016263023018836975, + 0.00768346106633544, + 0.0034810598008334637 + ], + "credit_map": [ + 0.0033710328862071037, + 0.003687898861244321, + 0.0043566543608903885, + 0.0040632192976772785, + 0.0036189190577715635, + 0.005037823226302862, + 0.00584762915968895, + 0.006078644655644894, + 0.024875648319721222, + 0.02940005622804165, + 0.29213786125183105, + 0.25813302397727966, + 0.02755291759967804, + 0.03594443202018738, + 0.043842364102602005, + 0.02988559380173683, + 0.005163293797522783, + 0.004519984126091003, + 0.007148965261876583, + 0.004977028351277113, + 0.0055586970411241055, + 0.01512295100837946, + 0.004562949296087027, + 0.005128809716552496, + 0.004973916802555323, + 0.01615557074546814, + 0.022405149415135384, + 0.0063204215839505196, + 0.003830528352409601, + 0.0032130545005202293, + 0.00536204082891345, + 0.004161075223237276, + 0.0036939112469553947, + 0.0036949727218598127, + 0.0033225619699805975, + 0.005642229691147804, + 0.002732192864641547, + 0.002509373240172863, + 0.0060698166489601135, + 0.0043082996271550655, + 0.0024379619862884283, + 0.0020197294652462006, + 0.0026747030206024647, + 0.0027129941154271364, + 0.0017830798169597983, + 0.004850292578339577, + 0.0027360159438103437, + 0.003489400027319789, + 0.004692332819104195, + 0.002103158738464117, + 0.0024640613701194525, + 0.0028166831471025944, + 0.0019358935533091426, + 0.002632016083225608, + 0.0025082977954298258, + 0.00230423710308969, + 0.0022450927644968033, + 0.0016780381556600332, + 0.004805449862033129, + 0.002321781124919653, + 0.0012824906734749675, + 0.0010053744772449136, + 0.0016377356369048357, + 0.00128705776296556, + 0.000993827125057578, + 0.0007477214094251394, + 0.0010260092094540596, + 0.0010266229510307312, + 0.0008364313980564475, + 0.0008447589934803545, + 0.0009769708849489689, + 0.001231117988936603, + 0.0004668624314945191, + 0.0003795177908614278, + 0.0004820033209398389, + 0.0005335669848136604, + 0.0004002294153906405, + 0.0003110193938482553, + 0.0002774594759102911, + 0.0006304005510173738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.14285315573215485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.041207391768693924, + 0.5369603633880615, + 0.03651629015803337, + 0.017135929316282272, + 0.18294450640678406, + 0.07179953902959824, + 0.03440408781170845, + 0.04246005415916443, + 0.028093133121728897, + 0.00847867876291275 + ], + "credit_map": [ + 0.003783868858590722, + 0.003835896961390972, + 0.008465439081192017, + 0.005397793836891651, + 0.004227522760629654, + 0.004682453814893961, + 0.00667537422850728, + 0.004139040131121874, + 0.018416481092572212, + 0.022373586893081665, + 0.2109382003545761, + 0.18520909547805786, + 0.02056840993463993, + 0.026284152641892433, + 0.031192844733595848, + 0.021977588534355164, + 0.004275558050721884, + 0.002984160091727972, + 0.004989404231309891, + 0.002276771701872349, + 0.003544884966686368, + 0.009929086081683636, + 0.002674960531294346, + 0.005841465201228857, + 0.001522804843261838, + 0.0026333683636039495, + 0.0028405680786818266, + 0.002674030838534236, + 0.001954158768057823, + 0.0016281248535960913, + 0.0023123263381421566, + 0.0015705461846664548, + 0.03221723437309265, + 0.0164644755423069, + 0.030195899307727814, + 0.015563873574137688, + 0.012364518828690052, + 0.019478000700473785, + 0.02458249032497406, + 0.03207801282405853, + 0.0084241246804595, + 0.0067956759594380856, + 0.011852182447910309, + 0.007543371524661779, + 0.006056840997189283, + 0.01552591659128666, + 0.0077458214946091175, + 0.007855602540075779, + 0.006673044059425592, + 0.004306373652070761, + 0.003358890302479267, + 0.004809366073459387, + 0.003527228021994233, + 0.004645566456019878, + 0.003485105698928237, + 0.0035985142458230257, + 0.004304464440792799, + 0.004638609476387501, + 0.00916311889886856, + 0.006277803797274828, + 0.004048163071274757, + 0.0025352132506668568, + 0.004429509397596121, + 0.00706317275762558, + 0.003143617184832692, + 0.002274221507832408, + 0.0033769819419831038, + 0.004516339395195246, + 0.0020496705546975136, + 0.002676901873201132, + 0.00456191087141633, + 0.005493489094078541, + 0.000987286097370088, + 0.0011991356732323766, + 0.0009200583444908261, + 0.0013132118619978428, + 0.0011656605638563633, + 0.0008596647530794144, + 0.000876197125762701, + 0.0011574642267078161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.005950118415057659 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins at the center of a swirling energy vortex and zooms out at explosive speed, revealing the magnificent spectacle of the dawn of the universe: Amidst the boundless blackness of the void, countless shards of energy, composed of fiery red and deep blue, violently scatter in Van Gogh's signature whirlpools. These energies are not smooth gases, but entities with the texture of thick oil paintings, each twisting stream composed of a three-dimensional brushstroke. In the exquisite detail of 8K ultra-high-definition resolution, viewers can clearly see the microscopic ridges and gullies formed by the accumulation of pigments. The fiery red energy burns at the edges, colliding and merging with the deep blue, stirring up violent waves of violet and magenta. The entire image is filled with dynamic tension, as the vast vortex of energy expands outward with unstoppable force, tearing apart the tranquil background and creating an atmosphere of primordial, chaotic, and unparalleled grandeur.", + "seed": 14887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_12/index_9885_seed_14887_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": -0.3827072709798813, + "vlm_logprob": -0.3827072709798813, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.3827072709798813, + "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\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.050186049193143845 + ], + "opposite_logprob_tokens": [ + -3.0501861572265625 + ], + "logprob_reward": -0.050186049193143845, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.050186049193143845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32630231976509094, + 0.01701493002474308, + 0.007215666119009256, + 0.008385231718420982, + 0.014670299366116524, + 0.02672414295375347, + 0.019402803853154182, + 0.5424147248268127, + 0.018390415236353874, + 0.019479520618915558 + ], + "credit_map": [ + 0.05296827480196953, + 0.010139023885130882, + 0.01110028475522995, + 0.007969208993017673, + 0.07411644607782364, + 0.13730430603027344, + 0.009301962330937386, + 0.02340279519557953, + 0.004657427780330181, + 0.0012147779343649745, + 0.001184232416562736, + 0.0007732905214652419, + 0.005728048272430897, + 0.001239109318703413, + 0.0009593063150532544, + 0.0012587372912093997, + 0.0009200623608194292, + 0.001246063970029354, + 0.0008748177206143737, + 0.0007358374423347414, + 0.0008046049624681473, + 0.0006778122624382377, + 0.0009177651954814792, + 0.0010387024376541376, + 0.0011213583638891578, + 0.00091364182299003, + 0.00076654291478917, + 0.0010676681995391846, + 0.0006062398315407336, + 0.000935120100621134, + 0.0014366074465215206, + 0.0015380537370219827, + 0.002464140998199582, + 0.0016065703239291906, + 0.002118333475664258, + 0.002006927505135536, + 0.0009473605314269662, + 0.0014261649921536446, + 0.0015465597389265895, + 0.0025542420335114002, + 0.006500134244561195, + 0.0019824698101729155, + 0.007095459382981062, + 0.0024513842072337866, + 0.0018781650578603148, + 0.0023061607498675585, + 0.0018846727907657623, + 0.002625697525218129, + 0.0028888878878206015, + 0.0026312656700611115, + 0.003602065145969391, + 0.0043213823810219765, + 0.001215206808410585, + 0.0014148236950859427, + 0.0013345173792913556, + 0.0019946557004004717, + 0.028097810223698616, + 0.2134464830160141, + 0.1780509501695633, + 0.037001486867666245, + 0.0333089716732502, + 0.022301435470581055, + 0.0138743557035923, + 0.016333214938640594, + 0.002937265206128359, + 0.0028333067893981934, + 0.0014644070761278272, + 0.0023235336411744356, + 0.0020263302139937878, + 0.0020063116680830717, + 0.002002531662583351, + 0.0027967304922640324, + 0.0025194301269948483, + 0.0014071966288611293, + 0.0028825087938457727, + 0.0018171756528317928, + 0.0017573900986462831, + 0.004772739484906197, + 0.001689152792096138, + 0.002633926924318075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.050186049193143845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.8048802614212036 + ], + "opposite_logprob_tokens": [ + -0.179880291223526 + ], + "logprob_reward": -1.8048802614212036, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048802614212036, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6100730895996094, + 0.08017370104789734, + 0.02105068601667881, + 0.013425693847239017, + 0.01889011636376381, + 0.01497390866279602, + 0.012632765807211399, + 0.17675736546516418, + 0.02575223334133625, + 0.026270415633916855 + ], + "credit_map": [ + 0.10866786539554596, + 0.017936697229743004, + 0.02296106517314911, + 0.015711598098278046, + 0.10763974487781525, + 0.26578840613365173, + 0.017189430072903633, + 0.054178327322006226, + 0.010704667307436466, + 0.007911259308457375, + 0.005663097370415926, + 0.006935358513146639, + 0.022242344915866852, + 0.007849440909922123, + 0.005173372104763985, + 0.013694162480533123, + 0.003299211384728551, + 0.003415843239054084, + 0.0029503311961889267, + 0.0015198809560388327, + 0.002034998731687665, + 0.002239566994830966, + 0.001469151582568884, + 0.004121703561395407, + 0.0017339924816042185, + 0.001734093064442277, + 0.0017265165224671364, + 0.0015913993120193481, + 0.001269014086574316, + 0.0012727752327919006, + 0.002065348671749234, + 0.002032554941251874, + 0.002806824166327715, + 0.0031130558345466852, + 0.0027742506936192513, + 0.0030306370463222265, + 0.0013361251913011074, + 0.002003100235015154, + 0.0016944140661507845, + 0.0021317116916179657, + 0.0038876181934028864, + 0.0014897930668666959, + 0.0031115838792175055, + 0.0011829291470348835, + 0.0011371326399967074, + 0.0018159841420128942, + 0.0010168786393478513, + 0.0013319897698238492, + 0.0014928810996934772, + 0.0011787646217271686, + 0.0026433425955474377, + 0.0017783803632482886, + 0.0007578599615953863, + 0.001723790424875915, + 0.0010130936279892921, + 0.0020446537528187037, + 0.008495867252349854, + 0.0725998729467392, + 0.055073462426662445, + 0.01138943713158369, + 0.01141435094177723, + 0.007024644874036312, + 0.004840762820094824, + 0.005918990354984999, + 0.002889347728341818, + 0.005782846827059984, + 0.0024219725746661425, + 0.0021347568836063147, + 0.0035264508333057165, + 0.0025551726575940847, + 0.0023166171740740538, + 0.004125069361180067, + 0.003666156902909279, + 0.0031694036442786455, + 0.0031470966059714556, + 0.0027079228311777115, + 0.002523277187719941, + 0.0039536054246127605, + 0.003300788113847375, + 0.0038021684158593416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -1.8048802614212036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016101213172078133 + ], + "opposite_logprob_tokens": [ + -4.141101360321045 + ], + "logprob_reward": -0.016101213172078133, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016101213172078133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3466169834136963, + 0.1893499493598938, + 0.049653925001621246, + 0.02453371323645115, + 0.07916529476642609, + 0.045418478548526764, + 0.04955918341875076, + 0.13364064693450928, + 0.02756388671696186, + 0.05449788272380829 + ], + "credit_map": [ + 0.06750337779521942, + 0.012058435007929802, + 0.015692180022597313, + 0.010336977429687977, + 0.04228846728801727, + 0.14929112792015076, + 0.012299386784434319, + 0.03714705631136894, + 0.030519718304276466, + 0.019215645268559456, + 0.013160322792828083, + 0.012748779729008675, + 0.05933447182178497, + 0.017704522237181664, + 0.011168640106916428, + 0.02549787424504757, + 0.00820186734199524, + 0.007860940881073475, + 0.004317520186305046, + 0.0041260793805122375, + 0.0051034423522651196, + 0.006226899102330208, + 0.004253984894603491, + 0.009563196450471878, + 0.003277901792898774, + 0.003922312520444393, + 0.0029769735410809517, + 0.002145350445061922, + 0.0023380687925964594, + 0.0033379183150827885, + 0.0030037665273994207, + 0.0035314245615154505, + 0.015685169026255608, + 0.010762069374322891, + 0.00990103930234909, + 0.009028255939483643, + 0.005717028398066759, + 0.0066191027872264385, + 0.008511798456311226, + 0.012940836139023304, + 0.00955168902873993, + 0.004663059953600168, + 0.007996750064194202, + 0.003963266499340534, + 0.004234481602907181, + 0.006517739500850439, + 0.0038088231813162565, + 0.0046826740726828575, + 0.0057015069760382175, + 0.00712663633748889, + 0.005464579910039902, + 0.012534376233816147, + 0.0038383223582059145, + 0.003976117353886366, + 0.004377302713692188, + 0.006540348287671804, + 0.006430874112993479, + 0.054818011820316315, + 0.041597362607717514, + 0.008581108413636684, + 0.008659007027745247, + 0.005385325290262699, + 0.0037856511771678925, + 0.004383319988846779, + 0.0032073622569441795, + 0.006249403581023216, + 0.0031381931621581316, + 0.0028222950641065836, + 0.0032172000501304865, + 0.002330900402739644, + 0.0028766156174242496, + 0.0037219191435724497, + 0.006141878664493561, + 0.006385959219187498, + 0.00759953772649169, + 0.0060286130756139755, + 0.005522099789232016, + 0.009501847438514233, + 0.005978927481919527, + 0.007339025381952524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016101213172078133 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372783586382866 + ], + "opposite_logprob_tokens": [ + -3.4087278842926025 + ], + "logprob_reward": -0.03372783586382866, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372783586382866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6100541353225708, + 0.054065048694610596, + 0.07511305063962936, + 0.03704584017395973, + 0.016143647953867912, + 0.017287475988268852, + 0.013677474111318588, + 0.11906366050243378, + 0.01722160167992115, + 0.040328074246644974 + ], + "credit_map": [ + 0.0360993891954422, + 0.025081781670451164, + 0.016898782923817635, + 0.014480119571089745, + 0.24503237009048462, + 0.2196492850780487, + 0.026505522429943085, + 0.026306884363293648, + 0.012986392714083195, + 0.004036158788949251, + 0.005244405008852482, + 0.0032497262582182884, + 0.010896424762904644, + 0.004602279979735613, + 0.004284481052309275, + 0.008765178732573986, + 0.013266423717141151, + 0.011131015606224537, + 0.010764485225081444, + 0.0052683488465845585, + 0.007348050829023123, + 0.009893509559333324, + 0.005458850879222155, + 0.011982360854744911, + 0.005270094610750675, + 0.00564156100153923, + 0.005020332057029009, + 0.004202237352728844, + 0.004114265087991953, + 0.004147298168390989, + 0.0033501044381409883, + 0.0052999453619122505, + 0.0020920413080602884, + 0.0024762859102338552, + 0.0021828000899404287, + 0.0017628624336794019, + 0.0013050619745627046, + 0.0020075063221156597, + 0.0015173343708738685, + 0.002799755660817027, + 0.004141537938266993, + 0.0017085941508412361, + 0.0033113022800534964, + 0.0018553446279838681, + 0.0015880521386861801, + 0.001756005804054439, + 0.0011406830744817853, + 0.0017859548097476363, + 0.0016818555304780602, + 0.0014884545234963298, + 0.002464404795318842, + 0.0033299766946583986, + 0.0009114576387219131, + 0.0010202231351286173, + 0.0010211900807917118, + 0.00175991072319448, + 0.006065355613827705, + 0.04736291617155075, + 0.038850996643304825, + 0.008377360180020332, + 0.0073748016729950905, + 0.0042279609479010105, + 0.002793861087411642, + 0.004010395146906376, + 0.0019726580940186977, + 0.003075085813179612, + 0.0014996256213635206, + 0.0018753790063783526, + 0.0018493527313694358, + 0.002356595592573285, + 0.0019107622792944312, + 0.0026821403298527002, + 0.005853808484971523, + 0.004374812822788954, + 0.005604855250567198, + 0.004693115595728159, + 0.0043197739869356155, + 0.004603879991918802, + 0.005782858468592167, + 0.0050949654541909695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.03372783586382866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640995249152184 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640995249152184, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640995249152184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3234737515449524, + 0.06081155315041542, + 0.04621034115552902, + 0.03496680408716202, + 0.046339813619852066, + 0.041085172444581985, + 0.029941197484731674, + 0.3066307306289673, + 0.04877985268831253, + 0.061760809272527695 + ], + "credit_map": [ + 0.020458679646253586, + 0.015606359578669071, + 0.009348603896796703, + 0.009600144810974598, + 0.1237366646528244, + 0.11466462910175323, + 0.01042005605995655, + 0.019638605415821075, + 0.007584455423057079, + 0.006233100779354572, + 0.004357162397354841, + 0.005414414219558239, + 0.014326298609375954, + 0.006860867142677307, + 0.004180052317678928, + 0.011855198070406914, + 0.0075993710197508335, + 0.007957259193062782, + 0.005915741436183453, + 0.003118326188996434, + 0.004572004079818726, + 0.005212946329265833, + 0.0043818713165819645, + 0.007452822756022215, + 0.004528888501226902, + 0.004013136960566044, + 0.004987732507288456, + 0.003463380504399538, + 0.003090976970270276, + 0.0034266547299921513, + 0.006658582016825676, + 0.004797452129423618, + 0.007658993359655142, + 0.006181536708027124, + 0.006887705996632576, + 0.004917228128761053, + 0.0031295358203351498, + 0.004615528974682093, + 0.004938893020153046, + 0.00801039393991232, + 0.00978599302470684, + 0.004427377134561539, + 0.00643469812348485, + 0.0033594295382499695, + 0.0039146216586232185, + 0.005026417784392834, + 0.0034629853907972574, + 0.004673648625612259, + 0.0026934684719890356, + 0.003709811717271805, + 0.004406476393342018, + 0.006581021938472986, + 0.0027869786135852337, + 0.002267720876261592, + 0.0032932099420577288, + 0.004202510695904493, + 0.016431642696261406, + 0.11774424463510513, + 0.10375485569238663, + 0.02280893735587597, + 0.01835070736706257, + 0.011312484741210938, + 0.007463414222002029, + 0.00876442901790142, + 0.00494421599432826, + 0.011422273702919483, + 0.005904012359678745, + 0.005280888639390469, + 0.006241084076464176, + 0.00389120657928288, + 0.004954516422003508, + 0.006141655147075653, + 0.007947690784931183, + 0.0055655911564826965, + 0.008457507938146591, + 0.0052944207563996315, + 0.006200992967933416, + 0.012653890065848827, + 0.0063589876517653465, + 0.009281722828745842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.008640995249152184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins at the center of a swirling energy vortex and zooms out at explosive speed, revealing the magnificent spectacle of the dawn of the universe: Amidst the boundless blackness of the void, countless shards of energy, composed of fiery red and deep blue, violently scatter in Van Gogh's signature whirlpools. These energies are not smooth gases, but entities with the texture of thick oil paintings, each twisting stream composed of a three-dimensional brushstroke. In the exquisite detail of 8K ultra-high-definition resolution, viewers can clearly see the microscopic ridges and gullies formed by the accumulation of pigments. The fiery red energy burns at the edges, colliding and merging with the deep blue, stirring up violent waves of violet and magenta. The entire image is filled with dynamic tension, as the vast vortex of energy expands outward with unstoppable force, tearing apart the tranquil background and creating an atmosphere of primordial, chaotic, and unparalleled grandeur.", + "seed": 14888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_12/index_9885_seed_14888_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": -0.34695686027407646, + "vlm_logprob": -0.34695686027407646, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.34695686027407646, + "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\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6348072290420532 + ], + "opposite_logprob_tokens": [ + -0.7598072290420532 + ], + "logprob_reward": -0.6348072290420532, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6348072290420532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1983570158481598, + 0.10396991670131683, + 0.08505741506814957, + 0.04485926404595375, + 0.07693632692098618, + 0.08811283856630325, + 0.210276260972023, + 0.11204425990581512, + 0.06283442676067352, + 0.017552295699715614 + ], + "credit_map": [ + 0.026594635099172592, + 0.02310130000114441, + 0.02737891487777233, + 0.025806505233049393, + 0.022105131298303604, + 0.025339407846331596, + 0.02253188192844391, + 0.02549922838807106, + 0.009705429896712303, + 0.011018171906471252, + 0.01078055240213871, + 0.011335227638483047, + 0.01369723491370678, + 0.01180923730134964, + 0.016621548682451248, + 0.019002506509423256, + 0.008576114661991596, + 0.008382024243474007, + 0.01066827867180109, + 0.012614252045750618, + 0.00976570975035429, + 0.009161877445876598, + 0.013965143822133541, + 0.011924011632800102, + 0.004610501695424318, + 0.005636497400701046, + 0.006267197895795107, + 0.005196354817599058, + 0.004523101262748241, + 0.004938561003655195, + 0.006537263281643391, + 0.007149785291403532, + 0.006822484545409679, + 0.008728792890906334, + 0.005999811924993992, + 0.013203544542193413, + 0.008031025528907776, + 0.008622324094176292, + 0.013601342216134071, + 0.011926994659006596, + 0.013581955805420876, + 0.01676221378147602, + 0.010170701891183853, + 0.006607021205127239, + 0.016903823241591454, + 0.008057812228798866, + 0.009043624624609947, + 0.0069856783375144005, + 0.014488287270069122, + 0.02211403101682663, + 0.007587458938360214, + 0.017290551215410233, + 0.01926543191075325, + 0.03718888759613037, + 0.07260743528604507, + 0.01973416842520237, + 0.01129523478448391, + 0.010409777984023094, + 0.007009060587733984, + 0.010149558074772358, + 0.03754950314760208, + 0.008870520628988743, + 0.010503556579351425, + 0.016257043927907944, + 0.007993033155798912, + 0.008934248238801956, + 0.011235655285418034, + 0.009149149060249329, + 0.006216185633093119, + 0.005741824861615896, + 0.006793248001486063, + 0.006771077401936054, + 0.001430097152478993, + 0.0015163116622716188, + 0.0025586478877812624, + 0.003318795934319496, + 0.0017982192803174257, + 0.0022582567762583494, + 0.0018044596072286367, + 0.0028675049543380737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.6348072290420532 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0539000034332275 + ], + "opposite_logprob_tokens": [ + -0.4289000332355499 + ], + "logprob_reward": -1.0539000034332275, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539000034332275, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21122929453849792, + 0.07855744659900665, + 0.07299938052892685, + 0.16920679807662964, + 0.07256463170051575, + 0.11161354184150696, + 0.18731708824634552, + 0.04463304579257965, + 0.029346169903874397, + 0.022532589733600616 + ], + "credit_map": [ + 0.022216686978936195, + 0.025503642857074738, + 0.024900106713175774, + 0.02289300598204136, + 0.02608790062367916, + 0.03199363499879837, + 0.023783516138792038, + 0.03385082632303238, + 0.008804846554994583, + 0.006942322012037039, + 0.011352584697306156, + 0.00831273477524519, + 0.00858449935913086, + 0.009305879473686218, + 0.009745802730321884, + 0.015508782118558884, + 0.006464092992246151, + 0.005626416765153408, + 0.00813219789415598, + 0.008066948503255844, + 0.00799519568681717, + 0.010236654430627823, + 0.01244485005736351, + 0.014033032581210136, + 0.016002964228391647, + 0.025302985683083534, + 0.028234172612428665, + 0.018432138487696648, + 0.022050222381949425, + 0.018375886604189873, + 0.020323559641838074, + 0.02048487588763237, + 0.007485337555408478, + 0.008324680849909782, + 0.005203756969422102, + 0.013713178224861622, + 0.007484021130949259, + 0.007783933077007532, + 0.013289128430187702, + 0.009280601516366005, + 0.022136492654681206, + 0.016565578058362007, + 0.01193470973521471, + 0.01244295947253704, + 0.012902799062430859, + 0.011972600594162941, + 0.010845836251974106, + 0.01281257439404726, + 0.015512586571276188, + 0.018213804811239243, + 0.013902639038860798, + 0.01417063269764185, + 0.017190782353281975, + 0.03160922974348068, + 0.05846318230032921, + 0.018254240974783897, + 0.004585455637425184, + 0.0056170872412621975, + 0.005663381423801184, + 0.004739576485008001, + 0.006075119134038687, + 0.008993664756417274, + 0.004361182451248169, + 0.004597583319991827, + 0.0031720770057290792, + 0.004052565898746252, + 0.0047460030764341354, + 0.003914177417755127, + 0.002925919834524393, + 0.0029923939146101475, + 0.0035821532364934683, + 0.003960882313549519, + 0.0020721331238746643, + 0.002885436872020364, + 0.0028855116106569767, + 0.0032273854594677687, + 0.003577921539545059, + 0.0023189298808574677, + 0.002851130673661828, + 0.0027141415048390627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -1.0539000034332275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645722642540932 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645722642540932, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645722642540932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08598274737596512, + 0.09817047417163849, + 0.0611446350812912, + 0.08149231970310211, + 0.04708511382341385, + 0.09196656197309494, + 0.37174054980278015, + 0.07003426551818848, + 0.05780033394694328, + 0.03458303213119507 + ], + "credit_map": [ + 0.009352613240480423, + 0.008539460599422455, + 0.007590415887534618, + 0.008597306907176971, + 0.010143715888261795, + 0.016821227967739105, + 0.007518856320530176, + 0.017419157549738884, + 0.007060157135128975, + 0.006693081930279732, + 0.009532813914120197, + 0.006889938842505217, + 0.009767500683665276, + 0.017873652279376984, + 0.010671611875295639, + 0.02968171238899231, + 0.004048313945531845, + 0.0036001058761030436, + 0.00683171721175313, + 0.005183682311326265, + 0.005918256472796202, + 0.014900154434144497, + 0.00523030012845993, + 0.015432102605700493, + 0.009606786072254181, + 0.00994827039539814, + 0.013224069029092789, + 0.007812192663550377, + 0.0080157071352005, + 0.009850508533418179, + 0.010898226872086525, + 0.01213656086474657, + 0.005597269162535667, + 0.004332246724516153, + 0.005286391358822584, + 0.003947124816477299, + 0.005298227537423372, + 0.007083228789269924, + 0.00556009728461504, + 0.009980527684092522, + 0.011714420281350613, + 0.009701227769255638, + 0.005651616025716066, + 0.007496776059269905, + 0.008903305977582932, + 0.02150603011250496, + 0.006946919020265341, + 0.02004626952111721, + 0.026728235185146332, + 0.03632746636867523, + 0.020001022145152092, + 0.025838488712906837, + 0.030642302706837654, + 0.06922677159309387, + 0.12046744674444199, + 0.0425087995827198, + 0.00829812791198492, + 0.0051857903599739075, + 0.016582051292061806, + 0.00599467009305954, + 0.005868620704859495, + 0.011997964233160019, + 0.00511515699326992, + 0.010991882532835007, + 0.00767551688477397, + 0.0049683270044624805, + 0.012023478746414185, + 0.00689246179535985, + 0.0049632699228823185, + 0.0073031592182815075, + 0.005610120948404074, + 0.008364001289010048, + 0.003429266158491373, + 0.0035742383915930986, + 0.004850120283663273, + 0.004742851480841637, + 0.004282847046852112, + 0.005337639711797237, + 0.0035393957514315844, + 0.004826672375202179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.008645722642540932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806537553668022 + ], + "opposite_logprob_tokens": [ + -3.529806613922119 + ], + "logprob_reward": -0.029806537553668022, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806537553668022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11989320814609528, + 0.06958640366792679, + 0.0726759135723114, + 0.07364904135465622, + 0.054805874824523926, + 0.06002466380596161, + 0.14253033697605133, + 0.2018740326166153, + 0.17870758473873138, + 0.02625291422009468 + ], + "credit_map": [ + 0.014618362300097942, + 0.013234838843345642, + 0.0181575957685709, + 0.01654716394841671, + 0.013985083438456059, + 0.013126008212566376, + 0.014081002213060856, + 0.016143159940838814, + 0.004491525236517191, + 0.006530229467898607, + 0.005817491561174393, + 0.011529341340065002, + 0.006432484835386276, + 0.008215725421905518, + 0.015033048577606678, + 0.011536558158695698, + 0.0048243096098303795, + 0.0064697046764194965, + 0.007984683848917484, + 0.009689505212008953, + 0.005567007232457399, + 0.011185860261321068, + 0.008812629617750645, + 0.018142210319638252, + 0.007716563064604998, + 0.008285874500870705, + 0.01342066377401352, + 0.008083743043243885, + 0.008045276626944542, + 0.008561923168599606, + 0.010579494759440422, + 0.008955503813922405, + 0.007168312091380358, + 0.00535194855183363, + 0.005562858190387487, + 0.006317770574241877, + 0.006440030410885811, + 0.008721815422177315, + 0.00880917627364397, + 0.006433960981667042, + 0.00829478818923235, + 0.007362573407590389, + 0.004841112997382879, + 0.004803403280675411, + 0.005886665545403957, + 0.012677103281021118, + 0.005160903558135033, + 0.010998113080859184, + 0.009472215548157692, + 0.014535090886056423, + 0.007116834633052349, + 0.010896947234869003, + 0.012744160369038582, + 0.02497122809290886, + 0.0461680106818676, + 0.016625836491584778, + 0.023585814982652664, + 0.01712457649409771, + 0.019858364015817642, + 0.015051131136715412, + 0.06505418568849564, + 0.023744141682982445, + 0.016322752460837364, + 0.021133068948984146, + 0.023370858281850815, + 0.01543413009494543, + 0.03210343420505524, + 0.018186504021286964, + 0.017471127212047577, + 0.021622946485877037, + 0.0205946397036314, + 0.029923953115940094, + 0.003910065162926912, + 0.0022852057591080666, + 0.002953938441351056, + 0.002524895826354623, + 0.006136464886367321, + 0.004098111763596535, + 0.0019939427729696035, + 0.0023502898402512074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.029806537553668022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot centered on a swirling energy vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera zoom out at explosive speed to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the energy shards composed of fiery red and deep blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy streams have the texture of thick oil paintings with visible three-dimensional brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue energies collide to create waves of violet and magenta?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624808698892593 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624808698892593, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624808698892593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13432259857654572, + 0.1021508276462555, + 0.06226638704538345, + 0.15287360548973083, + 0.07550611346960068, + 0.08577951788902283, + 0.17938131093978882, + 0.09483592957258224, + 0.05178234726190567, + 0.06110137701034546 + ], + "credit_map": [ + 0.015523010864853859, + 0.011302923783659935, + 0.011535858735442162, + 0.012889621779322624, + 0.014278245158493519, + 0.019168101251125336, + 0.0134571623057127, + 0.03616767376661301, + 0.007562501821666956, + 0.006849059835076332, + 0.011524679139256477, + 0.010138588026165962, + 0.010685480199754238, + 0.011480879038572311, + 0.01343865692615509, + 0.03047098033130169, + 0.005691969767212868, + 0.008235471323132515, + 0.009154967963695526, + 0.00799147691577673, + 0.006436915602535009, + 0.011036688461899757, + 0.006974637974053621, + 0.0067442539148032665, + 0.016246674582362175, + 0.022524883970618248, + 0.026211794465780258, + 0.018456390127539635, + 0.01849159598350525, + 0.011979127302765846, + 0.021811669692397118, + 0.01715146377682686, + 0.009282710961997509, + 0.005428193602710962, + 0.005991413723677397, + 0.00605882378295064, + 0.007005117367953062, + 0.018076244741678238, + 0.007461609784513712, + 0.016202006489038467, + 0.0093288104981184, + 0.008680804632604122, + 0.006976054981350899, + 0.005309498403221369, + 0.008771713823080063, + 0.021350393071770668, + 0.0059781670570373535, + 0.019384076818823814, + 0.016214264556765556, + 0.02023855783045292, + 0.01928788237273693, + 0.027530526742339134, + 0.023469477891921997, + 0.015422305092215538, + 0.03775852173566818, + 0.01945977471768856, + 0.012465663254261017, + 0.009894388727843761, + 0.015333877876400948, + 0.011531831696629524, + 0.012719874270260334, + 0.012573410756886005, + 0.010387061163783073, + 0.00992982741445303, + 0.007369745057076216, + 0.006609281990677118, + 0.00793405156582594, + 0.006760779768228531, + 0.005445836577564478, + 0.005596958100795746, + 0.0061385296285152435, + 0.005927166901528835, + 0.008624416776001453, + 0.005131610203534365, + 0.006839854642748833, + 0.009799734689295292, + 0.007058135233819485, + 0.009167322888970375, + 0.0045182728208601475, + 0.009962028823792934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.007624808698892593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_tensor([9980, 9980])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_tensor([9980, 9980])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d14b2b8756009229ba69183cd9f339597dc34405 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_tensor([9980, 9980])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium profile shot, positioned at a slightly low angle to convey a sense of gentle majesty, gracefully moves sideways, perfectly synchronizing with the central subject: a lanky, gentle giant composed entirely of living wood and foliage. His form is a meticulous and artistic masterpiece of Art Nouveau; his torso and long limbs are crafted from smooth, polished wood that bends and flows in elegant, sinuous lines, with organic patterns artfully carved into the grain. Dense green foliage covers his shoulders and waist, each leaf delicately swirling with veins. He sways with a soothing, hypnotic, and complex pendulum; his wooden torso initiates a slow, rhythmic tilt, while his slender limbs follow in graceful, delayed arcs, creating a mesmerizing, continuous wave of movement that flows through his entire body. His background is a stylized natural backdrop, a carpet of intricate fern patterns and oversized bell-shaped flowers with pearlescent petals. The trees in the background twist their branches into elaborate, organic arabesques, silhouetted against a sky painted in soft, warm colors. The entire scene is bathed in the warm, diffuse glow of the afternoon sun, casting long, soft shadows that accentuate the rich, earthy browns of the wood and the myriad green hues of the foliage, ultimately creating an atmosphere of deep tranquility and enchantment that captures the essence of artistic, fluid naturalism.", + "seed": 29620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_9980_seed_29620_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": -0.04267497723922133, + "vlm_logprob": -0.04267497723922133, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04267497723922133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04168069362640381 + ], + "opposite_logprob_tokens": [ + -3.2916808128356934 + ], + "logprob_reward": -0.04168069362640381, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04168069362640381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2660441994667053, + 0.1398930847644806, + 0.09623242914676666, + 0.10413094609975815, + 0.049714911729097366, + 0.08024685084819794, + 0.15858840942382812, + 0.028011789545416832, + 0.049816664308309555, + 0.027320707216858864 + ], + "credit_map": [ + 0.0244569294154644, + 0.035145681351423264, + 0.022509383037686348, + 0.02394372969865799, + 0.04693200811743736, + 0.018188659101724625, + 0.030557142570614815, + 0.06431067734956741, + 0.015262524597346783, + 0.014697589911520481, + 0.028950167819857597, + 0.019267309457063675, + 0.02190338633954525, + 0.010839142836630344, + 0.016283567994832993, + 0.012689403258264065, + 0.011403897777199745, + 0.011968951672315598, + 0.012245637364685535, + 0.01054962445050478, + 0.011192074976861477, + 0.01652788557112217, + 0.010701548308134079, + 0.01164280902594328, + 0.008684933185577393, + 0.015174489468336105, + 0.00907544232904911, + 0.01323793362826109, + 0.021912410855293274, + 0.009375992231070995, + 0.013886446133255959, + 0.012783295474946499, + 0.005247066728770733, + 0.005337896756827831, + 0.007640573661774397, + 0.005474174860864878, + 0.006370364688336849, + 0.006185004487633705, + 0.007070472929626703, + 0.006389360409229994, + 0.008790813386440277, + 0.00787491537630558, + 0.012713496573269367, + 0.00802266038954258, + 0.008754918351769447, + 0.011007729917764664, + 0.00904893409460783, + 0.014033384621143341, + 0.01408664882183075, + 0.02514856494963169, + 0.014774387702345848, + 0.02039146237075329, + 0.020969998091459274, + 0.016945507377386093, + 0.022274065762758255, + 0.023997781798243523, + 0.004066400695592165, + 0.002807863289490342, + 0.002420503180474043, + 0.004422928672283888, + 0.004558582790195942, + 0.0034999928902834654, + 0.0035840366035699844, + 0.0026514832861721516, + 0.005214097443968058, + 0.0076861027628183365, + 0.0048910221084952354, + 0.007250037044286728, + 0.005601902958005667, + 0.006005005910992622, + 0.007523011416196823, + 0.005645484197884798, + 0.0015156699810177088, + 0.00558354239910841, + 0.0031559530179947615, + 0.004434363450855017, + 0.0029768587555736303, + 0.002841960871592164, + 0.0038281166926026344, + 0.0029842399526387453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.04168069362640381 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371100872755051 + ], + "opposite_logprob_tokens": [ + -3.4087109565734863 + ], + "logprob_reward": -0.03371100872755051, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371100872755051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35706913471221924, + 0.12080994993448257, + 0.041126273572444916, + 0.06636188179254532, + 0.035088781267404556, + 0.11672352999448776, + 0.1308162361383438, + 0.048393625766038895, + 0.041517943143844604, + 0.042092569172382355 + ], + "credit_map": [ + 0.020505055785179138, + 0.029772678390145302, + 0.026853328570723534, + 0.014959233812987804, + 0.03263465687632561, + 0.02157682552933693, + 0.06923295557498932, + 0.14153441786766052, + 0.01201755739748478, + 0.016412068158388138, + 0.01924039050936699, + 0.017752621322870255, + 0.01850048638880253, + 0.010739597491919994, + 0.012881036847829819, + 0.013266189023852348, + 0.004003157373517752, + 0.004693725146353245, + 0.005275576841086149, + 0.0059571717865765095, + 0.004848992917686701, + 0.005131837911903858, + 0.006249230355024338, + 0.004966581240296364, + 0.006488560698926449, + 0.006027249153703451, + 0.0072881379164755344, + 0.006869371980428696, + 0.007824152708053589, + 0.01147295068949461, + 0.008338061161339283, + 0.012053396552801132, + 0.0032044921535998583, + 0.004195645917207003, + 0.004666302818804979, + 0.0036740493960678577, + 0.004894126206636429, + 0.0046238163486123085, + 0.005063116550445557, + 0.004767232574522495, + 0.01357754785567522, + 0.010131514631211758, + 0.016824766993522644, + 0.011974697932600975, + 0.011641615070402622, + 0.016828017309308052, + 0.018959952518343925, + 0.016785424202680588, + 0.010403201915323734, + 0.019064361229538918, + 0.01265998650342226, + 0.01718158647418022, + 0.01682470552623272, + 0.013553517870604992, + 0.018254537135362625, + 0.02287433296442032, + 0.005867437459528446, + 0.006842007394880056, + 0.007386433891952038, + 0.006376524921506643, + 0.0058509246446192265, + 0.005666072946041822, + 0.006321947555989027, + 0.004082278348505497, + 0.004526233766227961, + 0.006926806177943945, + 0.003043619217351079, + 0.0060601793229579926, + 0.004477516748011112, + 0.0037716091610491276, + 0.008051122538745403, + 0.004660855513066053, + 0.0028214005287736654, + 0.011768324300646782, + 0.004184459336102009, + 0.004237418994307518, + 0.00300150946713984, + 0.00591374933719635, + 0.00469565112143755, + 0.005470056552439928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.03371100872755051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000669836997986 + ], + "opposite_logprob_tokens": [ + -2.6950066089630127 + ], + "logprob_reward": -0.07000669836997986, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000669836997986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30810344219207764, + 0.06022505834698677, + 0.03535902500152588, + 0.1241789311170578, + 0.04142320901155472, + 0.05946929007768631, + 0.24885182082653046, + 0.041693370789289474, + 0.03384245187044144, + 0.046853452920913696 + ], + "credit_map": [ + 0.02528342977166176, + 0.030825603753328323, + 0.0288145300000906, + 0.020015759393572807, + 0.02382650040090084, + 0.015461587347090244, + 0.04356938973069191, + 0.12030661851167679, + 0.005886109545826912, + 0.0069549959152936935, + 0.009288660250604153, + 0.007687581703066826, + 0.011347378604114056, + 0.006934545002877712, + 0.006415729410946369, + 0.0057100532576441765, + 0.002562203211709857, + 0.00440513389185071, + 0.004555061459541321, + 0.004457891918718815, + 0.003325273049995303, + 0.005839171353727579, + 0.006445887964218855, + 0.0037683991249650717, + 0.010847601108253002, + 0.018841266632080078, + 0.010522741824388504, + 0.019392021000385284, + 0.025096144527196884, + 0.013627779670059681, + 0.014437959529459476, + 0.01141341496258974, + 0.0027723698876798153, + 0.006763526704162359, + 0.004555858671665192, + 0.005986267700791359, + 0.003907610196620226, + 0.004037404898554087, + 0.007624372839927673, + 0.005775793921202421, + 0.006769898347556591, + 0.0060623278841376305, + 0.009632977657020092, + 0.00675300695002079, + 0.006488649174571037, + 0.008510512299835682, + 0.006279565393924713, + 0.008972350507974625, + 0.023901578038930893, + 0.03802188113331795, + 0.022847607731819153, + 0.03151152282953262, + 0.03262518346309662, + 0.02753637172281742, + 0.03493206575512886, + 0.0374755933880806, + 0.005136444699019194, + 0.005032903514802456, + 0.005921253934502602, + 0.006960217375308275, + 0.005533464252948761, + 0.0040243323892354965, + 0.005269215442240238, + 0.0038155382499098778, + 0.0036060987040400505, + 0.005090148653835058, + 0.0025276625528931618, + 0.004066717345267534, + 0.004607497248798609, + 0.0032873041927814484, + 0.00617802981287241, + 0.004478991497308016, + 0.003257234115153551, + 0.010468798689544201, + 0.004155430477112532, + 0.0064902831800282, + 0.00522783026099205, + 0.004571458324790001, + 0.007010908797383308, + 0.005671503953635693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.07000669836997986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948815029114485 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005948815029114485, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948815029114485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3440355658531189, + 0.0555572584271431, + 0.04979574680328369, + 0.08502724021673203, + 0.0179500300437212, + 0.0629301443696022, + 0.26785510778427124, + 0.04468277841806412, + 0.04271732643246651, + 0.029448797926306725 + ], + "credit_map": [ + 0.019851205870509148, + 0.036242179572582245, + 0.03824065625667572, + 0.017819013446569443, + 0.024289879947900772, + 0.019566461443901062, + 0.05825993791222572, + 0.12976624071598053, + 0.006879563443362713, + 0.007596857845783234, + 0.00942689273506403, + 0.0081763481721282, + 0.006928411778062582, + 0.003460405860096216, + 0.00650473777204752, + 0.00658403430134058, + 0.004546249285340309, + 0.0059974160976707935, + 0.005762849468737841, + 0.007452018558979034, + 0.006870516110211611, + 0.007782157976180315, + 0.005657187197357416, + 0.005727351177483797, + 0.009683044627308846, + 0.007410090416669846, + 0.00738880829885602, + 0.011394485831260681, + 0.012214741669595242, + 0.011668914929032326, + 0.0158299058675766, + 0.009437243454158306, + 0.002453609835356474, + 0.0025147751439362764, + 0.002040833467617631, + 0.002207681071013212, + 0.0018399417167529464, + 0.0019346307963132858, + 0.0026439689099788666, + 0.002314587589353323, + 0.008318099193274975, + 0.007493740878999233, + 0.010102147236466408, + 0.00616861879825592, + 0.006225860211998224, + 0.0075044576078653336, + 0.008034543134272099, + 0.009082671254873276, + 0.025330351665616035, + 0.04263294115662575, + 0.02580731362104416, + 0.03313799947500229, + 0.0375797338783741, + 0.027245525270700455, + 0.03589542210102081, + 0.0402258038520813, + 0.00727933319285512, + 0.006475343368947506, + 0.006159244105219841, + 0.004467903170734644, + 0.005404498428106308, + 0.005712143145501614, + 0.005734099540859461, + 0.003450213000178337, + 0.003180780680850148, + 0.007848050445318222, + 0.005446218885481358, + 0.007582755293697119, + 0.005431316792964935, + 0.004446989856660366, + 0.004462266340851784, + 0.004318946041166782, + 0.002195558277890086, + 0.003020132891833782, + 0.0036599403247237206, + 0.004794525448232889, + 0.0034435619600117207, + 0.003574565751478076, + 0.005420692730695009, + 0.0033398165833204985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.005948815029114485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062027670443058014 + ], + "opposite_logprob_tokens": [ + -2.8120276927948 + ], + "logprob_reward": -0.062027670443058014, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062027670443058014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41816020011901855, + 0.062448546290397644, + 0.07604244351387024, + 0.036019694060087204, + 0.019044378772377968, + 0.03734789043664932, + 0.17374591529369354, + 0.07798685133457184, + 0.03517851606011391, + 0.06402558833360672 + ], + "credit_map": [ + 0.027679793536663055, + 0.04213523864746094, + 0.044983357191085815, + 0.027196984738111496, + 0.026727942749857903, + 0.030584117397665977, + 0.07378589361906052, + 0.14506687223911285, + 0.005139934830367565, + 0.00705235218629241, + 0.018205411732196808, + 0.006032223347574472, + 0.009428071789443493, + 0.003963748458772898, + 0.005507112946361303, + 0.00711968494579196, + 0.008059928193688393, + 0.008336995728313923, + 0.010021653026342392, + 0.006515516433864832, + 0.006871817167848349, + 0.010463547892868519, + 0.017752641811966896, + 0.008020336739718914, + 0.0033364074770361185, + 0.003928279969841242, + 0.0037160497158765793, + 0.004801352508366108, + 0.005416585132479668, + 0.004095560871064663, + 0.004739975556731224, + 0.005985480733215809, + 0.002866291906684637, + 0.0024421405978500843, + 0.0023596263490617275, + 0.002295574638992548, + 0.00174721737857908, + 0.002013122895732522, + 0.00297652930021286, + 0.0023438746575266123, + 0.003286329796537757, + 0.003173000644892454, + 0.005269190762192011, + 0.003900497918948531, + 0.0034539205953478813, + 0.003824332496151328, + 0.009076439775526524, + 0.005364175420254469, + 0.01569071039557457, + 0.02824154682457447, + 0.01770370826125145, + 0.02122858539223671, + 0.025745365768671036, + 0.0163893923163414, + 0.02250499464571476, + 0.026241600513458252, + 0.008739209733903408, + 0.008789016865193844, + 0.010023441165685654, + 0.014248199760913849, + 0.010044623166322708, + 0.009037956595420837, + 0.009467055089771748, + 0.007637341506779194, + 0.004242526367306709, + 0.0037493568379431963, + 0.002757730195298791, + 0.004830128978937864, + 0.004819834139198065, + 0.004770911298692226, + 0.00610370934009552, + 0.003904316108673811, + 0.0048642209731042385, + 0.008986189030110836, + 0.007736236322671175, + 0.00806412659585476, + 0.007864183746278286, + 0.007086111232638359, + 0.010956499725580215, + 0.008468014188110828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.062027670443058014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium profile shot, positioned at a slightly low angle to convey a sense of gentle majesty, gracefully moves sideways, perfectly synchronizing with the central subject: a lanky, gentle giant composed entirely of living wood and foliage. His form is a meticulous and artistic masterpiece of Art Nouveau; his torso and long limbs are crafted from smooth, polished wood that bends and flows in elegant, sinuous lines, with organic patterns artfully carved into the grain. Dense green foliage covers his shoulders and waist, each leaf delicately swirling with veins. He sways with a soothing, hypnotic, and complex pendulum; his wooden torso initiates a slow, rhythmic tilt, while his slender limbs follow in graceful, delayed arcs, creating a mesmerizing, continuous wave of movement that flows through his entire body. His background is a stylized natural backdrop, a carpet of intricate fern patterns and oversized bell-shaped flowers with pearlescent petals. The trees in the background twist their branches into elaborate, organic arabesques, silhouetted against a sky painted in soft, warm colors. The entire scene is bathed in the warm, diffuse glow of the afternoon sun, casting long, soft shadows that accentuate the rich, earthy browns of the wood and the myriad green hues of the foliage, ultimately creating an atmosphere of deep tranquility and enchantment that captures the essence of artistic, fluid naturalism.", + "seed": 29621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_9980_seed_29621_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": -0.018520624842494726, + "vlm_logprob": -0.018520624842494726, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.018520624842494726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01616760715842247 + ], + "opposite_logprob_tokens": [ + -4.391167640686035 + ], + "logprob_reward": -0.01616760715842247, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01616760715842247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3189946413040161, + 0.050436437129974365, + 0.08291970193386078, + 0.10383258759975433, + 0.040400974452495575, + 0.045847076922655106, + 0.07301612198352814, + 0.07482634484767914, + 0.14945298433303833, + 0.06027321144938469 + ], + "credit_map": [ + 0.03891109302639961, + 0.04923991858959198, + 0.040002837777137756, + 0.04454239830374718, + 0.03118087723851204, + 0.05082791671156883, + 0.02508716844022274, + 0.03920239210128784, + 0.005090172402560711, + 0.008912318386137486, + 0.00638557830825448, + 0.007577439304441214, + 0.0058229961432516575, + 0.006255645304918289, + 0.0046389782801270485, + 0.005753306206315756, + 0.0069750952534377575, + 0.010198481380939484, + 0.01088523119688034, + 0.01175492350012064, + 0.009834092110395432, + 0.010421274229884148, + 0.010558097623288631, + 0.012292500585317612, + 0.009574370458722115, + 0.01414088811725378, + 0.011105512268841267, + 0.02355043962597847, + 0.008377601392567158, + 0.013035589829087257, + 0.013535388745367527, + 0.010512782260775566, + 0.0034538700710982084, + 0.004968858323991299, + 0.00454011932015419, + 0.004610393662005663, + 0.004933592863380909, + 0.005377093330025673, + 0.007371985819190741, + 0.005145058501511812, + 0.00477123586460948, + 0.0055332304909825325, + 0.004698534961789846, + 0.00447995075955987, + 0.00958652887493372, + 0.005554626230150461, + 0.0063360342755913734, + 0.004886933136731386, + 0.007101408671587706, + 0.008925557136535645, + 0.009776249527931213, + 0.011819392442703247, + 0.008098608814179897, + 0.008237408474087715, + 0.010085297748446465, + 0.008972194977104664, + 0.004689091816544533, + 0.012023351155221462, + 0.005875562317669392, + 0.007318346295505762, + 0.012306216172873974, + 0.011823197826743126, + 0.008282925002276897, + 0.012507646344602108, + 0.011006615124642849, + 0.012705783359706402, + 0.019311366602778435, + 0.019777385517954826, + 0.01661825180053711, + 0.022674232721328735, + 0.024742720648646355, + 0.022616611793637276, + 0.006301244720816612, + 0.007042312063276768, + 0.012652907520532608, + 0.008988991379737854, + 0.005786383058875799, + 0.006570889614522457, + 0.007086492143571377, + 0.00584398815408349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.01616760715842247 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207405038177967 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207405038177967, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207405038177967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20600149035453796, + 0.06796839833259583, + 0.12747208774089813, + 0.1090497076511383, + 0.08392377942800522, + 0.04279445484280586, + 0.09231498837471008, + 0.12365124374628067, + 0.10799705982208252, + 0.0388268306851387 + ], + "credit_map": [ + 0.03144000843167305, + 0.041785869747400284, + 0.02698175609111786, + 0.02716301940381527, + 0.01780945435166359, + 0.019743669778108597, + 0.019275829195976257, + 0.02180187776684761, + 0.006711283233016729, + 0.009793084114789963, + 0.011721223592758179, + 0.008535685017704964, + 0.0073348586447536945, + 0.012916863895952702, + 0.0052445512264966965, + 0.005710839293897152, + 0.009433078579604626, + 0.010901347734034061, + 0.02196437120437622, + 0.014456324279308319, + 0.014071962796151638, + 0.01807285100221634, + 0.024088067933917046, + 0.014484086073935032, + 0.009471576660871506, + 0.013512366451323032, + 0.011419340036809444, + 0.022008150815963745, + 0.009978128597140312, + 0.015710260719060898, + 0.015094601549208164, + 0.01185526978224516, + 0.009445948526263237, + 0.00657634949311614, + 0.008266943506896496, + 0.007158213295042515, + 0.02050797827541828, + 0.013065292499959469, + 0.011640272103250027, + 0.007262783125042915, + 0.0033197824377566576, + 0.0057252696715295315, + 0.006086455192416906, + 0.0061468095518648624, + 0.007953585125505924, + 0.0035874280147254467, + 0.0040857731364667416, + 0.005889351014047861, + 0.009463552385568619, + 0.01307228859513998, + 0.013256221078336239, + 0.01406781654804945, + 0.012087997980415821, + 0.007845068350434303, + 0.013046573847532272, + 0.009475459344685078, + 0.008260155096650124, + 0.01851649582386017, + 0.010322448797523975, + 0.011618101969361305, + 0.021504152566194534, + 0.016099384054541588, + 0.015175377950072289, + 0.02215511165559292, + 0.00917753204703331, + 0.011131389066576958, + 0.010462024249136448, + 0.013725037686526775, + 0.014749085530638695, + 0.01889808289706707, + 0.01619708351790905, + 0.01365682203322649, + 0.004113351460546255, + 0.005556956864893436, + 0.006034538149833679, + 0.006450374610722065, + 0.0035694455727934837, + 0.0036464110016822815, + 0.005915937013924122, + 0.003539812285453081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.014207405038177967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343757286667824 + ], + "opposite_logprob_tokens": [ + -3.651343822479248 + ], + "logprob_reward": -0.026343757286667824, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343757286667824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16352631151676178, + 0.1290365308523178, + 0.07471320778131485, + 0.11001135408878326, + 0.07241904735565186, + 0.2020503133535385, + 0.038923412561416626, + 0.04459657147526741, + 0.07329656928777695, + 0.09142667800188065 + ], + "credit_map": [ + 0.017680106684565544, + 0.023463107645511627, + 0.01912529021501541, + 0.021158693358302116, + 0.014527041465044022, + 0.02825993299484253, + 0.012323074042797089, + 0.02698906697332859, + 0.011411500163376331, + 0.021020252257585526, + 0.01776627078652382, + 0.025090696290135384, + 0.013949374668300152, + 0.011342528276145458, + 0.015834476798772812, + 0.012621436268091202, + 0.005637533031404018, + 0.00726791936904192, + 0.01287595834583044, + 0.0057561662979424, + 0.007997432723641396, + 0.0120399110019207, + 0.014542702585458755, + 0.008595584891736507, + 0.009390865452587605, + 0.013599681667983532, + 0.010832758620381355, + 0.02245165966451168, + 0.01810217648744583, + 0.014200696721673012, + 0.011322597973048687, + 0.010110916569828987, + 0.0071302903816103935, + 0.008223189041018486, + 0.00828735064715147, + 0.007119651418179274, + 0.012397609651088715, + 0.011673948727548122, + 0.00962128397077322, + 0.007965723983943462, + 0.01875346712768078, + 0.02544926106929779, + 0.017683038488030434, + 0.023103486746549606, + 0.04844815284013748, + 0.01767582632601261, + 0.01959884911775589, + 0.03133822977542877, + 0.0028627084102481604, + 0.005843020975589752, + 0.003320018295198679, + 0.0053818970918655396, + 0.0076727718114852905, + 0.004007099661976099, + 0.004790824372321367, + 0.005045069847255945, + 0.004105100873857737, + 0.0048059639520943165, + 0.0044968402944505215, + 0.004317394457757473, + 0.007148413453251123, + 0.006879821419715881, + 0.005587413441389799, + 0.007255624979734421, + 0.006017542444169521, + 0.007898128591477871, + 0.008789746090769768, + 0.006504375487565994, + 0.010559390299022198, + 0.009826069697737694, + 0.012642885558307171, + 0.011058429256081581, + 0.008451771922409534, + 0.01497526466846466, + 0.014748090878129005, + 0.012461083941161633, + 0.010850222781300545, + 0.006954011041671038, + 0.013753059320151806, + 0.009233178570866585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.026343757286667824 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021962826140224934 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021962826140224934, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021962826140224934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1430904120206833, + 0.06261061877012253, + 0.06089054420590401, + 0.06859710067510605, + 0.1955854296684265, + 0.14260895550251007, + 0.07008875161409378, + 0.1592065840959549, + 0.046064265072345734, + 0.05125734955072403 + ], + "credit_map": [ + 0.019423259422183037, + 0.017971670255064964, + 0.018147064372897148, + 0.01873701810836792, + 0.014986447989940643, + 0.01854289136826992, + 0.013266491703689098, + 0.022015556693077087, + 0.007243078667670488, + 0.008168956264853477, + 0.009355859830975533, + 0.009112922474741936, + 0.008122525177896023, + 0.006801888812333345, + 0.0058517782017588615, + 0.007953610271215439, + 0.004721291363239288, + 0.007386299315840006, + 0.008282323367893696, + 0.007884818129241467, + 0.006234071217477322, + 0.0072968569584190845, + 0.010012885555624962, + 0.009071998298168182, + 0.004084879998117685, + 0.004584353882819414, + 0.00670017721131444, + 0.01429113931953907, + 0.01041983999311924, + 0.012103014625608921, + 0.009189590811729431, + 0.007224103901535273, + 0.015310140326619148, + 0.014206918887794018, + 0.019623495638370514, + 0.019974103197455406, + 0.02094479650259018, + 0.050028834491968155, + 0.03205074369907379, + 0.023446401581168175, + 0.01394119393080473, + 0.016639530658721924, + 0.01323546003550291, + 0.016030820086598396, + 0.025646395981311798, + 0.015522816218435764, + 0.017403867095708847, + 0.024188870564103127, + 0.006886082235723734, + 0.00625262875109911, + 0.009575266391038895, + 0.009553714655339718, + 0.008127788081765175, + 0.007814401760697365, + 0.011843330226838589, + 0.010035540908575058, + 0.019115092232823372, + 0.02183416113257408, + 0.014396091923117638, + 0.022048236802220345, + 0.024114130064845085, + 0.015291446819901466, + 0.0222201906144619, + 0.020187221467494965, + 0.0031996886245906353, + 0.005300976801663637, + 0.0056021371856331825, + 0.004637883044779301, + 0.006405725609511137, + 0.004580239299684763, + 0.010823856107890606, + 0.005513758864253759, + 0.005864177830517292, + 0.00523168733343482, + 0.007755896542221308, + 0.006911762990057468, + 0.008345022797584534, + 0.005817080847918987, + 0.00717334495857358, + 0.004158374387770891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.0021962826140224934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368807211518288 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.03368807211518288, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368807211518288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.539272129535675, + 0.05613945797085762, + 0.04024872928857803, + 0.059488445520401, + 0.0971788838505745, + 0.06608427315950394, + 0.0436374694108963, + 0.04678560048341751, + 0.028120091184973717, + 0.023044928908348083 + ], + "credit_map": [ + 0.053282786160707474, + 0.07172409445047379, + 0.06323591619729996, + 0.07517164945602417, + 0.058329563587903976, + 0.10330086201429367, + 0.03997353836894035, + 0.07425368577241898, + 0.006327104289084673, + 0.008019364438951015, + 0.008614526130259037, + 0.009850699454545975, + 0.004237543325871229, + 0.004824006464332342, + 0.007570123765617609, + 0.006696087773889303, + 0.0024123305920511484, + 0.0047813779674470425, + 0.010580691508948803, + 0.003842826234176755, + 0.002450717380270362, + 0.00891447626054287, + 0.003966411110013723, + 0.003299898002296686, + 0.0042719049379229546, + 0.0061422004364430904, + 0.005734726786613464, + 0.01020834781229496, + 0.00815342366695404, + 0.010061502456665039, + 0.00870926771312952, + 0.0062070670537650585, + 0.00798511691391468, + 0.006875256076455116, + 0.008450096473097801, + 0.010471104644238949, + 0.01254159677773714, + 0.025037216022610664, + 0.014793614856898785, + 0.01102488487958908, + 0.00584828108549118, + 0.006128900684416294, + 0.005935704801231623, + 0.005371721927076578, + 0.008148987777531147, + 0.008333321660757065, + 0.010079428553581238, + 0.016237927600741386, + 0.0037846900522708893, + 0.004429860506206751, + 0.005026019178330898, + 0.0064758057706058025, + 0.0056360382586717606, + 0.0062500969506800175, + 0.005088018719106913, + 0.006946939043700695, + 0.005847325548529625, + 0.006595900747925043, + 0.003946302924305201, + 0.006191805470734835, + 0.004884213209152222, + 0.005191308446228504, + 0.0067217652685940266, + 0.007406980264931917, + 0.002411003690212965, + 0.00336497207172215, + 0.003997890744358301, + 0.0035870030988007784, + 0.003927826415747404, + 0.0027487576007843018, + 0.004815550521016121, + 0.0032670877408236265, + 0.0016156306955963373, + 0.0030522311571985483, + 0.00329494453035295, + 0.0035928331781178713, + 0.0027095915284007788, + 0.0027978243306279182, + 0.0034788623452186584, + 0.002503008348867297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.03368807211518288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium profile shot, positioned at a slightly low angle to convey a sense of gentle majesty, gracefully moves sideways, perfectly synchronizing with the central subject: a lanky, gentle giant composed entirely of living wood and foliage. His form is a meticulous and artistic masterpiece of Art Nouveau; his torso and long limbs are crafted from smooth, polished wood that bends and flows in elegant, sinuous lines, with organic patterns artfully carved into the grain. Dense green foliage covers his shoulders and waist, each leaf delicately swirling with veins. He sways with a soothing, hypnotic, and complex pendulum; his wooden torso initiates a slow, rhythmic tilt, while his slender limbs follow in graceful, delayed arcs, creating a mesmerizing, continuous wave of movement that flows through his entire body. His background is a stylized natural backdrop, a carpet of intricate fern patterns and oversized bell-shaped flowers with pearlescent petals. The trees in the background twist their branches into elaborate, organic arabesques, silhouetted against a sky painted in soft, warm colors. The entire scene is bathed in the warm, diffuse glow of the afternoon sun, casting long, soft shadows that accentuate the rich, earthy browns of the wood and the myriad green hues of the foliage, ultimately creating an atmosphere of deep tranquility and enchantment that captures the essence of artistic, fluid naturalism.", + "seed": 29622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_9980_seed_29622_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": -0.03372249687090516, + "vlm_logprob": -0.03372249687090516, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03372249687090516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04713480547070503 + ], + "opposite_logprob_tokens": [ + -3.1721348762512207 + ], + "logprob_reward": -0.04713480547070503, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04713480547070503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1374482959508896, + 0.21523089706897736, + 0.06216779723763466, + 0.034414004534482956, + 0.036309320479631424, + 0.1396503746509552, + 0.060493916273117065, + 0.07589245587587357, + 0.16694603860378265, + 0.07144691795110703 + ], + "credit_map": [ + 0.01552482321858406, + 0.014638313092291355, + 0.025817569345235825, + 0.017535043880343437, + 0.014001852832734585, + 0.018486177548766136, + 0.017952337861061096, + 0.013492168858647346, + 0.018020490184426308, + 0.021441970020532608, + 0.02831215038895607, + 0.028204306960105896, + 0.024738600477576256, + 0.030796805396676064, + 0.03995127975940704, + 0.023765280842781067, + 0.005322717595845461, + 0.006552835926413536, + 0.006850948091596365, + 0.014945673756301403, + 0.005381159018725157, + 0.006292229983955622, + 0.009426272474229336, + 0.00739595852792263, + 0.002702387049794197, + 0.004028453957289457, + 0.0050463322550058365, + 0.00799851305782795, + 0.0026818655896931887, + 0.003709641983732581, + 0.00426698662340641, + 0.003979824483394623, + 0.003417549654841423, + 0.0022095516324043274, + 0.005538237281143665, + 0.007043296936899424, + 0.0033237095922231674, + 0.004107429180294275, + 0.004995708353817463, + 0.005673839710652828, + 0.01653885655105114, + 0.016223596408963203, + 0.02163819782435894, + 0.017054352909326553, + 0.020914139226078987, + 0.014040849171578884, + 0.021786687895655632, + 0.011453695595264435, + 0.0017594785895198584, + 0.0019416471477597952, + 0.0025234718341380358, + 0.003016168950125575, + 0.0025286662857979536, + 0.002179952571168542, + 0.02401953749358654, + 0.022524993866682053, + 0.006280830129981041, + 0.006221732124686241, + 0.009466636925935745, + 0.011432652361690998, + 0.008203604258596897, + 0.010212219320237637, + 0.01325845718383789, + 0.010816330090165138, + 0.012604678981006145, + 0.013689291663467884, + 0.016649018973112106, + 0.03030272014439106, + 0.0260626133531332, + 0.017664119601249695, + 0.026307513937354088, + 0.02366608753800392, + 0.007256731390953064, + 0.0070557123981416225, + 0.009462273679673672, + 0.007108993362635374, + 0.008739037439227104, + 0.012619410641491413, + 0.01051780954003334, + 0.008686951361596584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.04713480547070503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0142061123624444 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.0142061123624444, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0142061123624444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10195079445838928, + 0.15937454998493195, + 0.07688715308904648, + 0.03994138911366463, + 0.04344126954674721, + 0.20405398309230804, + 0.08738980442285538, + 0.05036255344748497, + 0.1642286628484726, + 0.07236981391906738 + ], + "credit_map": [ + 0.007998191751539707, + 0.013873470015823841, + 0.012361329048871994, + 0.012294119223952293, + 0.01379409246146679, + 0.016474876552820206, + 0.012241149321198463, + 0.012913576327264309, + 0.014392443001270294, + 0.01312230620533228, + 0.02435475029051304, + 0.017357323318719864, + 0.02418484166264534, + 0.025273719802498817, + 0.022933604195713997, + 0.017755556851625443, + 0.005821337923407555, + 0.008043493144214153, + 0.008761106990277767, + 0.02023722417652607, + 0.006395331583917141, + 0.007015290204435587, + 0.012120780535042286, + 0.008492589928209782, + 0.002844673115760088, + 0.004511093720793724, + 0.003365890821442008, + 0.01000236440449953, + 0.003296987619251013, + 0.004427289590239525, + 0.007704219315201044, + 0.003788871457800269, + 0.004134421702474356, + 0.0036315054167062044, + 0.0047323694452643394, + 0.011393529362976551, + 0.00268588587641716, + 0.0029826860409229994, + 0.010991538874804974, + 0.0028893377166241407, + 0.025355402380228043, + 0.023177387192845345, + 0.0325055755674839, + 0.023126691579818726, + 0.03137288987636566, + 0.020058754831552505, + 0.032312825322151184, + 0.01614445634186268, + 0.0025954002048820257, + 0.0029940721578896046, + 0.0032885458786040545, + 0.004295815713703632, + 0.005598096176981926, + 0.0034939348697662354, + 0.03271286562085152, + 0.03241107240319252, + 0.0036425674334168434, + 0.003206399967893958, + 0.006979531608521938, + 0.00810268335044384, + 0.004823062568902969, + 0.007661581505089998, + 0.008626038208603859, + 0.007320690434426069, + 0.014434860087931156, + 0.016632044687867165, + 0.012886455282568932, + 0.023053109645843506, + 0.02521207183599472, + 0.017036981880664825, + 0.022953065112233162, + 0.032020073384046555, + 0.00680849002674222, + 0.009083324111998081, + 0.011967761442065239, + 0.00867602787911892, + 0.007504396606236696, + 0.008588316850364208, + 0.011106953024864197, + 0.008634542115032673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.0142061123624444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305106773972511 + ], + "opposite_logprob_tokens": [ + -3.16805100440979 + ], + "logprob_reward": -0.04305106773972511, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305106773972511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14951474964618683, + 0.18644291162490845, + 0.06777545809745789, + 0.07531605660915375, + 0.07019595801830292, + 0.07170607149600983, + 0.06241532787680626, + 0.037514202296733856, + 0.09715414047241211, + 0.181965172290802 + ], + "credit_map": [ + 0.01362277939915657, + 0.016834862530231476, + 0.02480422705411911, + 0.018288947641849518, + 0.017082931473851204, + 0.018305163830518723, + 0.018237674608826637, + 0.022338151931762695, + 0.01412541326135397, + 0.01862659864127636, + 0.020840628072619438, + 0.02091328799724579, + 0.022539980709552765, + 0.036016568541526794, + 0.02826421521604061, + 0.025116227567195892, + 0.005459690932184458, + 0.006748708430677652, + 0.008566895499825478, + 0.015318735502660275, + 0.006234542466700077, + 0.007652001921087503, + 0.010080071166157722, + 0.007714808918535709, + 0.006228698883205652, + 0.00851009227335453, + 0.009103995747864246, + 0.014082912355661392, + 0.0078768040984869, + 0.0055770487524569035, + 0.015949305146932602, + 0.00798719935119152, + 0.005780225619673729, + 0.006411421112716198, + 0.005340281408280134, + 0.01945224031805992, + 0.00496333185583353, + 0.00519626634195447, + 0.007164828013628721, + 0.015887362882494926, + 0.006094832438975573, + 0.005853600800037384, + 0.012108909897506237, + 0.010237978771328926, + 0.005463324952870607, + 0.0069729117676615715, + 0.014409180730581284, + 0.010565333068370819, + 0.0050463308580219746, + 0.0032997066155076027, + 0.0032486801501363516, + 0.0032624267041683197, + 0.00483903381973505, + 0.004201025702059269, + 0.018254615366458893, + 0.020263507962226868, + 0.003125464543700218, + 0.005192650947719812, + 0.0047143897973001, + 0.005002355203032494, + 0.004249846562743187, + 0.005057075060904026, + 0.005430076736956835, + 0.004742340184748173, + 0.006290890276432037, + 0.00874349009245634, + 0.013186417520046234, + 0.01436586119234562, + 0.010759271681308746, + 0.009278882294893265, + 0.019732091575860977, + 0.014797233045101166, + 0.013921935111284256, + 0.019759390503168106, + 0.03226763382554054, + 0.022701941430568695, + 0.020180337131023407, + 0.020332125946879387, + 0.029318127781152725, + 0.023483676835894585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.04305106773972511 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196639310568571 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.002196639310568571, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196639310568571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12690730392932892, + 0.16744236648082733, + 0.04891244322061539, + 0.09443908929824829, + 0.04199261963367462, + 0.1648852825164795, + 0.07894781231880188, + 0.03690057247877121, + 0.12349821627140045, + 0.11607431620359421 + ], + "credit_map": [ + 0.02148493006825447, + 0.012390798889100552, + 0.026145488023757935, + 0.015855982899665833, + 0.00994290690869093, + 0.010161200538277626, + 0.017691941931843758, + 0.013234049081802368, + 0.017080113291740417, + 0.014219683595001698, + 0.028906621038913727, + 0.02350662834942341, + 0.01726522296667099, + 0.018510546535253525, + 0.02131813019514084, + 0.026635412126779556, + 0.004485797602683306, + 0.007380588445812464, + 0.0067160422913730145, + 0.008051431737840176, + 0.004673749674111605, + 0.005518018268048763, + 0.006295687984675169, + 0.005791130010038614, + 0.005920030176639557, + 0.008434477262198925, + 0.011658364906907082, + 0.01802702620625496, + 0.008971160277724266, + 0.009119279682636261, + 0.022106466814875603, + 0.010202279314398766, + 0.003217894583940506, + 0.0041717966087162495, + 0.0035923433024436235, + 0.010846003890037537, + 0.004034857731312513, + 0.0030948726925998926, + 0.0104732820764184, + 0.002561570145189762, + 0.01856357418000698, + 0.01896638423204422, + 0.026351716369390488, + 0.020829180255532265, + 0.02720501273870468, + 0.015533643774688244, + 0.024314874783158302, + 0.013120886869728565, + 0.007157135754823685, + 0.005409208592027426, + 0.0062423390336334705, + 0.0070968130603432655, + 0.011540952138602734, + 0.005327051505446434, + 0.016358934342861176, + 0.01981537975370884, + 0.002513122744858265, + 0.003575701965019107, + 0.004805925767868757, + 0.005995339713990688, + 0.003632121719419956, + 0.004926410503685474, + 0.005993661005049944, + 0.005458288360387087, + 0.009288606233894825, + 0.013032409362494946, + 0.008809871040284634, + 0.016507906839251518, + 0.02238491363823414, + 0.013304688967764378, + 0.018717028200626373, + 0.02145279198884964, + 0.010132585652172565, + 0.0156935453414917, + 0.020516186952590942, + 0.01364101842045784, + 0.011975127272307873, + 0.012419072911143303, + 0.017051633447408676, + 0.014645151793956757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.002196639310568571 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202385947108269 + ], + "opposite_logprob_tokens": [ + -2.812023878097534 + ], + "logprob_reward": -0.06202385947108269, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202385947108269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.483463317155838, + 0.08338256180286407, + 0.037314046174287796, + 0.04323052987456322, + 0.014502080157399178, + 0.04760778695344925, + 0.024897431954741478, + 0.019538259133696556, + 0.09974554926156998, + 0.1463184803724289 + ], + "credit_map": [ + 0.04792251065373421, + 0.061520934104919434, + 0.08765511959791183, + 0.058301929384469986, + 0.03922932595014572, + 0.052188217639923096, + 0.08163636177778244, + 0.0550089031457901, + 0.008276602253317833, + 0.006480342708528042, + 0.013733536936342716, + 0.010771294124424458, + 0.01091677788645029, + 0.009382527321577072, + 0.011004318483173847, + 0.012817162089049816, + 0.0034221759997308254, + 0.0037753074429929256, + 0.004518847446888685, + 0.005688557401299477, + 0.004140899050980806, + 0.004901701584458351, + 0.006096812896430492, + 0.00476974667981267, + 0.003245293628424406, + 0.004478860646486282, + 0.004618003498762846, + 0.007298658136278391, + 0.0038336997386068106, + 0.00585426390171051, + 0.008047189563512802, + 0.005854563321918249, + 0.0012949209194630384, + 0.0022276220843195915, + 0.00169765402097255, + 0.0022238194942474365, + 0.0017689756350591779, + 0.0014368713600561023, + 0.0021237977780401707, + 0.0017284174682572484, + 0.004253696650266647, + 0.0041662463918328285, + 0.007589688990265131, + 0.0060212742537260056, + 0.008206674829125404, + 0.005296784453094006, + 0.00593549245968461, + 0.006137928459793329, + 0.0011538207763805985, + 0.0011575287207961082, + 0.0017767781391739845, + 0.0018042688025161624, + 0.0017986091552302241, + 0.0014165618922561407, + 0.007703341078013182, + 0.008086522109806538, + 0.0015425224555656314, + 0.0021473823580890894, + 0.002623508917167783, + 0.0030073183588683605, + 0.002811035141348839, + 0.0021112284157425165, + 0.0030244633089751005, + 0.002270799595862627, + 0.007367745507508516, + 0.007206287235021591, + 0.011465677991509438, + 0.015968082472682, + 0.01618809625506401, + 0.01194870937615633, + 0.01498804334551096, + 0.014612901024520397, + 0.01137569546699524, + 0.01734689064323902, + 0.024960564449429512, + 0.01796247623860836, + 0.01530248299241066, + 0.015725288540124893, + 0.0236508809030056, + 0.019994188100099564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.06202385947108269 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium profile shot, positioned at a slightly low angle to convey a sense of gentle majesty, gracefully moves sideways, perfectly synchronizing with the central subject: a lanky, gentle giant composed entirely of living wood and foliage. His form is a meticulous and artistic masterpiece of Art Nouveau; his torso and long limbs are crafted from smooth, polished wood that bends and flows in elegant, sinuous lines, with organic patterns artfully carved into the grain. Dense green foliage covers his shoulders and waist, each leaf delicately swirling with veins. He sways with a soothing, hypnotic, and complex pendulum; his wooden torso initiates a slow, rhythmic tilt, while his slender limbs follow in graceful, delayed arcs, creating a mesmerizing, continuous wave of movement that flows through his entire body. His background is a stylized natural backdrop, a carpet of intricate fern patterns and oversized bell-shaped flowers with pearlescent petals. The trees in the background twist their branches into elaborate, organic arabesques, silhouetted against a sky painted in soft, warm colors. The entire scene is bathed in the warm, diffuse glow of the afternoon sun, casting long, soft shadows that accentuate the rich, earthy browns of the wood and the myriad green hues of the foliage, ultimately creating an atmosphere of deep tranquility and enchantment that captures the essence of artistic, fluid naturalism.", + "seed": 29623, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_9980_seed_29623_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": -0.032070249086245894, + "vlm_logprob": -0.032070249086245894, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.032070249086245894, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024675492197275162 + ], + "opposite_logprob_tokens": [ + -3.8996756076812744 + ], + "logprob_reward": -0.024675492197275162, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024675492197275162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.075358085334301, + 0.04437611997127533, + 0.02981070801615715, + 0.02916470356285572, + 0.00938287191092968, + 0.11023742705583572, + 0.016786549240350723, + 0.0174823347479105, + 0.6273941397666931, + 0.04000704362988472 + ], + "credit_map": [ + 0.005467093084007502, + 0.005465207155793905, + 0.005699335131794214, + 0.008428444154560566, + 0.007498784456402063, + 0.006577666383236647, + 0.018310705199837685, + 0.017910854890942574, + 0.0034979989286512136, + 0.002703531878069043, + 0.005170893855392933, + 0.006302577443420887, + 0.004376836121082306, + 0.006310577969998121, + 0.00904051773250103, + 0.0069731902331113815, + 0.0030755179468542337, + 0.003039752598851919, + 0.0031166779808700085, + 0.0038988084997981787, + 0.0033066130708903074, + 0.0051971375942230225, + 0.0032790564000606537, + 0.004897145554423332, + 0.0031933903228491545, + 0.002953511895611882, + 0.003984773997217417, + 0.004062518011778593, + 0.003721205983310938, + 0.0025819786824285984, + 0.005555643234401941, + 0.0031116832979023457, + 0.001023257034830749, + 0.00153551425319165, + 0.0009940469171851873, + 0.0012916544219478965, + 0.0013651560293510556, + 0.0008903054404072464, + 0.0010175704956054688, + 0.0012653686571866274, + 0.02024589665234089, + 0.0069883535616099834, + 0.049085844308137894, + 0.009541894309222698, + 0.001985103590413928, + 0.004295041784644127, + 0.006551514379680157, + 0.011543787084519863, + 0.0026883312966674566, + 0.001789440750144422, + 0.0029264194890856743, + 0.0032254939433187246, + 0.0018905004253610969, + 0.001093187602236867, + 0.002048013499006629, + 0.0011251643300056458, + 0.0011706668883562088, + 0.0017258210573345423, + 0.0021501381415873766, + 0.003132554702460766, + 0.002190955216065049, + 0.0010717608965933323, + 0.0044694202952086926, + 0.0015710192965343595, + 0.04839830845594406, + 0.014388887211680412, + 0.03082733415067196, + 0.03435036167502403, + 0.021181214600801468, + 0.039219729602336884, + 0.26807135343551636, + 0.17095698416233063, + 0.002373694209381938, + 0.0027854738291352987, + 0.011137344874441624, + 0.01251071598380804, + 0.002313528675585985, + 0.0015106460778042674, + 0.0045301709324121475, + 0.0028454686980694532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.024675492197275162 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792149998247623 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792149998247623, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792149998247623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1726381629705429, + 0.10310696065425873, + 0.05466620624065399, + 0.047650046646595, + 0.02592712640762329, + 0.11598692834377289, + 0.046101734042167664, + 0.17095494270324707, + 0.19665542244911194, + 0.06631243228912354 + ], + "credit_map": [ + 0.01448810938745737, + 0.01285454723984003, + 0.01712578721344471, + 0.020900195464491844, + 0.021008403971791267, + 0.016805164515972137, + 0.03584503009915352, + 0.03361092507839203, + 0.007548444904386997, + 0.009309601038694382, + 0.014874055981636047, + 0.012834129855036736, + 0.009099237620830536, + 0.012442772276699543, + 0.017408574000000954, + 0.019590148702263832, + 0.006139320321381092, + 0.00545324943959713, + 0.005728638730943203, + 0.008880148641765118, + 0.007000156212598085, + 0.007391517050564289, + 0.007635928224772215, + 0.006437244359403849, + 0.005242560058832169, + 0.0038346732035279274, + 0.00571781862527132, + 0.005835976917296648, + 0.005244978237897158, + 0.004260311834514141, + 0.011452117934823036, + 0.006061606574803591, + 0.0036060973070561886, + 0.0038789701648056507, + 0.0031296273227781057, + 0.0031557062175124884, + 0.004752289038151503, + 0.0029171896167099476, + 0.0023415624164044857, + 0.00214568292722106, + 0.02389482781291008, + 0.006980688311159611, + 0.04142080247402191, + 0.011981136165559292, + 0.002678666030988097, + 0.010280633345246315, + 0.008114897646009922, + 0.010635275393724442, + 0.004216447472572327, + 0.0045721097849309444, + 0.0062194461934268475, + 0.007813525386154652, + 0.006104428321123123, + 0.0052331662736833096, + 0.006553852465003729, + 0.005388758610934019, + 0.011056714691221714, + 0.016291707754135132, + 0.01546978298574686, + 0.055053725838661194, + 0.01215214654803276, + 0.013511864468455315, + 0.035508282482624054, + 0.011910727247595787, + 0.016849447041749954, + 0.02039146050810814, + 0.016877854242920876, + 0.01993142068386078, + 0.01906374655663967, + 0.013871471397578716, + 0.048857204616069794, + 0.04081283509731293, + 0.005091401748359203, + 0.007443603593856096, + 0.01298818364739418, + 0.014674731530249119, + 0.006339875515550375, + 0.0037629029247909784, + 0.01073975395411253, + 0.005271976813673973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.009792149998247623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369026258587837 + ], + "opposite_logprob_tokens": [ + -3.4086902141571045 + ], + "logprob_reward": -0.03369026258587837, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369026258587837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07883422076702118, + 0.08748866617679596, + 0.048725675791502, + 0.06208766996860504, + 0.02458076924085617, + 0.22408641874790192, + 0.037479743361473083, + 0.08670706301927567, + 0.2385282814502716, + 0.11148150265216827 + ], + "credit_map": [ + 0.009637122973799706, + 0.009671434760093689, + 0.010321849025785923, + 0.009561554528772831, + 0.009062103927135468, + 0.010989954695105553, + 0.009738452732563019, + 0.009851737879216671, + 0.01010456494987011, + 0.009207314811646938, + 0.010426389053463936, + 0.010495450347661972, + 0.010442939586937428, + 0.012785719707608223, + 0.012086409144103527, + 0.011939872987568378, + 0.006519705057144165, + 0.0046035004779696465, + 0.004992304835468531, + 0.005057055968791246, + 0.0068629104644060135, + 0.00947279017418623, + 0.005449909716844559, + 0.005767498165369034, + 0.007269117049872875, + 0.00603104755282402, + 0.011708175763487816, + 0.0072874641045928, + 0.009828019887208939, + 0.005793227814137936, + 0.00784163549542427, + 0.006328975316137075, + 0.00258789025247097, + 0.002947547473013401, + 0.0030120275914669037, + 0.0027371435426175594, + 0.004897945560514927, + 0.0030372419860213995, + 0.002615684410557151, + 0.002745287958532572, + 0.03478158637881279, + 0.014136680401861668, + 0.10237955302000046, + 0.019995396956801414, + 0.0034976184833794832, + 0.012265454977750778, + 0.012961889617145061, + 0.024068210273981094, + 0.005377409514039755, + 0.004277006722986698, + 0.0076028225012123585, + 0.004190780222415924, + 0.004359688144177198, + 0.004100430756807327, + 0.0041810981929302216, + 0.0033905035816133022, + 0.005361433606594801, + 0.0061312709003686905, + 0.005647372920066118, + 0.033899880945682526, + 0.006390213035047054, + 0.005095308180898428, + 0.018630487844347954, + 0.005551092326641083, + 0.018971944227814674, + 0.01194855198264122, + 0.016319742426276207, + 0.016333188861608505, + 0.010505599901080132, + 0.017013533040881157, + 0.09470189362764359, + 0.05273381620645523, + 0.006233519874513149, + 0.004588137846440077, + 0.037370726466178894, + 0.04250635579228401, + 0.0035042501986026764, + 0.003305467776954174, + 0.006615933030843735, + 0.007357096299529076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.03369026258587837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1760791689157486, + 0.042969148606061935, + 0.030377304181456566, + 0.05135932192206383, + 0.055204588919878006, + 0.18201453983783722, + 0.0528474822640419, + 0.06828843802213669, + 0.13595739006996155, + 0.20490260422229767 + ], + "credit_map": [ + 0.012858150526881218, + 0.016277547925710678, + 0.01776980049908161, + 0.017409855499863625, + 0.019864138215780258, + 0.015021231025457382, + 0.04356935992836952, + 0.03330908343195915, + 0.005185682792216539, + 0.0033664293587207794, + 0.005530993454158306, + 0.004392711911350489, + 0.004619004670530558, + 0.006263591814786196, + 0.0059761833399534225, + 0.007634551730006933, + 0.0053223962895572186, + 0.002552617806941271, + 0.0035036522895097733, + 0.003802068531513214, + 0.0031267416197806597, + 0.0038533390033990145, + 0.003663046518340707, + 0.004553442820906639, + 0.005328820087015629, + 0.006425022147595882, + 0.006487248931080103, + 0.0073523251339793205, + 0.008677400648593903, + 0.004503489006310701, + 0.007623412646353245, + 0.004961603321135044, + 0.0057684676721692085, + 0.00755991879850626, + 0.007437257561832666, + 0.008645354770123959, + 0.008249434642493725, + 0.006273066159337759, + 0.005291747394949198, + 0.005979340989142656, + 0.03517324477434158, + 0.010979250073432922, + 0.07541005313396454, + 0.015516084618866444, + 0.0038493459578603506, + 0.011434332467615604, + 0.012014508247375488, + 0.017637720331549644, + 0.008550970815122128, + 0.006617453880608082, + 0.006393148098140955, + 0.007682489231228828, + 0.006382660008966923, + 0.0032504198607057333, + 0.005974007770419121, + 0.007996330969035625, + 0.005084395874291658, + 0.006136204116046429, + 0.005870627239346504, + 0.026502391323447227, + 0.0033577759750187397, + 0.002610333962365985, + 0.013891496695578098, + 0.004835216794162989, + 0.010023614391684532, + 0.010800523683428764, + 0.011620236560702324, + 0.011187230236828327, + 0.007830040529370308, + 0.010680194944143295, + 0.04917531833052635, + 0.024640237912535667, + 0.012332100421190262, + 0.009505254216492176, + 0.06489352881908417, + 0.07677417248487473, + 0.007057853974401951, + 0.006393291521817446, + 0.013526488095521927, + 0.014419917948544025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0031914988067001104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a lanky figure composed entirely of living wood and foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's torso and limbs crafted from smooth, polished wood with elegant, sinuous lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the subject sway with a slow, rhythmic pendulum motion where limbs follow in graceful, delayed arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a stylized natural scene featuring intricate fern patterns and oversized bell-shaped flowers with pearlescent petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, diffuse glow of the afternoon sun casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890018418431282 + ], + "opposite_logprob_tokens": [ + -2.4640018939971924 + ], + "logprob_reward": -0.0890018418431282, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890018418431282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05168658494949341, + 0.03859247639775276, + 0.027124332264065742, + 0.03927354887127876, + 0.009755113162100315, + 0.06418298184871674, + 0.024932418018579483, + 0.04481814056634903, + 0.09812448173761368, + 0.6015099287033081 + ], + "credit_map": [ + 0.00467772176489234, + 0.0039879800751805305, + 0.00519987428560853, + 0.005368435755372047, + 0.00407140888273716, + 0.004802507348358631, + 0.012898451648652554, + 0.010680208913981915, + 0.003982617054134607, + 0.003436187980696559, + 0.0045508770272135735, + 0.005779766943305731, + 0.0063486648723483086, + 0.0044491831213235855, + 0.004800843540579081, + 0.0052443379536271095, + 0.004170107655227184, + 0.003012053668498993, + 0.002723798155784607, + 0.002885582158342004, + 0.002313267905265093, + 0.0047959936782717705, + 0.0028248068410903215, + 0.004398722667247057, + 0.002658546669408679, + 0.003374616615474224, + 0.004821545910090208, + 0.007260257378220558, + 0.005626281723380089, + 0.0035540980752557516, + 0.006652776151895523, + 0.005325426813215017, + 0.001444116234779358, + 0.0012073736870661378, + 0.0009059312869794667, + 0.001278512878343463, + 0.0013964539393782616, + 0.0012148875975981355, + 0.001124572241678834, + 0.001183265820145607, + 0.010122774168848991, + 0.004202641546726227, + 0.029004044830799103, + 0.005804689135402441, + 0.001274492242373526, + 0.0028791669756174088, + 0.0038416532333940268, + 0.007053522393107414, + 0.005155425984412432, + 0.003090338781476021, + 0.0034183496609330177, + 0.002831983845680952, + 0.00499587832018733, + 0.0019183491822332144, + 0.0021201104391366243, + 0.0014019819209352136, + 0.0029739425517618656, + 0.004535825457423925, + 0.003916461020708084, + 0.016542095690965652, + 0.002422432182356715, + 0.00246421224437654, + 0.009072660468518734, + 0.0028905163053423166, + 0.008151955902576447, + 0.009608713909983635, + 0.007933473214507103, + 0.008065540343523026, + 0.007187222596257925, + 0.007067548111081123, + 0.030749594792723656, + 0.019360439851880074, + 0.033331163227558136, + 0.02679591439664364, + 0.19949519634246826, + 0.23667381703853607, + 0.0174412839114666, + 0.019806278869509697, + 0.027976645156741142, + 0.039989590644836426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.0890018418431282 + } + ], + "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/0000028/group_14/index_tensor([8041, 8041])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_14/index_tensor([8041, 8041])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e3099b57e5a94e8e26060b1b7de1c7278f630d63 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_14/index_tensor([8041, 8041])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, dynamic tracking shot follows a sleek, jet-black coupe as it hurtles forward at breakneck speed on an empty highway shrouded in darkness. Its sleek metallic paintwork gleams with a cold sheen in the darkness, reflecting the scattered ambient light. Its sharp LED headlights cast two piercing, straight beams of white light, slicing the black asphalt ahead, slightly slippery from the light rain, clearly illuminating its rough texture and the gleaming water marks. As the vehicle speeds along, guardrails and blurred signage flash by in the beams. The wheels spin at high speed, kicking up a fine mist that becomes barely visible in the red taillights. The scene, dominated by deep blue and pure black, is punctuated by the brilliant white of the headlights and the crimson of the taillights, creating a cinematic atmosphere of intense speed, tension, and the loneliness and power of a solitary night traveler.", + "seed": 13328, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_14/index_8041_seed_13328_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": -0.054285401245579123, + "vlm_logprob": -0.054285401245579123, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.054285401245579123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14518004655838013 + ], + "opposite_logprob_tokens": [ + -2.0201799869537354 + ], + "logprob_reward": -0.14518004655838013, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14518004655838013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08514079451560974, + 0.32515934109687805, + 0.029201889410614967, + 0.05181746557354927, + 0.00961381383240223, + 0.06986054033041, + 0.05455603823065758, + 0.054322998970746994, + 0.104631207883358, + 0.21569596230983734 + ], + "credit_map": [ + 0.008293078280985355, + 0.007280595134943724, + 0.006913338787853718, + 0.006504440680146217, + 0.012444178573787212, + 0.014351489953696728, + 0.01491733267903328, + 0.014436342753469944, + 0.006539213005453348, + 0.014712836593389511, + 0.006838615983724594, + 0.0065502808429300785, + 0.13984768092632294, + 0.1367497444152832, + 0.008316191844642162, + 0.005604766774922609, + 0.0037683402188122272, + 0.0031362920999526978, + 0.005006768275052309, + 0.0018195039592683315, + 0.003853726899251342, + 0.0038473117165267467, + 0.003257390344515443, + 0.004512554034590721, + 0.007251674309372902, + 0.004283876623958349, + 0.014845918864011765, + 0.0033649615943431854, + 0.0045301299542188644, + 0.00734347989782691, + 0.00371442548930645, + 0.006482995115220547, + 0.001188223366625607, + 0.0009484968613833189, + 0.0015153198037296534, + 0.0011017614742740989, + 0.0010097796330228448, + 0.0014886952703818679, + 0.0008836601627990603, + 0.0014778777258470654, + 0.005508518777787685, + 0.007859504781663418, + 0.01072959415614605, + 0.007344888988882303, + 0.007263637147843838, + 0.011018798686563969, + 0.009468170814216137, + 0.010667423717677593, + 0.011823325417935848, + 0.0042253960855305195, + 0.011218352243304253, + 0.0046963803470134735, + 0.004054076503962278, + 0.01021580770611763, + 0.0025207046419382095, + 0.005801996681839228, + 0.007634684443473816, + 0.005312308203428984, + 0.0068888855166733265, + 0.008273451589047909, + 0.005347005557268858, + 0.006162915378808975, + 0.004112901631742716, + 0.010590851306915283, + 0.010076816193759441, + 0.010507194325327873, + 0.02978852391242981, + 0.019825376570224762, + 0.007773891091346741, + 0.007703100331127644, + 0.007822033949196339, + 0.011134274303913116, + 0.02309422381222248, + 0.03221692889928818, + 0.047691088169813156, + 0.03894590958952904, + 0.012080246582627296, + 0.016106460243463516, + 0.02997458167374134, + 0.01558651588857174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.14518004655838013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820380985736847 + ], + "opposite_logprob_tokens": [ + -4.0182037353515625 + ], + "logprob_reward": -0.01820380985736847, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820380985736847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11286648362874985, + 0.2566524147987366, + 0.043217986822128296, + 0.070366770029068, + 0.04162843897938728, + 0.07294046878814697, + 0.06948096305131912, + 0.06438175588846207, + 0.07750321924686432, + 0.19096145033836365 + ], + "credit_map": [ + 0.012018909677863121, + 0.010446921922266483, + 0.00783549901098013, + 0.009238972328603268, + 0.017531774938106537, + 0.019345253705978394, + 0.017561815679073334, + 0.018887341022491455, + 0.01241261139512062, + 0.01128642912954092, + 0.008615228347480297, + 0.006531669292598963, + 0.09953392297029495, + 0.08817548304796219, + 0.015055429190397263, + 0.015041646547615528, + 0.00418171426281333, + 0.004432336892932653, + 0.0068886615335941315, + 0.0026566204614937305, + 0.005531830247491598, + 0.007547970861196518, + 0.005033787339925766, + 0.00694506149739027, + 0.0073149194940924644, + 0.00735269021242857, + 0.009633955545723438, + 0.004449084866791964, + 0.0062951804138720036, + 0.011777314357459545, + 0.005052671767771244, + 0.018490953370928764, + 0.004149490501731634, + 0.0027704627718776464, + 0.005414758343249559, + 0.0034212239552289248, + 0.0036850138567388058, + 0.007023475132882595, + 0.003349743550643325, + 0.011814272962510586, + 0.005227462854236364, + 0.006509930826723576, + 0.016016703099012375, + 0.005984173156321049, + 0.009626343846321106, + 0.011365736834704876, + 0.007945210672914982, + 0.010264907963573933, + 0.021125905215740204, + 0.004686414729803801, + 0.013635224662721157, + 0.0037886667996644974, + 0.00393286719918251, + 0.015278683975338936, + 0.0026741866022348404, + 0.00435901153832674, + 0.007043037097901106, + 0.008630478754639626, + 0.009615517221391201, + 0.014246698468923569, + 0.005727433133870363, + 0.005724840797483921, + 0.005883730947971344, + 0.007510017137974501, + 0.005829562433063984, + 0.0114876264706254, + 0.013516110368072987, + 0.016346868127584457, + 0.00799864623695612, + 0.006785837467759848, + 0.007728526834398508, + 0.007810044102370739, + 0.023139139637351036, + 0.032620880752801895, + 0.04090103134512901, + 0.03031107783317566, + 0.01157926581799984, + 0.014835601672530174, + 0.02188129350543022, + 0.01569315791130066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.01820380985736847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024944401811808348 + ], + "opposite_logprob_tokens": [ + -6.0024943351745605 + ], + "logprob_reward": -0.0024944401811808348, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024944401811808348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1330789178609848, + 0.18500803411006927, + 0.07319962978363037, + 0.07764709740877151, + 0.06642361730337143, + 0.08178773522377014, + 0.1446208655834198, + 0.039211027324199677, + 0.11433739960193634, + 0.08468563109636307 + ], + "credit_map": [ + 0.012655504047870636, + 0.01869613490998745, + 0.013872426003217697, + 0.01896091178059578, + 0.017015505582094193, + 0.0190996453166008, + 0.01618708111345768, + 0.016591714695096016, + 0.009534051641821861, + 0.014127466827630997, + 0.01058625616133213, + 0.008685280568897724, + 0.05606517195701599, + 0.05743495747447014, + 0.015002678148448467, + 0.01357218250632286, + 0.010399454273283482, + 0.01221397239714861, + 0.010968007147312164, + 0.0064763268455863, + 0.009084711782634258, + 0.010105506516993046, + 0.0077208662405610085, + 0.006230790168046951, + 0.010408897884190083, + 0.009567421860992908, + 0.011561624705791473, + 0.005906104110181332, + 0.00804742518812418, + 0.011341908946633339, + 0.006112799514085054, + 0.014700922183692455, + 0.009979594498872757, + 0.00855683907866478, + 0.006401754915714264, + 0.006162269972264767, + 0.006716920994222164, + 0.009878232143819332, + 0.0056642331182956696, + 0.013063779100775719, + 0.007375699933618307, + 0.009452959522604942, + 0.008566695265471935, + 0.006990332622081041, + 0.01245585922151804, + 0.012520508840680122, + 0.009215130470693111, + 0.015210561454296112, + 0.046425800770521164, + 0.010115733370184898, + 0.027424689382314682, + 0.007997323758900166, + 0.009087424725294113, + 0.028777306899428368, + 0.005840969737619162, + 0.008951622992753983, + 0.003335882443934679, + 0.004251605365425348, + 0.004895360209047794, + 0.007270680274814367, + 0.0033888020552694798, + 0.004388468340039253, + 0.004969456698745489, + 0.006710776127874851, + 0.011987296864390373, + 0.01741640269756317, + 0.03093915432691574, + 0.018679127097129822, + 0.010409563779830933, + 0.00924896914511919, + 0.00678290193900466, + 0.008873997256159782, + 0.008470618166029453, + 0.011321686208248138, + 0.015925027430057526, + 0.014719120226800442, + 0.00774845015257597, + 0.006227367557585239, + 0.010552199557423592, + 0.009721166454255581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.0024944401811808348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525154173374176 + ], + "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": [ + 0.12837888300418854, + 0.11643237620592117, + 0.06723655015230179, + 0.030904380604624748, + 0.06426770985126495, + 0.22009649872779846, + 0.08589446544647217, + 0.0744495689868927, + 0.12318811565637589, + 0.08915137499570847 + ], + "credit_map": [ + 0.01455777883529663, + 0.01750415936112404, + 0.013024220243096352, + 0.01414534542709589, + 0.016389120370149612, + 0.01871429942548275, + 0.01653227210044861, + 0.01751171238720417, + 0.008791448548436165, + 0.005806751549243927, + 0.005920708645135164, + 0.003563164034858346, + 0.039553914219141006, + 0.03377511352300644, + 0.008588124066591263, + 0.010433170944452286, + 0.006809903774410486, + 0.005612418986856937, + 0.012446972541511059, + 0.0038897504564374685, + 0.009541033767163754, + 0.011205364018678665, + 0.008874626830220222, + 0.00885649025440216, + 0.003968429286032915, + 0.0025086926762014627, + 0.006594651844352484, + 0.0021226746030151844, + 0.0028433981351554394, + 0.004637641832232475, + 0.002776733133941889, + 0.005452164448797703, + 0.008461901918053627, + 0.005481533706188202, + 0.0067360359244048595, + 0.005611420143395662, + 0.005079949740320444, + 0.010271227918565273, + 0.00498248590156436, + 0.017643164843320847, + 0.016147736459970474, + 0.020823493599891663, + 0.031200170516967773, + 0.03635617345571518, + 0.022179679945111275, + 0.027438262477517128, + 0.03291970491409302, + 0.03303131088614464, + 0.019793225452303886, + 0.006287580821663141, + 0.0159648098051548, + 0.007572703994810581, + 0.005823367740958929, + 0.014859303832054138, + 0.00420944532379508, + 0.011384047567844391, + 0.009970489889383316, + 0.007908863015472889, + 0.011035303585231304, + 0.011935010552406311, + 0.0065342728048563, + 0.008226951584219933, + 0.004936298821121454, + 0.013902386650443077, + 0.012787247076630592, + 0.015874650329351425, + 0.032327644526958466, + 0.023437459021806717, + 0.010345835238695145, + 0.009518074803054333, + 0.008818522095680237, + 0.010078704915940762, + 0.007788847666233778, + 0.013546464964747429, + 0.012436355464160442, + 0.02125369943678379, + 0.005936529953032732, + 0.0056625488214194775, + 0.012988605536520481, + 0.009538335725665092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.00525154173374176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029716789722443 + ], + "opposite_logprob_tokens": [ + -2.350297212600708 + ], + "logprob_reward": -0.10029716789722443, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029716789722443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06331367790699005, + 0.3529955744743347, + 0.06414198875427246, + 0.03845895454287529, + 0.031771544367074966, + 0.14786136150360107, + 0.057891350239515305, + 0.06520690023899078, + 0.05001044273376465, + 0.128348246216774 + ], + "credit_map": [ + 0.005289044231176376, + 0.009085308760404587, + 0.00513669615611434, + 0.0056131454184651375, + 0.010332598350942135, + 0.011973956599831581, + 0.00725530693307519, + 0.008627623319625854, + 0.006191394757479429, + 0.011402065865695477, + 0.006005024071782827, + 0.006057063117623329, + 0.1563895344734192, + 0.1507173776626587, + 0.008290904574096203, + 0.007942221127450466, + 0.005684578791260719, + 0.005410867743194103, + 0.008328553289175034, + 0.0036103385500609875, + 0.008714722469449043, + 0.009758156724274158, + 0.007006230764091015, + 0.01562853530049324, + 0.004806492943316698, + 0.006493428256362677, + 0.0052825212478637695, + 0.0032111655455082655, + 0.005006913095712662, + 0.005267588887363672, + 0.0035754467826336622, + 0.004815399646759033, + 0.005148481111973524, + 0.0029021373484283686, + 0.004446025472134352, + 0.0026430084835737944, + 0.0034643837716430426, + 0.004590644035488367, + 0.004216761328279972, + 0.004360102582722902, + 0.014069897122681141, + 0.016848305240273476, + 0.022568250074982643, + 0.017123930156230927, + 0.018460409715771675, + 0.01566547527909279, + 0.023126592859625816, + 0.019998501986265182, + 0.015527625568211079, + 0.004574527498334646, + 0.009969358332455158, + 0.003435918129980564, + 0.003773634321987629, + 0.011506721377372742, + 0.00260821171104908, + 0.0064953528344631195, + 0.008843768388032913, + 0.005517670884728432, + 0.007936198264360428, + 0.010557269677519798, + 0.006406479515135288, + 0.006334464065730572, + 0.004756653215736151, + 0.01485439483076334, + 0.004290743265300989, + 0.006476219743490219, + 0.008857570588588715, + 0.012028234079480171, + 0.00437937444075942, + 0.004368503578007221, + 0.004524126648902893, + 0.005085669457912445, + 0.015812009572982788, + 0.022174393758177757, + 0.031017184257507324, + 0.019642001017928123, + 0.008138044737279415, + 0.008505861274898052, + 0.013459400273859501, + 0.009599346667528152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.10029716789722443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a sleek, jet-black coupe as it hurtles forward at breakneck speed on an empty highway shrouded in darkness. Its sleek metallic paintwork gleams with a cold sheen in the darkness, reflecting the scattered ambient light. Its sharp LED headlights cast two piercing, straight beams of white light, slicing the black asphalt ahead, slightly slippery from the light rain, clearly illuminating its rough texture and the gleaming water marks. As the vehicle speeds along, guardrails and blurred signage flash by in the beams. The wheels spin at high speed, kicking up a fine mist that becomes barely visible in the red taillights. The scene, dominated by deep blue and pure black, is punctuated by the brilliant white of the headlights and the crimson of the taillights, creating a cinematic atmosphere of intense speed, tension, and the loneliness and power of a solitary night traveler.", + "seed": 13329, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_14/index_8041_seed_13329_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": -0.07285820962861181, + "vlm_logprob": -0.07285820962861181, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.07285820962861181, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08735960721969604 + ], + "opposite_logprob_tokens": [ + -2.587359666824341 + ], + "logprob_reward": -0.08735960721969604, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08735960721969604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08501607924699783, + 0.06555870920419693, + 0.028196152299642563, + 0.020097648724913597, + 0.05914703756570816, + 0.14346647262573242, + 0.15251655876636505, + 0.12326590716838837, + 0.15212878584861755, + 0.17060668766498566 + ], + "credit_map": [ + 0.010228198952972889, + 0.011420524679124355, + 0.009450702928006649, + 0.01069383230060339, + 0.012371686287224293, + 0.010841860435903072, + 0.010218649171292782, + 0.009790627285838127, + 0.008241875097155571, + 0.008882452733814716, + 0.006191438063979149, + 0.008164046332240105, + 0.011064304038882256, + 0.009725484065711498, + 0.006581678055226803, + 0.006707432679831982, + 0.0030810199677944183, + 0.0017126231687143445, + 0.002584754256531596, + 0.0018659444758668542, + 0.0025953955482691526, + 0.007591198664158583, + 0.0022602363023906946, + 0.00650497991591692, + 0.0014138869009912014, + 0.0022311671636998653, + 0.003280484350398183, + 0.0023842728696763515, + 0.0036121723242104053, + 0.0025060386396944523, + 0.0020308003295212984, + 0.002638827543705702, + 0.007741767447441816, + 0.004157078452408314, + 0.0065481653437018394, + 0.005853729322552681, + 0.007589982356876135, + 0.008714277297258377, + 0.005903436802327633, + 0.012638597749173641, + 0.014049604535102844, + 0.010833965614438057, + 0.009238681755959988, + 0.02133641205728054, + 0.01318108756095171, + 0.03327646479010582, + 0.015692155808210373, + 0.02585810422897339, + 0.01512198243290186, + 0.012628457508981228, + 0.006575847510248423, + 0.009639183059334755, + 0.046432334929704666, + 0.016696348786354065, + 0.027446886524558067, + 0.01797551102936268, + 0.009182918816804886, + 0.010953808203339577, + 0.014295958913862705, + 0.01254109013825655, + 0.019989024847745895, + 0.017964676022529602, + 0.022224877029657364, + 0.01611354947090149, + 0.011397227644920349, + 0.01693921536207199, + 0.01185111328959465, + 0.018474137410521507, + 0.02678493782877922, + 0.019922101870179176, + 0.024915864691138268, + 0.021844198927283287, + 0.024288317188620567, + 0.02957288920879364, + 0.019166775047779083, + 0.022971810773015022, + 0.014554220251739025, + 0.01857161708176136, + 0.01665199175477028, + 0.024829067289829254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.08735960721969604 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308418184518814 + ], + "opposite_logprob_tokens": [ + -3.168084144592285 + ], + "logprob_reward": -0.04308418184518814, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308418184518814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15543141961097717, + 0.06664252281188965, + 0.0339110903441906, + 0.03460637480020523, + 0.07528699934482574, + 0.17560411989688873, + 0.0777333676815033, + 0.06942978501319885, + 0.1692398488521576, + 0.14211446046829224 + ], + "credit_map": [ + 0.02719765342772007, + 0.01763726957142353, + 0.01300015114247799, + 0.0163237527012825, + 0.025078922510147095, + 0.023987622931599617, + 0.015622875653207302, + 0.016583172604441643, + 0.008310495875775814, + 0.009650692343711853, + 0.00897703506052494, + 0.009021257981657982, + 0.009662018157541752, + 0.008299699053168297, + 0.00636741379275918, + 0.0063539124093949795, + 0.005569134373217821, + 0.002655658172443509, + 0.003983237314969301, + 0.0031065489165484905, + 0.005061705131083727, + 0.005765180103480816, + 0.004065806046128273, + 0.0037038219161331654, + 0.003208212088793516, + 0.0025788978673517704, + 0.005687072407454252, + 0.003220801241695881, + 0.003726789727807045, + 0.004521173425018787, + 0.004352089017629623, + 0.007311339024454355, + 0.006133120972663164, + 0.013031368143856525, + 0.007733362261205912, + 0.008649079129099846, + 0.013978837057948112, + 0.009427962824702263, + 0.008308504708111286, + 0.008024761453270912, + 0.017315195873379707, + 0.015520645305514336, + 0.013467574492096901, + 0.03352009505033493, + 0.01617267169058323, + 0.025608180090785027, + 0.020052827894687653, + 0.0339469350874424, + 0.008102728053927422, + 0.010365675203502178, + 0.0044812303967773914, + 0.0073974947445094585, + 0.021037450060248375, + 0.007673035841435194, + 0.010226421058177948, + 0.008449330925941467, + 0.006499070208519697, + 0.007787044160068035, + 0.009149887599050999, + 0.008607645519077778, + 0.009418632835149765, + 0.00992842111736536, + 0.009296543896198273, + 0.008742541074752808, + 0.020736945793032646, + 0.02259165234863758, + 0.01741783134639263, + 0.02058662474155426, + 0.02200336754322052, + 0.019866278395056725, + 0.02079908363521099, + 0.025238066911697388, + 0.021714642643928528, + 0.0231125820428133, + 0.014897891320288181, + 0.020549895241856575, + 0.013081965036690235, + 0.01649593934416771, + 0.014065982773900032, + 0.018195563927292824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.04308418184518814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954858381301165 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954858381301165, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954858381301165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13031278550624847, + 0.08924289047718048, + 0.08174117654561996, + 0.052346423268318176, + 0.049582723528146744, + 0.15636958181858063, + 0.1327420324087143, + 0.08062589168548584, + 0.07158822566270828, + 0.1554483026266098 + ], + "credit_map": [ + 0.020342376083135605, + 0.017320752143859863, + 0.017244761809706688, + 0.0075180926360189915, + 0.022485971450805664, + 0.021406937390565872, + 0.013384773395955563, + 0.010609129443764687, + 0.01586899161338806, + 0.01153877004981041, + 0.008192707784473896, + 0.010045704431831837, + 0.015097136609256268, + 0.012139496393501759, + 0.007412184029817581, + 0.008947894908487797, + 0.00633836816996336, + 0.009012449532747269, + 0.006490370258688927, + 0.00819823332130909, + 0.014156919904053211, + 0.013923754915595055, + 0.007620431017130613, + 0.016000647097826004, + 0.004559815861284733, + 0.008007226511836052, + 0.006863398943096399, + 0.00424276664853096, + 0.008124792017042637, + 0.0064989798702299595, + 0.005448488984256983, + 0.00860095489770174, + 0.0060014985501766205, + 0.006066249217838049, + 0.00671321852132678, + 0.004719828255474567, + 0.007237820420414209, + 0.006369876675307751, + 0.005170535296201706, + 0.007303694728761911, + 0.010868433862924576, + 0.009695470333099365, + 0.007598965894430876, + 0.019113941118121147, + 0.015177061781287193, + 0.03945932537317276, + 0.015507654286921024, + 0.038948725908994675, + 0.008856438100337982, + 0.006968591827899218, + 0.005433931481093168, + 0.008084319531917572, + 0.015466159209609032, + 0.011937655508518219, + 0.04251641407608986, + 0.0334785133600235, + 0.00786934420466423, + 0.006110216490924358, + 0.007608448620885611, + 0.008527649566531181, + 0.010707818903028965, + 0.018396485596895218, + 0.012186544016003609, + 0.00921938382089138, + 0.007923850789666176, + 0.008746557869017124, + 0.00627986341714859, + 0.00732989190146327, + 0.009539837017655373, + 0.011347441002726555, + 0.009682027623057365, + 0.010738754644989967, + 0.017457976937294006, + 0.023632749915122986, + 0.018579544499516487, + 0.024070963263511658, + 0.012987587600946426, + 0.020027443766593933, + 0.015395565889775753, + 0.023296456784009933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.005954858381301165 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634398825466633 + ], + "opposite_logprob_tokens": [ + -3.651344060897827 + ], + "logprob_reward": -0.02634398825466633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634398825466633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09513094276189804, + 0.053854942321777344, + 0.04411221668124199, + 0.0671376883983612, + 0.07149893790483475, + 0.1989147812128067, + 0.0667274221777916, + 0.10247109830379486, + 0.22159099578857422, + 0.07856099307537079 + ], + "credit_map": [ + 0.018083931878209114, + 0.008906071074306965, + 0.008744126185774803, + 0.0067812604829669, + 0.015799256041646004, + 0.013637530617415905, + 0.008542636409401894, + 0.014636127278208733, + 0.007658021524548531, + 0.00783273670822382, + 0.0049016401171684265, + 0.007117978297173977, + 0.009145131334662437, + 0.0073198541067540646, + 0.004632072057574987, + 0.0052475035190582275, + 0.005940282717347145, + 0.003407916985452175, + 0.0057685477659106255, + 0.003474129131063819, + 0.007567552849650383, + 0.0077571384608745575, + 0.004833472426980734, + 0.005363175179809332, + 0.005995578598231077, + 0.007748026866465807, + 0.011547333560883999, + 0.004721679259091616, + 0.012409546412527561, + 0.008465887047350407, + 0.007053551729768515, + 0.009196092374622822, + 0.006684086751192808, + 0.01115228421986103, + 0.007196927443146706, + 0.00819871760904789, + 0.01263089757412672, + 0.008363529108464718, + 0.008208416402339935, + 0.009064076468348503, + 0.011764725670218468, + 0.012320061214268208, + 0.010211633518338203, + 0.016225863248109818, + 0.013533010147511959, + 0.06066874414682388, + 0.014439143240451813, + 0.05975159630179405, + 0.0056135570630431175, + 0.0065986393019557, + 0.00430042389780283, + 0.006099340971559286, + 0.018405064940452576, + 0.006589632481336594, + 0.011114907450973988, + 0.008005854673683643, + 0.00798544380813837, + 0.008446352556347847, + 0.01197867002338171, + 0.012884153984487057, + 0.01636049523949623, + 0.013992968946695328, + 0.017263242974877357, + 0.013559765182435513, + 0.0340224951505661, + 0.03205697983503342, + 0.022595427930355072, + 0.02100003883242607, + 0.022737963125109673, + 0.03473770245909691, + 0.023053361102938652, + 0.031387023627758026, + 0.007241167593747377, + 0.013610057532787323, + 0.009507190436124802, + 0.01428979355841875, + 0.006943098735064268, + 0.010428618639707565, + 0.0055023422464728355, + 0.011038725264370441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.02634398825466633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154841244220734 + ], + "opposite_logprob_tokens": [ + -1.701548457145691 + ], + "logprob_reward": -0.20154841244220734, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154841244220734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11689187586307526, + 0.039255090057849884, + 0.027220822870731354, + 0.055378906428813934, + 0.07300274819135666, + 0.2718585729598999, + 0.09701390564441681, + 0.08521914482116699, + 0.12537658214569092, + 0.10878237336874008 + ], + "credit_map": [ + 0.014095391146838665, + 0.012720257975161076, + 0.014769310131669044, + 0.0072362665086984634, + 0.01974003203213215, + 0.02130833826959133, + 0.01173296757042408, + 0.015289313159883022, + 0.006053285673260689, + 0.004740944132208824, + 0.0039344471879303455, + 0.004292310681194067, + 0.0061492109671235085, + 0.005101354327052832, + 0.003694477491080761, + 0.005289058666676283, + 0.003766563953831792, + 0.0038697959389537573, + 0.00436951732262969, + 0.00267836288549006, + 0.0034034072887152433, + 0.0033085444010794163, + 0.003065223339945078, + 0.002759407041594386, + 0.006727694999426603, + 0.006062766537070274, + 0.009733181446790695, + 0.0044460720382630825, + 0.008201302960515022, + 0.007521338295191526, + 0.006073895841836929, + 0.00661265617236495, + 0.00829224195331335, + 0.008143838495016098, + 0.008625824935734272, + 0.007476308848708868, + 0.011074764654040337, + 0.009813928045332432, + 0.008150143548846245, + 0.011425698176026344, + 0.017650088295340538, + 0.014590471982955933, + 0.015972603112459183, + 0.036411624401807785, + 0.0236351415514946, + 0.0699370875954628, + 0.02921224758028984, + 0.06444930285215378, + 0.008939920924603939, + 0.007864114828407764, + 0.0055413683876395226, + 0.007532911375164986, + 0.03033146634697914, + 0.011890151537954807, + 0.01289614662528038, + 0.012017833068966866, + 0.007540064863860607, + 0.008423417806625366, + 0.009793645702302456, + 0.010291097685694695, + 0.012797788716852665, + 0.01589459553360939, + 0.011114860884845257, + 0.009363673627376556, + 0.015650292858481407, + 0.013509350828826427, + 0.012090758420526981, + 0.016307657584547997, + 0.013157942332327366, + 0.020817754790186882, + 0.013974289409816265, + 0.01986854337155819, + 0.010053355246782303, + 0.015149284154176712, + 0.010587306693196297, + 0.021545635536313057, + 0.008659306913614273, + 0.014848567545413971, + 0.01111206691712141, + 0.016826849430799484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.20154841244220734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a sleek, jet-black coupe as it hurtles forward at breakneck speed on an empty highway shrouded in darkness. Its sleek metallic paintwork gleams with a cold sheen in the darkness, reflecting the scattered ambient light. Its sharp LED headlights cast two piercing, straight beams of white light, slicing the black asphalt ahead, slightly slippery from the light rain, clearly illuminating its rough texture and the gleaming water marks. As the vehicle speeds along, guardrails and blurred signage flash by in the beams. The wheels spin at high speed, kicking up a fine mist that becomes barely visible in the red taillights. The scene, dominated by deep blue and pure black, is punctuated by the brilliant white of the headlights and the crimson of the taillights, creating a cinematic atmosphere of intense speed, tension, and the loneliness and power of a solitary night traveler.", + "seed": 13330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_14/index_8041_seed_13330_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": -0.03621559231542051, + "vlm_logprob": -0.03621559231542051, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03621559231542051, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03777904063463211 + ], + "opposite_logprob_tokens": [ + -3.4127790927886963 + ], + "logprob_reward": -0.03777904063463211, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03777904063463211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060641441494226456, + 0.06593769788742065, + 0.07293198257684708, + 0.04018735513091087, + 0.04526004567742348, + 0.06142120435833931, + 0.28417232632637024, + 0.0996171161532402, + 0.20073530077934265, + 0.06909552216529846 + ], + "credit_map": [ + 0.007297431584447622, + 0.007330131251364946, + 0.0048434617929160595, + 0.005202411208301783, + 0.01136094518005848, + 0.008647209033370018, + 0.00506171677261591, + 0.010898139327764511, + 0.010252916254103184, + 0.007653202395886183, + 0.006319140084087849, + 0.004653288051486015, + 0.011749687604606152, + 0.012743811123073101, + 0.00447724899277091, + 0.00808840524405241, + 0.006627499125897884, + 0.006118413060903549, + 0.006000108551234007, + 0.012913115322589874, + 0.006722735706716776, + 0.006943467538803816, + 0.016975855454802513, + 0.010630791075527668, + 0.004578649532049894, + 0.005263826344162226, + 0.0038589034229516983, + 0.0029204366728663445, + 0.006217055022716522, + 0.00638657296076417, + 0.004418057855218649, + 0.006543853785842657, + 0.0061777508817613125, + 0.005414861720055342, + 0.006888578180223703, + 0.004412261303514242, + 0.004581565037369728, + 0.006114850752055645, + 0.005733616184443235, + 0.005936558824032545, + 0.009361205622553825, + 0.008494868874549866, + 0.005018331576138735, + 0.006311967968940735, + 0.010800954885780811, + 0.00810615997761488, + 0.006760014686733484, + 0.0065676989033818245, + 0.018223734572529793, + 0.04355717822909355, + 0.02006443776190281, + 0.025424176827073097, + 0.06341635435819626, + 0.04700417071580887, + 0.024376556277275085, + 0.042105723172426224, + 0.00817582756280899, + 0.010013710707426071, + 0.008320274762809277, + 0.01641811802983284, + 0.012109288945794106, + 0.012598009780049324, + 0.020127590745687485, + 0.011854302138090134, + 0.015381556935608387, + 0.04596499353647232, + 0.014816860668361187, + 0.031912997364997864, + 0.02377648465335369, + 0.02313925139605999, + 0.022393617779016495, + 0.023349547758698463, + 0.007784450426697731, + 0.00902143307030201, + 0.008412599563598633, + 0.009379993192851543, + 0.007418288849294186, + 0.009583832696080208, + 0.006597578525543213, + 0.010897349566221237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.03777904063463211 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01256337109953165 + ], + "opposite_logprob_tokens": [ + -4.387563228607178 + ], + "logprob_reward": -0.01256337109953165, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256337109953165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07310467213392258, + 0.09033460170030594, + 0.04500489681959152, + 0.02801833488047123, + 0.03237948939204216, + 0.05799957364797592, + 0.2951790392398834, + 0.23514018952846527, + 0.09924764931201935, + 0.043591566383838654 + ], + "credit_map": [ + 0.01197826024144888, + 0.009472211822867393, + 0.007408009376376867, + 0.004526402801275253, + 0.011205398477613926, + 0.011112614534795284, + 0.005871905945241451, + 0.011529866605997086, + 0.019548075273633003, + 0.010709083639085293, + 0.007716198917478323, + 0.004291328135877848, + 0.014089749194681644, + 0.017555832862854004, + 0.004185199271887541, + 0.012239133007824421, + 0.003582178382202983, + 0.0060936687514185905, + 0.00639577629044652, + 0.005589607637375593, + 0.007659723982214928, + 0.0048654633574187756, + 0.005919456481933594, + 0.004899022169411182, + 0.0037389169447124004, + 0.0037233890034258366, + 0.002560790628194809, + 0.002491343766450882, + 0.0044149113819003105, + 0.004472048487514257, + 0.002702124882489443, + 0.003914807923138142, + 0.004155628848820925, + 0.0037298649549484253, + 0.004559311084449291, + 0.0030290570575743914, + 0.004008897114545107, + 0.004260352812707424, + 0.004363253712654114, + 0.004273123573511839, + 0.007423541508615017, + 0.009240195155143738, + 0.003979608416557312, + 0.00521808909252286, + 0.011050395667552948, + 0.00799309741705656, + 0.005031139589846134, + 0.00806350726634264, + 0.02108001336455345, + 0.051194608211517334, + 0.020978644490242004, + 0.02593427337706089, + 0.06572646647691727, + 0.04665849357843399, + 0.02312248758971691, + 0.04048404097557068, + 0.01949414424598217, + 0.02348586730659008, + 0.021968044340610504, + 0.027465447783470154, + 0.03837393596768379, + 0.0331285297870636, + 0.034658100455999374, + 0.0365661159157753, + 0.008326346054673195, + 0.01055894698947668, + 0.006536956876516342, + 0.011624026112258434, + 0.013740958645939827, + 0.014777527190744877, + 0.011861701495945454, + 0.021821198984980583, + 0.0035639095585793257, + 0.005891127977520227, + 0.005541055463254452, + 0.008876015432178974, + 0.00478001544252038, + 0.0052497233264148235, + 0.004949046764522791, + 0.004740670323371887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.01256337109953165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036237069871276617 + ], + "opposite_logprob_tokens": [ + -5.6286234855651855 + ], + "logprob_reward": -0.0036237069871276617, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036237069871276617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11944600939750671, + 0.11213017255067825, + 0.12361034750938416, + 0.027748577296733856, + 0.0566442534327507, + 0.05989336967468262, + 0.14743182063102722, + 0.08092450350522995, + 0.18915049731731415, + 0.08302045613527298 + ], + "credit_map": [ + 0.016076717525720596, + 0.021221382543444633, + 0.010419108904898167, + 0.006914905738085508, + 0.023731259629130363, + 0.016061395406723022, + 0.01095825806260109, + 0.014062984846532345, + 0.015048282220959663, + 0.01706741377711296, + 0.006483846809715033, + 0.008355040103197098, + 0.018597323447465897, + 0.016858430579304695, + 0.009838211350142956, + 0.019881615415215492, + 0.014513771049678326, + 0.014731623232364655, + 0.014809984713792801, + 0.009746743366122246, + 0.017953120172023773, + 0.014272719621658325, + 0.008105507120490074, + 0.02947688288986683, + 0.003490364644676447, + 0.0034735973458737135, + 0.0027813855558633804, + 0.002418862422928214, + 0.0036993289832025766, + 0.004115134011954069, + 0.004146147519350052, + 0.0036237563472241163, + 0.005575801711529493, + 0.008199495263397694, + 0.006987364497035742, + 0.005045859608799219, + 0.006393021903932095, + 0.007935718633234501, + 0.007172069512307644, + 0.009334922768175602, + 0.0046579656191170216, + 0.008701320737600327, + 0.004045071080327034, + 0.01115204393863678, + 0.009598592296242714, + 0.00607214355841279, + 0.010257188230752945, + 0.005409040488302708, + 0.011675862595438957, + 0.025835013017058372, + 0.010321195237338543, + 0.011272210627794266, + 0.03563721105456352, + 0.021284088492393494, + 0.014166769571602345, + 0.01723947748541832, + 0.010997932404279709, + 0.00801530759781599, + 0.00593654764816165, + 0.009692743420600891, + 0.012968168593943119, + 0.011580157093703747, + 0.010475996881723404, + 0.011257650330662727, + 0.010320968925952911, + 0.018313493579626083, + 0.00970141775906086, + 0.019401386380195618, + 0.02993360534310341, + 0.04638822376728058, + 0.026162799447774887, + 0.028928589075803757, + 0.010796434246003628, + 0.011056046932935715, + 0.009474734775722027, + 0.00936430785804987, + 0.008343804627656937, + 0.009544817730784416, + 0.009634041227400303, + 0.014806266874074936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.0036237069871276617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194829389452934 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014194829389452934, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194829389452934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07117223739624023, + 0.16204933822155, + 0.12278154492378235, + 0.03636578470468521, + 0.055344775319099426, + 0.05449177324771881, + 0.19821593165397644, + 0.06693856418132782, + 0.17706803977489471, + 0.055571991950273514 + ], + "credit_map": [ + 0.010263247415423393, + 0.00814038421958685, + 0.007199276238679886, + 0.005952909123152494, + 0.009794789366424084, + 0.00904029794037342, + 0.007629510015249252, + 0.013151822611689568, + 0.02729611285030842, + 0.024556754156947136, + 0.00990944541990757, + 0.009592616930603981, + 0.03548555448651314, + 0.025939788669347763, + 0.010724162682890892, + 0.018544897437095642, + 0.010285558179020882, + 0.011280501261353493, + 0.011116649024188519, + 0.020078159868717194, + 0.014719786122441292, + 0.011176512576639652, + 0.023103615269064903, + 0.021020766347646713, + 0.003828752553090453, + 0.005481517408043146, + 0.0035510023590177298, + 0.0026454636827111244, + 0.006562836933881044, + 0.00539885088801384, + 0.002614411525428295, + 0.006282947491854429, + 0.00469358591362834, + 0.005813096184283495, + 0.004645111504942179, + 0.004639370366930962, + 0.007020830176770687, + 0.00881928764283657, + 0.005276416894048452, + 0.014437077566981316, + 0.005635207984596491, + 0.007048634812235832, + 0.003304866375401616, + 0.00396460946649313, + 0.009044281207025051, + 0.00864661019295454, + 0.004840858746320009, + 0.012006706558167934, + 0.014789030887186527, + 0.032765548676252365, + 0.00930322427302599, + 0.015761125832796097, + 0.045269157737493515, + 0.029691116884350777, + 0.017371194437146187, + 0.03326552361249924, + 0.008825703524053097, + 0.006529953330755234, + 0.005204635672271252, + 0.005832860246300697, + 0.01300040353089571, + 0.009367835707962513, + 0.009590789675712585, + 0.008586377836763859, + 0.00960150733590126, + 0.027183664962649345, + 0.012721933424472809, + 0.01890946365892887, + 0.02154272235929966, + 0.027332833036780357, + 0.02760433219373226, + 0.0321715846657753, + 0.006638594903051853, + 0.007525975350290537, + 0.006904906593263149, + 0.009341630153357983, + 0.004769781604409218, + 0.006422423757612705, + 0.006679023150354624, + 0.007289654575288296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.014194829389452934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291701346635818 + ], + "opposite_logprob_tokens": [ + -2.237916946411133 + ], + "logprob_reward": -0.11291701346635818, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291701346635818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06484266370534897, + 0.1370401531457901, + 0.04070434346795082, + 0.027366645634174347, + 0.021259887143969536, + 0.022248659282922745, + 0.16031542420387268, + 0.11755261570215225, + 0.3552158772945404, + 0.05345366522669792 + ], + "credit_map": [ + 0.012100683525204659, + 0.006199730560183525, + 0.00658771488815546, + 0.0054835365153849125, + 0.007881850004196167, + 0.009708832018077374, + 0.006201216019690037, + 0.010679109022021294, + 0.021491624414920807, + 0.019387533888220787, + 0.009647995233535767, + 0.010082392953336239, + 0.028661269694566727, + 0.02109035663306713, + 0.010560799390077591, + 0.01611819490790367, + 0.005540130659937859, + 0.0048348987475037575, + 0.004075230564922094, + 0.002746089594438672, + 0.004936615936458111, + 0.00532522052526474, + 0.003812742419540882, + 0.00943342037498951, + 0.003352023893967271, + 0.0028510515112429857, + 0.0043478659354150295, + 0.001999033847823739, + 0.0038394706789404154, + 0.00465876841917634, + 0.0022440229076892138, + 0.004074412398040295, + 0.002550700679421425, + 0.0018529508961364627, + 0.00204956391826272, + 0.002809318946674466, + 0.0020207292400300503, + 0.003393780440092087, + 0.0031276473309844732, + 0.0034551979042589664, + 0.002617859048768878, + 0.0032852960284799337, + 0.0016894644359126687, + 0.002841157605871558, + 0.003926256205886602, + 0.002815334126353264, + 0.0024696856271475554, + 0.0026036102790385485, + 0.01189053151756525, + 0.028325047343969345, + 0.010644862428307533, + 0.012144770473241806, + 0.03572911024093628, + 0.024401964619755745, + 0.013234135694801807, + 0.023945016786456108, + 0.010425636544823647, + 0.010855719447135925, + 0.008895204402506351, + 0.011991450563073158, + 0.025423116981983185, + 0.01811235584318638, + 0.019531991332769394, + 0.012317163869738579, + 0.026127247139811516, + 0.037526797503232956, + 0.03074236772954464, + 0.044511523097753525, + 0.04299061745405197, + 0.052337951958179474, + 0.04189380258321762, + 0.07908561080694199, + 0.007904074154794216, + 0.007078295573592186, + 0.007223064545542002, + 0.007388255558907986, + 0.004688248969614506, + 0.006371496245265007, + 0.004745065234601498, + 0.008055170066654682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.11291701346635818 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a sleek, jet-black coupe as it hurtles forward at breakneck speed on an empty highway shrouded in darkness. Its sleek metallic paintwork gleams with a cold sheen in the darkness, reflecting the scattered ambient light. Its sharp LED headlights cast two piercing, straight beams of white light, slicing the black asphalt ahead, slightly slippery from the light rain, clearly illuminating its rough texture and the gleaming water marks. As the vehicle speeds along, guardrails and blurred signage flash by in the beams. The wheels spin at high speed, kicking up a fine mist that becomes barely visible in the red taillights. The scene, dominated by deep blue and pure black, is punctuated by the brilliant white of the headlights and the crimson of the taillights, creating a cinematic atmosphere of intense speed, tension, and the loneliness and power of a solitary night traveler.", + "seed": 13331, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_14/index_8041_seed_13331_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": -0.039350662287324666, + "vlm_logprob": -0.039350662287324666, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.039350662287324666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05388517677783966 + ], + "opposite_logprob_tokens": [ + -3.0538852214813232 + ], + "logprob_reward": -0.05388517677783966, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05388517677783966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14507947862148285, + 0.04665853828191757, + 0.04773946851491928, + 0.037866391241550446, + 0.03921784833073616, + 0.07117966562509537, + 0.07114555686712265, + 0.08019624650478363, + 0.24397829174995422, + 0.21693849563598633 + ], + "credit_map": [ + 0.01687796413898468, + 0.01474742405116558, + 0.014496640302240849, + 0.010668562725186348, + 0.028778528794646263, + 0.03241240605711937, + 0.013318967074155807, + 0.013778985477983952, + 0.004359171260148287, + 0.0056579275988042355, + 0.006190489511936903, + 0.00461413012817502, + 0.009110148064792156, + 0.006215579342097044, + 0.005029106978327036, + 0.005481986328959465, + 0.005657294765114784, + 0.004583450965583324, + 0.007425464689731598, + 0.0038607793394476175, + 0.005608858074992895, + 0.008243290707468987, + 0.004766334779560566, + 0.007593992166221142, + 0.003612673841416836, + 0.004491453990340233, + 0.003915245179086924, + 0.003637137357145548, + 0.0067997220903635025, + 0.009287852793931961, + 0.0027713957242667675, + 0.003350907238200307, + 0.0039861868135631084, + 0.003733048914000392, + 0.00391136622056365, + 0.0036816196516156197, + 0.0053869313560426235, + 0.007012985646724701, + 0.0037841435987502337, + 0.007721565663814545, + 0.008917946368455887, + 0.005199559964239597, + 0.008538490161299706, + 0.01188473030924797, + 0.006786765996366739, + 0.014195606112480164, + 0.005560666788369417, + 0.010095896199345589, + 0.0076525830663740635, + 0.010023790411651134, + 0.010003723204135895, + 0.010072835721075535, + 0.007359254639595747, + 0.010333527810871601, + 0.005991908255964518, + 0.009707938879728317, + 0.008589057251811028, + 0.01019231602549553, + 0.007692277897149324, + 0.00866383034735918, + 0.00820150226354599, + 0.01749897375702858, + 0.00821344181895256, + 0.011144853197038174, + 0.03312196210026741, + 0.03207146003842354, + 0.03011210449039936, + 0.04264027625322342, + 0.025716781616210938, + 0.024056417867541313, + 0.0298991110175848, + 0.026360157877206802, + 0.01562836766242981, + 0.025198183953762054, + 0.03771018236875534, + 0.027333935722708702, + 0.03286606818437576, + 0.017803408205509186, + 0.04246942698955536, + 0.017928924411535263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.05388517677783966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014213752001523972 + ], + "opposite_logprob_tokens": [ + -4.264213562011719 + ], + "logprob_reward": -0.014213752001523972, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014213752001523972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2454456388950348, + 0.08217933028936386, + 0.03553927689790726, + 0.04145260527729988, + 0.06255101412534714, + 0.06274078786373138, + 0.10650336742401123, + 0.1209055706858635, + 0.09218937903642654, + 0.15049295127391815 + ], + "credit_map": [ + 0.01690080389380455, + 0.027999328449368477, + 0.02203136682510376, + 0.014028133824467659, + 0.07006203383207321, + 0.05470039322972298, + 0.018214333802461624, + 0.021509265527129173, + 0.008627799339592457, + 0.009804137982428074, + 0.01222284696996212, + 0.005907815415412188, + 0.010446273721754551, + 0.018132874742150307, + 0.006247914861887693, + 0.010789670050144196, + 0.00291440193541348, + 0.004109588917344809, + 0.0085086515173316, + 0.002661795588210225, + 0.005207017995417118, + 0.004715269897133112, + 0.0029958009254187346, + 0.004426753148436546, + 0.004008536692708731, + 0.005738473031669855, + 0.005463296547532082, + 0.003867992665618658, + 0.0070616635493934155, + 0.005817892495542765, + 0.003457806771621108, + 0.006036944687366486, + 0.007148109842091799, + 0.005863603670150042, + 0.004438905045390129, + 0.005328667350113392, + 0.010219111107289791, + 0.014864601194858551, + 0.00430723000317812, + 0.010380798019468784, + 0.005215991288423538, + 0.006409432739019394, + 0.007968542166054249, + 0.011457549408078194, + 0.008832893334329128, + 0.006816760636866093, + 0.008558809757232666, + 0.007480816449970007, + 0.008535736240446568, + 0.012052284553647041, + 0.014198917895555496, + 0.014823255129158497, + 0.011313243769109249, + 0.01353307906538248, + 0.013073326088488102, + 0.01897353306412697, + 0.01379566453397274, + 0.01478463876992464, + 0.012037201784551144, + 0.025040453299880028, + 0.011142502538859844, + 0.015638839453458786, + 0.013839920982718468, + 0.014626359567046165, + 0.009859688580036163, + 0.01131290104240179, + 0.01039126142859459, + 0.019675791263580322, + 0.010708743706345558, + 0.009792839176952839, + 0.01220597606152296, + 0.008242190815508366, + 0.010767229832708836, + 0.015130698680877686, + 0.024355262517929077, + 0.03448330610990524, + 0.0124204708263278, + 0.014733564108610153, + 0.024620501324534416, + 0.013981932774186134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.014213752001523972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954266060143709 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954266060143709, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954266060143709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17288266122341156, + 0.12309831380844116, + 0.05184740945696831, + 0.12958772480487823, + 0.10285283625125885, + 0.07161270827054977, + 0.05179654061794281, + 0.09315194934606552, + 0.06612223386764526, + 0.1370476931333542 + ], + "credit_map": [ + 0.02094140090048313, + 0.015883464366197586, + 0.011992593295872211, + 0.010113677941262722, + 0.04796924069523811, + 0.03447016328573227, + 0.01597043126821518, + 0.015541665256023407, + 0.01118861511349678, + 0.014470396563410759, + 0.019653085619211197, + 0.00860928650945425, + 0.01665305159986019, + 0.024606622755527496, + 0.013028347864747047, + 0.014888901263475418, + 0.005662885960191488, + 0.004875150043517351, + 0.010152592323720455, + 0.0035645810421556234, + 0.006289774551987648, + 0.01061621867120266, + 0.0045327250845730305, + 0.00615347595885396, + 0.011305264197289944, + 0.011576532386243343, + 0.013808383606374264, + 0.009479748085141182, + 0.02903926745057106, + 0.02597038820385933, + 0.011980198323726654, + 0.01642792299389839, + 0.011000603437423706, + 0.013953685760498047, + 0.008367537520825863, + 0.014325815252959728, + 0.013027554377913475, + 0.017364555969834328, + 0.011810568161308765, + 0.013002514839172363, + 0.008277950808405876, + 0.007475711405277252, + 0.009074892848730087, + 0.014321930706501007, + 0.0057939826510846615, + 0.01186759490519762, + 0.0085136191919446, + 0.006287024822086096, + 0.006951058283448219, + 0.006984473206102848, + 0.007571034133434296, + 0.005569059401750565, + 0.005814907606691122, + 0.007827280089259148, + 0.003991370555013418, + 0.007087357807904482, + 0.007769748102873564, + 0.012069134041666985, + 0.008290010504424572, + 0.017382239922881126, + 0.008245113305747509, + 0.013336218893527985, + 0.013296123594045639, + 0.012763352133333683, + 0.0076514133252203465, + 0.010046404786407948, + 0.00781535729765892, + 0.009902426041662693, + 0.008188340812921524, + 0.00799605343490839, + 0.006499280221760273, + 0.008022954687476158, + 0.010345209389925003, + 0.012960252352058887, + 0.019724678248167038, + 0.03384832292795181, + 0.011740771122276783, + 0.01287087332457304, + 0.023870665580034256, + 0.011686911806464195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.005954266060143709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780227206647396 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780227206647396, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780227206647396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29735067486763, + 0.07357621937990189, + 0.03388931602239609, + 0.06529337167739868, + 0.046188920736312866, + 0.09092557430267334, + 0.05222845822572708, + 0.16892565786838531, + 0.11978393793106079, + 0.05183783546090126 + ], + "credit_map": [ + 0.024781132116913795, + 0.03090369515120983, + 0.014437548816204071, + 0.012253456749022007, + 0.07171732932329178, + 0.0783579871058464, + 0.020409436896443367, + 0.044490087777376175, + 0.007095701526850462, + 0.009635563008487225, + 0.010337545536458492, + 0.0058752805925905704, + 0.009382342919707298, + 0.010771915316581726, + 0.009121027775108814, + 0.011356844566762447, + 0.0038019963540136814, + 0.003677107859402895, + 0.00500616105273366, + 0.002921651815995574, + 0.0040404279716312885, + 0.004846130032092333, + 0.0036836876533925533, + 0.005912153050303459, + 0.006488497834652662, + 0.008405515924096107, + 0.009436075575649738, + 0.004275748506188393, + 0.010045178234577179, + 0.01050921157002449, + 0.004458676557987928, + 0.011674465611577034, + 0.005232455674558878, + 0.0038193005602806807, + 0.005312563385814428, + 0.00440227473154664, + 0.00559280626475811, + 0.009815146215260029, + 0.004902272950857878, + 0.007112096529453993, + 0.01155722513794899, + 0.007086937315762043, + 0.010537606664001942, + 0.011488035321235657, + 0.007510554976761341, + 0.017167573794722557, + 0.008123454637825489, + 0.01745419017970562, + 0.00517453346401453, + 0.004579364322125912, + 0.008238685317337513, + 0.006902043242007494, + 0.005499993450939655, + 0.008038029074668884, + 0.00479465164244175, + 0.00900116004049778, + 0.018251121044158936, + 0.016486210748553276, + 0.015220318920910358, + 0.026729770004749298, + 0.017707515507936478, + 0.033874236047267914, + 0.017870066687464714, + 0.022786419838666916, + 0.011687097139656544, + 0.012570736929774284, + 0.014321524649858475, + 0.018383080139756203, + 0.015386505052447319, + 0.0180087611079216, + 0.01280590146780014, + 0.0166203323751688, + 0.004974621348083019, + 0.004081304185092449, + 0.006670593284070492, + 0.010065102949738503, + 0.006319985259324312, + 0.007694134954363108, + 0.006493059452623129, + 0.005539035424590111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.009780227206647396 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, jet-black coupe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the car's sharp LED headlights casting two piercing beams of white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an empty highway shrouded in darkness with light rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's wheels spin at high speed, kicking up a fine mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129198893904686 + ], + "opposite_logprob_tokens": [ + -2.237919807434082 + ], + "logprob_reward": -0.1129198893904686, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129198893904686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18991118669509888, + 0.0788997933268547, + 0.06344939023256302, + 0.10241658985614777, + 0.04240363463759422, + 0.08231624960899353, + 0.07443750649690628, + 0.08164512366056442, + 0.11685401201248169, + 0.16766658425331116 + ], + "credit_map": [ + 0.015059561468660831, + 0.025893032550811768, + 0.015512294135987759, + 0.014853070490062237, + 0.04076987877488136, + 0.03936557099223137, + 0.018955405801534653, + 0.019502365961670876, + 0.008925158530473709, + 0.013289004564285278, + 0.008731037378311157, + 0.008112682029604912, + 0.011723743751645088, + 0.011330465786159039, + 0.00862827803939581, + 0.008159429766237736, + 0.007610786706209183, + 0.00721770478412509, + 0.009624306112527847, + 0.005540615878999233, + 0.007922717370092869, + 0.009014840237796307, + 0.006063133478164673, + 0.010455287992954254, + 0.010102803818881512, + 0.012879889458417892, + 0.013479749672114849, + 0.007280353922396898, + 0.01981988549232483, + 0.025815485045313835, + 0.00540949497371912, + 0.007628922816365957, + 0.006470512133091688, + 0.004414451774209738, + 0.006380872800946236, + 0.003102516755461693, + 0.004638167098164558, + 0.006836943794041872, + 0.0029244692996144295, + 0.007635701447725296, + 0.009531576186418533, + 0.007662403862923384, + 0.013165928423404694, + 0.010457088239490986, + 0.009477877989411354, + 0.013400916941463947, + 0.005948793608695269, + 0.012671664357185364, + 0.006159653887152672, + 0.007045719772577286, + 0.009117839857935905, + 0.006734796799719334, + 0.008119231089949608, + 0.01124944631010294, + 0.00785774365067482, + 0.018153073266148567, + 0.007919840514659882, + 0.007150221150368452, + 0.0076552038080990314, + 0.012025550939142704, + 0.007726190611720085, + 0.020349640399217606, + 0.009366820566356182, + 0.009451660327613354, + 0.014672176912426949, + 0.010418266989290714, + 0.012422467581927776, + 0.024158861488103867, + 0.011282320134341717, + 0.013623753562569618, + 0.017126059159636497, + 0.013150101527571678, + 0.01092361006885767, + 0.015617512166500092, + 0.03375617787241936, + 0.03337781876325607, + 0.011939095333218575, + 0.017652420327067375, + 0.034070808440446854, + 0.010329139418900013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.1129198893904686 + } + ], + "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/0000028/group_15/index_tensor([9047, 9047])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_15/index_tensor([9047, 9047])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2eaac0da4b853de14a147e162460e17202857d23 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_15/index_tensor([9047, 9047])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera slowly and steadily approaches a man standing in a dimly lit, luxurious room. He wears a perfectly tailored dark charcoal wool suit jacket over a crisp white dress shirt. A single, dramatic key light from the side sculpts his form, casting deep shadows and creating a powerful black-white contrast, highlighting the sharp line of his jaw and his focused, determined expression. His dark hair is neatly styled. As the camera advances, his hands rise with deliberate grace, resting on his neck. He carefully straightens the collar of his starched white shirt, his fingertips making tiny, precise adjustments to ensure perfect symmetry. The intense light captures the fine twill weave of the shirt fabric and the subtle texture of his skin, creating a tactile visual experience. The background remains a soft, dark blur, ensuring that all attention is focused on this ritual of preparation. The camera's slow, relentless push culminates in extreme close-up, filling the frame with the texture of the collar and the steady, controlled pressure of his fingers, creating an atmosphere of focused attention and quiet, powerful confidence.", + "seed": 7617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_15/index_9047_seed_7617_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": -0.01936327322036959, + "vlm_logprob": -0.01936327322036959, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.01936327322036959, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025263119488954544 + ], + "opposite_logprob_tokens": [ + -3.9002630710601807 + ], + "logprob_reward": -0.025263119488954544, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025263119488954544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03743290901184082, + 0.013023082166910172, + 0.012709117494523525, + 0.021026335656642914, + 0.01578308641910553, + 0.11340375989675522, + 0.15638341009616852, + 0.0351455919444561, + 0.3906826078891754, + 0.2044101059436798 + ], + "credit_map": [ + 0.007675246801227331, + 0.003681972622871399, + 0.00565474946051836, + 0.003779228776693344, + 0.0033966624177992344, + 0.0042642648331820965, + 0.0031950846314430237, + 0.0057856980711221695, + 0.0013681167038157582, + 0.0018571303226053715, + 0.0017161931609734893, + 0.0015723361866548657, + 0.0015761468093842268, + 0.0013582211686298251, + 0.001739163650199771, + 0.0018357746303081512, + 0.0014747674576938152, + 0.001422386383637786, + 0.001495277974754572, + 0.0016030247788876295, + 0.001866897800937295, + 0.0014653555117547512, + 0.0017176744295284152, + 0.0016637323424220085, + 0.0023514050990343094, + 0.0018643744988366961, + 0.003016700269654393, + 0.0026875780895352364, + 0.0026709754019975662, + 0.0028868727385997772, + 0.00304150371812284, + 0.002506926190108061, + 0.001921140355989337, + 0.0019942240323871374, + 0.0025113460142165422, + 0.001599809154868126, + 0.0018415633821859956, + 0.0022199121303856373, + 0.0017469162121415138, + 0.0019481745548546314, + 0.008546754717826843, + 0.006147523410618305, + 0.02453181892633438, + 0.0092107392847538, + 0.010202251374721527, + 0.013927556574344635, + 0.025568896904587746, + 0.015268217772245407, + 0.015218837186694145, + 0.013026158325374126, + 0.023916762322187424, + 0.01529016625136137, + 0.016589423641562462, + 0.02200138196349144, + 0.028193505480885506, + 0.02214718423783779, + 0.0020876582711935043, + 0.0030535662081092596, + 0.004846017807722092, + 0.002939520636573434, + 0.00880039669573307, + 0.003367014229297638, + 0.0033320451620966196, + 0.006719372235238552, + 0.02756650745868683, + 0.014346044510602951, + 0.023933764547109604, + 0.007430199068039656, + 0.1457921266555786, + 0.03710317239165306, + 0.0166179072111845, + 0.11789287626743317, + 0.018682224676012993, + 0.009401002898812294, + 0.011936879716813564, + 0.006352818105369806, + 0.07258439064025879, + 0.024021264165639877, + 0.013658812269568443, + 0.04777270928025246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.025263119488954544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821231961250305 + ], + "opposite_logprob_tokens": [ + -3.5298211574554443 + ], + "logprob_reward": -0.029821231961250305, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821231961250305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03477764502167702, + 0.017423266544938087, + 0.03081919252872467, + 0.060320209711790085, + 0.0465611107647419, + 0.09995053708553314, + 0.05512358248233795, + 0.13840535283088684, + 0.30189356207847595, + 0.21472550928592682 + ], + "credit_map": [ + 0.004636091645807028, + 0.004114658106118441, + 0.00481424992904067, + 0.00424477644264698, + 0.003802413819357753, + 0.0038323616608977318, + 0.0038764113560318947, + 0.005456680431962013, + 0.001609133556485176, + 0.0023485177662223577, + 0.0012703813845291734, + 0.0022328069899231195, + 0.001972673460841179, + 0.003088961588218808, + 0.0022418727166950703, + 0.0026589201297611, + 0.002504176925867796, + 0.003416267456486821, + 0.002483411692082882, + 0.004068211652338505, + 0.00487253675237298, + 0.003455941565334797, + 0.0056594302877783775, + 0.004359216894954443, + 0.0055050114169716835, + 0.005262499209493399, + 0.008847260847687721, + 0.006563856266438961, + 0.0063221994787454605, + 0.00738035561516881, + 0.012815670110285282, + 0.007623352576047182, + 0.003866745624691248, + 0.0035543465055525303, + 0.008923972956836224, + 0.006102412939071655, + 0.004417561460286379, + 0.006753366440534592, + 0.007295489776879549, + 0.005647218320518732, + 0.00967480055987835, + 0.010519450530409813, + 0.01268499344587326, + 0.011418427340686321, + 0.015231060795485973, + 0.011457021348178387, + 0.013950126245617867, + 0.015014653094112873, + 0.005551411770284176, + 0.0054198154248297215, + 0.007576583884656429, + 0.006830538157373667, + 0.00839927140623331, + 0.006847321055829525, + 0.00714134331792593, + 0.007357301190495491, + 0.007015960291028023, + 0.007652861066162586, + 0.008999713696539402, + 0.007973849773406982, + 0.04558900371193886, + 0.014963901601731777, + 0.012234233319759369, + 0.03397582471370697, + 0.022513549774885178, + 0.024005118757486343, + 0.04069187492132187, + 0.013153493404388428, + 0.08268441259860992, + 0.030584054067730904, + 0.020843664184212685, + 0.06741734594106674, + 0.01971025951206684, + 0.009854135103523731, + 0.009875655174255371, + 0.005750665906816721, + 0.07831308245658875, + 0.025841867551207542, + 0.013105630874633789, + 0.05227421596646309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.029821231961250305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033712390810251236 + ], + "opposite_logprob_tokens": [ + -3.408712387084961 + ], + "logprob_reward": -0.033712390810251236, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033712390810251236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18395182490348816, + 0.030244074761867523, + 0.04283023253083229, + 0.046844661235809326, + 0.06730122119188309, + 0.05533798411488533, + 0.04019744321703911, + 0.031431958079338074, + 0.3421858847141266, + 0.1596747189760208 + ], + "credit_map": [ + 0.03357602655887604, + 0.02315525896847248, + 0.02520628832280636, + 0.020106913521885872, + 0.016233140602707863, + 0.01955413818359375, + 0.01589825563132763, + 0.03022179752588272, + 0.0039219181053340435, + 0.0036766917910426855, + 0.004730775952339172, + 0.004846200812608004, + 0.002937917597591877, + 0.003100855741649866, + 0.0027647523675113916, + 0.004264962859451771, + 0.004354397300630808, + 0.0043046255595982075, + 0.0047185118310153484, + 0.005317611154168844, + 0.0054339258931577206, + 0.006440937519073486, + 0.006227914243936539, + 0.0060323080979287624, + 0.004365904256701469, + 0.005689479410648346, + 0.004850689321756363, + 0.00538049777969718, + 0.006804824341088533, + 0.005398055072873831, + 0.008570811711251736, + 0.005784396547824144, + 0.007036012597382069, + 0.005163844674825668, + 0.015211429446935654, + 0.007854578085243702, + 0.0049638464115560055, + 0.008569657802581787, + 0.00967534352093935, + 0.008826506324112415, + 0.0034689554013311863, + 0.004317137412726879, + 0.005757556762546301, + 0.005328460596501827, + 0.007848850451409817, + 0.006248886231333017, + 0.015090908855199814, + 0.007277228403836489, + 0.003849049797281623, + 0.004492522682994604, + 0.0073910863138735294, + 0.0053366394713521, + 0.004415631294250488, + 0.004651648458093405, + 0.005494122393429279, + 0.0045667411759495735, + 0.0029398323968052864, + 0.0028989261481910944, + 0.004199361894279718, + 0.003484301036223769, + 0.004161443095654249, + 0.0042656403966248035, + 0.00481909466907382, + 0.004663357976824045, + 0.02495717629790306, + 0.013581370934844017, + 0.012635158374905586, + 0.006660711020231247, + 0.129344642162323, + 0.03234095498919487, + 0.01752905547618866, + 0.10513681918382645, + 0.014075024984776974, + 0.0073273926973342896, + 0.010302471928298473, + 0.005272761452943087, + 0.05500601977109909, + 0.018595878034830093, + 0.011073751375079155, + 0.03802142292261124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.033712390810251236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631315384060144 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631315384060144, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631315384060144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009150108322501183, + 0.006962159648537636, + 0.006759507581591606, + 0.007701701018959284, + 0.008081494830548763, + 0.017554009333252907, + 0.009851235896348953, + 0.037696097046136856, + 0.817933976650238, + 0.07830966264009476 + ], + "credit_map": [ + 0.001726734102703631, + 0.0008335908059962094, + 0.001312306267209351, + 0.001106135779991746, + 0.0008718330063857138, + 0.000902302039321512, + 0.0010034083388745785, + 0.0013937983894720674, + 0.0007391106337308884, + 0.0007149681332521141, + 0.0005728575633838773, + 0.001025696750730276, + 0.0008554595406167209, + 0.0011542676948010921, + 0.0008852903265506029, + 0.0010145094711333513, + 0.0005647040088661015, + 0.0008641658932901919, + 0.0007811295217834413, + 0.000792275823187083, + 0.0007973399478942156, + 0.0009766904404386878, + 0.0011286137159913778, + 0.00085458904504776, + 0.0006470552762039006, + 0.0005486221634782851, + 0.0007193612982518971, + 0.0007324455655179918, + 0.0012403539149090648, + 0.0009727886645123363, + 0.0019514444284141064, + 0.0008896299987100065, + 0.000752622785512358, + 0.0004400976176839322, + 0.0018602414056658745, + 0.0007353682303801179, + 0.0008023728732950985, + 0.00130061327945441, + 0.0007925175596028566, + 0.0013976615155115724, + 0.0013080878416076303, + 0.0010156083153560758, + 0.002203558571636677, + 0.0014418860664591193, + 0.002235283376649022, + 0.002202568808570504, + 0.004682280123233795, + 0.0024647372774779797, + 0.0008815707988105714, + 0.0008273513521999121, + 0.002193810185417533, + 0.0011832070304080844, + 0.0011246667709201574, + 0.0013679254334419966, + 0.0008619611035101116, + 0.0014107439201325178, + 0.003457623301073909, + 0.002384849591180682, + 0.004031575750559568, + 0.003635821398347616, + 0.005704482086002827, + 0.005037711933255196, + 0.006402305793017149, + 0.007041728589683771, + 0.054133519530296326, + 0.028912335634231567, + 0.028886234387755394, + 0.014716928824782372, + 0.3174605667591095, + 0.07678049802780151, + 0.03561268746852875, + 0.2614312469959259, + 0.006936477031558752, + 0.005164913833141327, + 0.0061873518861830235, + 0.003437694162130356, + 0.023421769961714745, + 0.008965624496340752, + 0.007875907234847546, + 0.01631992682814598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.007631315384060144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003883084573317319 + ], + "opposite_logprob_tokens": [ + -7.875388145446777 + ], + "logprob_reward": -0.0003883084573317319, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003883084573317319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09085961431264877, + 0.03927784040570259, + 0.06468277424573898, + 0.021594969555735588, + 0.038237836211919785, + 0.08672469109296799, + 0.0400216281414032, + 0.027297573164105415, + 0.3417542278766632, + 0.24954892694950104 + ], + "credit_map": [ + 0.01936248689889908, + 0.011233204044401646, + 0.011215964332222939, + 0.011787451803684235, + 0.006488266866654158, + 0.007979091256856918, + 0.007743720430880785, + 0.015049419365823269, + 0.005208880640566349, + 0.005731602665036917, + 0.006100068800151348, + 0.005151590798050165, + 0.004131385125219822, + 0.003948525991290808, + 0.0038836572784930468, + 0.00512212747707963, + 0.007204323075711727, + 0.010142937302589417, + 0.00718686543405056, + 0.008464002050459385, + 0.009425419382750988, + 0.006771486718207598, + 0.006632698234170675, + 0.008855033665895462, + 0.002523115137591958, + 0.0027651088312268257, + 0.002574746496975422, + 0.0022717283573001623, + 0.0028442542534321547, + 0.002758478280156851, + 0.003098284127190709, + 0.0027592508122324944, + 0.004372551571577787, + 0.0035035524051636457, + 0.007787254638969898, + 0.003910047933459282, + 0.004170502070337534, + 0.0050293090753257275, + 0.00483932439237833, + 0.004625292960554361, + 0.00840345025062561, + 0.007836384698748589, + 0.012639944441616535, + 0.008986600674688816, + 0.010426292195916176, + 0.009138204157352448, + 0.01789189875125885, + 0.011401908472180367, + 0.0038112730253487825, + 0.003381325863301754, + 0.008408524096012115, + 0.004127406980842352, + 0.004659736528992653, + 0.004807556048035622, + 0.0051742675714194775, + 0.005651530344039202, + 0.0033984582405537367, + 0.0018951162928715348, + 0.0047268555499613285, + 0.002227321034297347, + 0.0033979881554841995, + 0.00338366930373013, + 0.004986329935491085, + 0.003281831042841077, + 0.02038850635290146, + 0.01320425234735012, + 0.01560772955417633, + 0.007765130139887333, + 0.12900418043136597, + 0.03272222727537155, + 0.0167575404047966, + 0.1063045933842659, + 0.022295977920293808, + 0.01111879013478756, + 0.017479844391345978, + 0.008055382408201694, + 0.0832493007183075, + 0.028610751032829285, + 0.021312033757567406, + 0.057426806539297104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0003883084573317319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily approaches a man standing in a dimly lit, luxurious room. He wears a perfectly tailored dark charcoal wool suit jacket over a crisp white dress shirt. A single, dramatic key light from the side sculpts his form, casting deep shadows and creating a powerful black-white contrast, highlighting the sharp line of his jaw and his focused, determined expression. His dark hair is neatly styled. As the camera advances, his hands rise with deliberate grace, resting on his neck. He carefully straightens the collar of his starched white shirt, his fingertips making tiny, precise adjustments to ensure perfect symmetry. The intense light captures the fine twill weave of the shirt fabric and the subtle texture of his skin, creating a tactile visual experience. The background remains a soft, dark blur, ensuring that all attention is focused on this ritual of preparation. The camera's slow, relentless push culminates in extreme close-up, filling the frame with the texture of the collar and the steady, controlled pressure of his fingers, creating an atmosphere of focused attention and quiet, powerful confidence.", + "seed": 7618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_15/index_9047_seed_7618_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": -0.018777139644953422, + "vlm_logprob": -0.018777139644953422, + "vlm_yes_no_margin": 4.899999952316284, + "vlm_token_credit": -0.018777139644953422, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024745620787143707 + ], + "opposite_logprob_tokens": [ + -4.149745464324951 + ], + "logprob_reward": -0.024745620787143707, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.024745620787143707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12299246340990067, + 0.09228350222110748, + 0.0721631869673729, + 0.0720997154712677, + 0.07226287573575974, + 0.09639047831296921, + 0.1132764145731926, + 0.189628005027771, + 0.09886211156845093, + 0.070041224360466 + ], + "credit_map": [ + 0.006895930040627718, + 0.010745862498879433, + 0.007074242923408747, + 0.015745583921670914, + 0.02209661342203617, + 0.019829850643873215, + 0.024949537590146065, + 0.015654858201742172, + 0.009594418108463287, + 0.012949623167514801, + 0.009272514842450619, + 0.009654782712459564, + 0.013652633875608444, + 0.01612710952758789, + 0.01119939237833023, + 0.009833037853240967, + 0.0075340778566896915, + 0.009173459373414516, + 0.007385882083326578, + 0.00684735830873251, + 0.007471503224223852, + 0.012869301252067089, + 0.008859149180352688, + 0.012022465467453003, + 0.007188749499619007, + 0.008356873877346516, + 0.006409076973795891, + 0.006971892435103655, + 0.008208955638110638, + 0.01152283325791359, + 0.008903922513127327, + 0.014537420123815536, + 0.0057166386395692825, + 0.007112821098417044, + 0.008066433481872082, + 0.010004010982811451, + 0.009125725366175175, + 0.012267515994608402, + 0.010350298136472702, + 0.009619438089430332, + 0.005267053842544556, + 0.01148983184248209, + 0.0060727340169250965, + 0.009108955040574074, + 0.013484890572726727, + 0.012546928599476814, + 0.027360882610082626, + 0.011059213429689407, + 0.013173398561775684, + 0.010379796847701073, + 0.0145872225984931, + 0.01111384853720665, + 0.018772441893815994, + 0.017118321731686592, + 0.010919817723333836, + 0.01721157878637314, + 0.03999350592494011, + 0.01147109642624855, + 0.04154229909181595, + 0.017435071989893913, + 0.015104621648788452, + 0.024845099076628685, + 0.01928057335317135, + 0.019955750554800034, + 0.010731310583651066, + 0.010311971418559551, + 0.008076922036707401, + 0.011635319329798222, + 0.014441782608628273, + 0.017724642530083656, + 0.011199429631233215, + 0.014740739949047565, + 0.00837345328181982, + 0.004391524009406567, + 0.010109937749803066, + 0.0051613058894872665, + 0.00845412164926529, + 0.01584555208683014, + 0.008961851708590984, + 0.008743486367166042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.024745620787143707 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043072763830423355 + ], + "opposite_logprob_tokens": [ + -3.1680727005004883 + ], + "logprob_reward": -0.043072763830423355, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043072763830423355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32316431403160095, + 0.06159988045692444, + 0.22958976030349731, + 0.03949550539255142, + 0.03961535915732384, + 0.11782890558242798, + 0.037541549652814865, + 0.0543844997882843, + 0.04752654954791069, + 0.0492536723613739 + ], + "credit_map": [ + 0.012693362310528755, + 0.02855774573981762, + 0.01437296811491251, + 0.044746436178684235, + 0.042167410254478455, + 0.050229817628860474, + 0.08639950305223465, + 0.043997086584568024, + 0.003370927646756172, + 0.00655444897711277, + 0.003693419275805354, + 0.007068735081702471, + 0.007977680303156376, + 0.010625150986015797, + 0.010512477718293667, + 0.01179704163223505, + 0.017355898395180702, + 0.02960660122334957, + 0.02111167460680008, + 0.028457816690206528, + 0.031556107103824615, + 0.034757424145936966, + 0.036342963576316833, + 0.030401283875107765, + 0.0020368911791592836, + 0.004752533510327339, + 0.0029538865201175213, + 0.004693140275776386, + 0.005743378307670355, + 0.0067381407134234905, + 0.005570000968873501, + 0.007007535547018051, + 0.0022748010233044624, + 0.0034544123336672783, + 0.0026886288542300463, + 0.0036395196802914143, + 0.00721231522038579, + 0.0069455429911613464, + 0.005697430111467838, + 0.00770271010696888, + 0.004884784109890461, + 0.01226745080202818, + 0.009159418754279613, + 0.008321959525346756, + 0.014552938751876354, + 0.02310962975025177, + 0.021583054214715958, + 0.023949673399329185, + 0.0035434411838650703, + 0.004267496522516012, + 0.004172697197645903, + 0.0032032669987529516, + 0.006150766275823116, + 0.005442016292363405, + 0.005689414218068123, + 0.005072453990578651, + 0.010263469070196152, + 0.003439567517489195, + 0.012391560710966587, + 0.005103221163153648, + 0.005445630755275488, + 0.007128572557121515, + 0.0047875638119876385, + 0.0058249118737876415, + 0.004375577904284, + 0.005884614307433367, + 0.0036392919719219208, + 0.005853624548763037, + 0.008191497065126896, + 0.005582026205956936, + 0.009111763909459114, + 0.004888152237981558, + 0.006170408334583044, + 0.0027267164550721645, + 0.007291961461305618, + 0.003837181255221367, + 0.005833733826875687, + 0.010365028865635395, + 0.005852151662111282, + 0.007176491431891918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.043072763830423355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197722733020782 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.018197722733020782, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018197722733020782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14480385184288025, + 0.21952997148036957, + 0.10938044637441635, + 0.05197560787200928, + 0.09746959060430527, + 0.04396640136837959, + 0.05277092382311821, + 0.05193965137004852, + 0.13939087092876434, + 0.08877266943454742 + ], + "credit_map": [ + 0.007992230355739594, + 0.017687425017356873, + 0.008235367015004158, + 0.014514995738863945, + 0.02239258587360382, + 0.02490130253136158, + 0.029535846784710884, + 0.0195440836250782, + 0.042446568608284, + 0.023377269506454468, + 0.029248740524053574, + 0.01938946172595024, + 0.01652083918452263, + 0.03589240461587906, + 0.024923399090766907, + 0.02773129940032959, + 0.005334699060767889, + 0.014996928162872791, + 0.010197642259299755, + 0.011018330231308937, + 0.016951855272054672, + 0.017696769908070564, + 0.01722833514213562, + 0.015955889597535133, + 0.002685774117708206, + 0.0065752738155424595, + 0.003953699488192797, + 0.004552173428237438, + 0.010596583597362041, + 0.00813454482704401, + 0.00966400746256113, + 0.005813549738377333, + 0.010478869080543518, + 0.012663492932915688, + 0.008623522706329823, + 0.00950708333402872, + 0.014859072864055634, + 0.016976645216345787, + 0.011925995349884033, + 0.012434910051524639, + 0.0021940614096820354, + 0.004244933370500803, + 0.002514158608391881, + 0.004041076637804508, + 0.006560106761753559, + 0.006011771969497204, + 0.012324740178883076, + 0.006075550802052021, + 0.0032141092233359814, + 0.006843606475740671, + 0.004697146825492382, + 0.006508124992251396, + 0.0100734643638134, + 0.007178499363362789, + 0.007848360575735569, + 0.006407608278095722, + 0.006327308248728514, + 0.004553736187517643, + 0.007823988795280457, + 0.004616212099790573, + 0.00827919039875269, + 0.007810384966433048, + 0.00563571322709322, + 0.006893114652484655, + 0.01740676909685135, + 0.010875439271330833, + 0.012751513160765171, + 0.014280671253800392, + 0.01663345843553543, + 0.02847832627594471, + 0.013426193036139011, + 0.025538496673107147, + 0.009439760819077492, + 0.005353880580514669, + 0.011980956420302391, + 0.006594302132725716, + 0.01056186854839325, + 0.020214887335896492, + 0.010230788961052895, + 0.014396226964890957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.018197722733020782 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630723994225264 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630723994225264, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630723994225264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10463947802782059, + 0.03756390139460564, + 0.06161702796816826, + 0.04721767455339432, + 0.05060268193483353, + 0.10754645615816116, + 0.0749846026301384, + 0.12156857550144196, + 0.2738924026489258, + 0.12036718428134918 + ], + "credit_map": [ + 0.005478906445205212, + 0.014526162296533585, + 0.0066522108390927315, + 0.012733740732073784, + 0.012133625335991383, + 0.01910586468875408, + 0.019835906103253365, + 0.014173069037497044, + 0.0035339216701686382, + 0.003656616900116205, + 0.002998506184667349, + 0.0029003871604800224, + 0.004232499282807112, + 0.006968315690755844, + 0.004078475758433342, + 0.009195183403789997, + 0.005480680149048567, + 0.0077753812074661255, + 0.004773851018399, + 0.005271992180496454, + 0.008156495168805122, + 0.010577086359262466, + 0.008320311084389687, + 0.011261235922574997, + 0.004750932101160288, + 0.005352652166038752, + 0.0040375166572630405, + 0.004057648126035929, + 0.005298517644405365, + 0.008120384998619556, + 0.006408833898603916, + 0.00919119082391262, + 0.005372706335037947, + 0.004229050129652023, + 0.00601165508851409, + 0.004588535521179438, + 0.006641227286309004, + 0.009066055528819561, + 0.0048119607381522655, + 0.00988149642944336, + 0.0037948277313262224, + 0.012534857727587223, + 0.008376953192055225, + 0.008568434044718742, + 0.014184810221195221, + 0.02033114805817604, + 0.022106759250164032, + 0.017648670822381973, + 0.006417314056307077, + 0.009388468228280544, + 0.0041475798934698105, + 0.0059882597997784615, + 0.01223787572234869, + 0.012122606858611107, + 0.011096328496932983, + 0.013586172834038734, + 0.025033200159668922, + 0.008793209679424763, + 0.025041090324521065, + 0.011772979982197285, + 0.010995489545166492, + 0.014899716712534428, + 0.01466287113726139, + 0.010370026342570782, + 0.02755863219499588, + 0.020643698051571846, + 0.02376016601920128, + 0.029825711622834206, + 0.030525613576173782, + 0.0625620037317276, + 0.027848325669765472, + 0.05116824805736542, + 0.015167954377830029, + 0.011505849659442902, + 0.017034858465194702, + 0.009809182956814766, + 0.013349080458283424, + 0.023334050551056862, + 0.015976039692759514, + 0.014190171845257282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.007630723994225264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0002388668799540028 + ], + "opposite_logprob_tokens": [ + -8.375238418579102 + ], + "logprob_reward": -0.0002388668799540028, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.0002388668799540028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12648753821849823, + 0.34752795100212097, + 0.0726669505238533, + 0.08775163441896439, + 0.040476616472005844, + 0.04779457300901413, + 0.037895314395427704, + 0.054112788289785385, + 0.07591204345226288, + 0.10937467217445374 + ], + "credit_map": [ + 0.008979085832834244, + 0.014350812882184982, + 0.0070154061540961266, + 0.015922745689749718, + 0.017609886825084686, + 0.021886177361011505, + 0.023442069068551064, + 0.01728134974837303, + 0.06549623608589172, + 0.03970199078321457, + 0.045319799333810806, + 0.03430917486548424, + 0.032223835587501526, + 0.058045558631420135, + 0.031336572021245956, + 0.041094765067100525, + 0.008091458119452, + 0.007926607504487038, + 0.006677562836557627, + 0.00683769304305315, + 0.009760257788002491, + 0.011993683874607086, + 0.010739863850176334, + 0.010639823041856289, + 0.009800502099096775, + 0.0072227600030601025, + 0.0075729708187282085, + 0.00737035321071744, + 0.00958156492561102, + 0.014706233516335487, + 0.010820942930877209, + 0.020676305517554283, + 0.005336232949048281, + 0.0034001627936959267, + 0.004598491359502077, + 0.00465952605009079, + 0.005628000944852829, + 0.005722455680370331, + 0.0048251524567604065, + 0.006306592840701342, + 0.002470036270096898, + 0.004719988442957401, + 0.003392928745597601, + 0.004420826677232981, + 0.006531786639243364, + 0.0071444204077124596, + 0.01280379481613636, + 0.006310791242867708, + 0.002907954156398773, + 0.0037459435407072306, + 0.0030613336712121964, + 0.004086820874363184, + 0.00906393863260746, + 0.004909211304038763, + 0.005606299266219139, + 0.004513814114034176, + 0.008870704099535942, + 0.003581268945708871, + 0.010443199425935745, + 0.005897237453609705, + 0.006455743219703436, + 0.006556645035743713, + 0.007090567145496607, + 0.005217425990849733, + 0.008980413898825645, + 0.006633787415921688, + 0.006384067237377167, + 0.00693716062232852, + 0.010003533214330673, + 0.01471641007810831, + 0.00887406524270773, + 0.013382607139647007, + 0.012135042808949947, + 0.006525357719510794, + 0.0097233010455966, + 0.005287348758429289, + 0.011101474985480309, + 0.035709500312805176, + 0.011920258402824402, + 0.016972390934824944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0002388668799540028 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily approaches a man standing in a dimly lit, luxurious room. He wears a perfectly tailored dark charcoal wool suit jacket over a crisp white dress shirt. A single, dramatic key light from the side sculpts his form, casting deep shadows and creating a powerful black-white contrast, highlighting the sharp line of his jaw and his focused, determined expression. His dark hair is neatly styled. As the camera advances, his hands rise with deliberate grace, resting on his neck. He carefully straightens the collar of his starched white shirt, his fingertips making tiny, precise adjustments to ensure perfect symmetry. The intense light captures the fine twill weave of the shirt fabric and the subtle texture of his skin, creating a tactile visual experience. The background remains a soft, dark blur, ensuring that all attention is focused on this ritual of preparation. The camera's slow, relentless push culminates in extreme close-up, filling the frame with the texture of the collar and the steady, controlled pressure of his fingers, creating an atmosphere of focused attention and quiet, powerful confidence.", + "seed": 7619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_15/index_9047_seed_7619_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": -0.06145086485194042, + "vlm_logprob": -0.06145086485194042, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06145086485194042, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02863580733537674 + ], + "opposite_logprob_tokens": [ + -3.7786357402801514 + ], + "logprob_reward": -0.02863580733537674, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02863580733537674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10838598012924194, + 0.11952486634254456, + 0.07633852958679199, + 0.12207327783107758, + 0.12267318367958069, + 0.18416738510131836, + 0.03573883697390556, + 0.08065937459468842, + 0.08898401260375977, + 0.06145450845360756 + ], + "credit_map": [ + 0.013005701825022697, + 0.009929914027452469, + 0.015437156893312931, + 0.007684915792196989, + 0.018759138882160187, + 0.014734386466443539, + 0.015011810697615147, + 0.013822960667312145, + 0.010297746397554874, + 0.006174993235617876, + 0.01012357883155346, + 0.009773398749530315, + 0.009275957942008972, + 0.026415210217237473, + 0.009363982826471329, + 0.03810000419616699, + 0.007131673861294985, + 0.003668221179395914, + 0.005732621997594833, + 0.0030272891744971275, + 0.008116763085126877, + 0.017276953905820847, + 0.006375161465257406, + 0.025009844452142715, + 0.005769733339548111, + 0.008236294612288475, + 0.011161025613546371, + 0.00937049463391304, + 0.015641896054148674, + 0.022807763889431953, + 0.018605049699544907, + 0.030481012538075447, + 0.008951577357947826, + 0.011346354149281979, + 0.016400277614593506, + 0.00698166573420167, + 0.018702859058976173, + 0.0212213434278965, + 0.015703512355685234, + 0.023365601897239685, + 0.02076091431081295, + 0.01289521623402834, + 0.02080150879919529, + 0.008116190321743488, + 0.02089979685842991, + 0.03406483307480812, + 0.018397655338048935, + 0.04823126643896103, + 0.003966844640672207, + 0.002818381180986762, + 0.005525079555809498, + 0.00377909024246037, + 0.0036997166462242603, + 0.004853416234254837, + 0.004728785715997219, + 0.006367519963532686, + 0.00679572531953454, + 0.007829722948372364, + 0.011288095265626907, + 0.005387049168348312, + 0.00925448164343834, + 0.011673493310809135, + 0.01636590249836445, + 0.012064904905855656, + 0.005341624841094017, + 0.010781955905258656, + 0.023374535143375397, + 0.004073931835591793, + 0.01163559965789318, + 0.008680838160216808, + 0.005143758840858936, + 0.019951771944761276, + 0.00439852848649025, + 0.00614542281255126, + 0.0134773263707757, + 0.004455467686057091, + 0.011376085691154003, + 0.008611498400568962, + 0.004909452050924301, + 0.008080731146037579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.02863580733537674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043093886226415634 + ], + "opposite_logprob_tokens": [ + -3.1680939197540283 + ], + "logprob_reward": -0.043093886226415634, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043093886226415634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18488629162311554, + 0.05732019990682602, + 0.055887941271066666, + 0.05091474577784538, + 0.24983327090740204, + 0.0401812307536602, + 0.06745002418756485, + 0.049566395580768585, + 0.20421965420246124, + 0.039740271866321564 + ], + "credit_map": [ + 0.023726921528577805, + 0.012552089057862759, + 0.015100708231329918, + 0.011005193926393986, + 0.022155003622174263, + 0.026479186490178108, + 0.037482790648937225, + 0.036384399980306625, + 0.004483614582568407, + 0.00357064139097929, + 0.005382265895605087, + 0.0070360503159463406, + 0.007939518429338932, + 0.009972866624593735, + 0.0067368000745773315, + 0.01219844352453947, + 0.004058065824210644, + 0.00331296119838953, + 0.004786534234881401, + 0.003885354846715927, + 0.004871913231909275, + 0.011405273340642452, + 0.008642684668302536, + 0.014925152063369751, + 0.0027706369291990995, + 0.0038611777126789093, + 0.0052469875663518906, + 0.003307984210550785, + 0.00696314312517643, + 0.009171294048428535, + 0.00801315251737833, + 0.011580371297895908, + 0.01713973842561245, + 0.01837325282394886, + 0.03305128589272499, + 0.015339565463364124, + 0.03720995411276817, + 0.04033965244889259, + 0.05519108101725578, + 0.03318873420357704, + 0.0032899531070142984, + 0.004003116395324469, + 0.006514726206660271, + 0.003730457741767168, + 0.006063689012080431, + 0.005702428985387087, + 0.00447854632511735, + 0.006398311350494623, + 0.005499652586877346, + 0.006442975252866745, + 0.010303040035068989, + 0.004290060140192509, + 0.009945854544639587, + 0.007792716380208731, + 0.010493389330804348, + 0.012682337313890457, + 0.005429110024124384, + 0.005487480666488409, + 0.009744974784553051, + 0.004098857752978802, + 0.0055325529538095, + 0.005160608794540167, + 0.005531072150915861, + 0.008581740781664848, + 0.012563231401145458, + 0.022474151104688644, + 0.03771284222602844, + 0.009626353159546852, + 0.029866283759474754, + 0.017210397869348526, + 0.015791837126016617, + 0.05897456780076027, + 0.0021473634988069534, + 0.004051717463880777, + 0.005304891150444746, + 0.0018544922349974513, + 0.009332114830613136, + 0.0061135548166930676, + 0.004522968549281359, + 0.006413167342543602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.043093886226415634 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498476326465607 + ], + "opposite_logprob_tokens": [ + -2.9299848079681396 + ], + "logprob_reward": -0.05498476326465607, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498476326465607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2166936993598938, + 0.3728153705596924, + 0.12808603048324585, + 0.04740702733397484, + 0.07401185482740402, + 0.028973082080483437, + 0.015831494703888893, + 0.042681992053985596, + 0.05613793060183525, + 0.01736149750649929 + ], + "credit_map": [ + 0.03368120267987251, + 0.014086266048252583, + 0.018859583884477615, + 0.010041521862149239, + 0.018492287024855614, + 0.038146551698446274, + 0.026821831241250038, + 0.05656446889042854, + 0.029935533180832863, + 0.021253712475299835, + 0.036449071019887924, + 0.04252820089459419, + 0.03641950339078903, + 0.07496929913759232, + 0.029980750754475594, + 0.10127930343151093, + 0.010490586049854755, + 0.009373224340379238, + 0.014127478934824467, + 0.013997795060276985, + 0.022057468071579933, + 0.019513292238116264, + 0.015376472845673561, + 0.02314971387386322, + 0.0030913681257516146, + 0.002811904065310955, + 0.0072012534365057945, + 0.0032387878745794296, + 0.006275056395679712, + 0.0094005037099123, + 0.0049722581170499325, + 0.010415894910693169, + 0.004380229860544205, + 0.006019513122737408, + 0.010147959925234318, + 0.004480333998799324, + 0.012302734889090061, + 0.012466896325349808, + 0.013601562939584255, + 0.010612623766064644, + 0.0021204000804573298, + 0.0018967259675264359, + 0.003366647055372596, + 0.0018776687793433666, + 0.00338556244969368, + 0.005111416801810265, + 0.003672083606943488, + 0.007542579434812069, + 0.0013799673179164529, + 0.0014688163064420223, + 0.0020582308061420918, + 0.0010964121902361512, + 0.0029173872899264097, + 0.0019275423837825656, + 0.002076961100101471, + 0.002906177658587694, + 0.003025548066943884, + 0.004320226609706879, + 0.005941607058048248, + 0.0030609890818595886, + 0.0045859538950026035, + 0.00684077525511384, + 0.007908456958830357, + 0.006998433265835047, + 0.00404323311522603, + 0.0063324617221951485, + 0.017222721129655838, + 0.002614089287817478, + 0.004959213547408581, + 0.006197723094373941, + 0.00364435906521976, + 0.011124123819172382, + 0.0012917398707941175, + 0.0014654126716777682, + 0.0027541634626686573, + 0.0012725855922326446, + 0.002526387572288513, + 0.0024327330756932497, + 0.0024939069990068674, + 0.0031245676800608635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.05498476326465607 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990349233150482 + ], + "opposite_logprob_tokens": [ + -1.804903507232666 + ], + "logprob_reward": -0.17990349233150482, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990349233150482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11233571171760559, + 0.05996467545628548, + 0.046237751841545105, + 0.04914035648107529, + 0.06569403409957886, + 0.2644595801830292, + 0.0996638759970665, + 0.1081262081861496, + 0.17720238864421844, + 0.017175475135445595 + ], + "credit_map": [ + 0.010579075664281845, + 0.006792660336941481, + 0.013085349462926388, + 0.00762287387624383, + 0.012965189293026924, + 0.01765289157629013, + 0.01795678213238716, + 0.02568088099360466, + 0.006270161364227533, + 0.004299726337194443, + 0.006086854729801416, + 0.006437847390770912, + 0.007433596067130566, + 0.009367674589157104, + 0.00789429247379303, + 0.012174516916275024, + 0.003529186826199293, + 0.003134693019092083, + 0.003417044412344694, + 0.0023591292556375265, + 0.004308153875172138, + 0.010359985753893852, + 0.0040433588437736034, + 0.015086200088262558, + 0.0035498046781867743, + 0.00406266376376152, + 0.004131488036364317, + 0.0034501818008720875, + 0.0058388253673911095, + 0.009731383994221687, + 0.004641003906726837, + 0.013735000044107437, + 0.0048548574559390545, + 0.005294454284012318, + 0.007914882153272629, + 0.003755410900339484, + 0.007337198592722416, + 0.01098097674548626, + 0.011216266080737114, + 0.014339984394609928, + 0.028291016817092896, + 0.017692746594548225, + 0.03540835902094841, + 0.01669795624911785, + 0.02581113763153553, + 0.044228147715330124, + 0.03754965215921402, + 0.05878053978085518, + 0.009949453175067902, + 0.007766261231154203, + 0.01638668216764927, + 0.005380243994295597, + 0.013955021277070045, + 0.0162618737667799, + 0.01240689866244793, + 0.01755743846297264, + 0.007855369709432125, + 0.008231009356677532, + 0.016504641622304916, + 0.010128242895007133, + 0.014502870850265026, + 0.01848103664815426, + 0.015454362146556377, + 0.01696866936981678, + 0.018686750903725624, + 0.030732300132513046, + 0.029320642352104187, + 0.01572975143790245, + 0.027626169845461845, + 0.021890565752983093, + 0.013614094816148281, + 0.019602106884121895, + 0.0010853091953322291, + 0.0016039933543652296, + 0.002823040122166276, + 0.0010791706154122949, + 0.0021377543453127146, + 0.0029922048561275005, + 0.0025691029150038958, + 0.0028848997317254543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.17990349233150482 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006363751017488539 + ], + "opposite_logprob_tokens": [ + -7.375636577606201 + ], + "logprob_reward": -0.0006363751017488539, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006363751017488539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21629612147808075, + 0.12207277119159698, + 0.13762253522872925, + 0.05447600409388542, + 0.11342845112085342, + 0.0806448683142662, + 0.04860770329833031, + 0.05200188234448433, + 0.12936687469482422, + 0.04548279196023941 + ], + "credit_map": [ + 0.028134401887655258, + 0.012660871259868145, + 0.02022307924926281, + 0.01138982642441988, + 0.02911008521914482, + 0.0357779823243618, + 0.030513711273670197, + 0.048486169427633286, + 0.010714145377278328, + 0.01030730176717043, + 0.010831071063876152, + 0.016626305878162384, + 0.01885337568819523, + 0.02014290727674961, + 0.01432518195360899, + 0.02027248591184616, + 0.010839142836630344, + 0.012245016172528267, + 0.02529754862189293, + 0.014456678181886673, + 0.016645031049847603, + 0.02075887657701969, + 0.015203847549855709, + 0.022176392376422882, + 0.0034809373319149017, + 0.004128757398575544, + 0.0056928652338683605, + 0.003672739490866661, + 0.00869895238429308, + 0.008628846146166325, + 0.008948402479290962, + 0.011224507354199886, + 0.010394812561571598, + 0.0076139625161886215, + 0.0136791430413723, + 0.007155213039368391, + 0.013128455728292465, + 0.020519690588116646, + 0.018907934427261353, + 0.02202923782169819, + 0.007127134129405022, + 0.005493553355336189, + 0.01228862814605236, + 0.006638925988227129, + 0.008059763349592686, + 0.013733135536313057, + 0.009359024465084076, + 0.017944704741239548, + 0.006023676600307226, + 0.004599675070494413, + 0.008688656613230705, + 0.005172181408852339, + 0.005021760240197182, + 0.006836338434368372, + 0.004872702062129974, + 0.007392709609121084, + 0.004213780630379915, + 0.004880260676145554, + 0.00665205018594861, + 0.003444717265665531, + 0.006407802924513817, + 0.009337000548839569, + 0.009984769858419895, + 0.007081498391926289, + 0.008718093857169151, + 0.010370121337473392, + 0.04289354383945465, + 0.005219582933932543, + 0.014882855117321014, + 0.01045116875320673, + 0.006776735186576843, + 0.030054781585931778, + 0.0035528745502233505, + 0.004158872179687023, + 0.008984727784991264, + 0.0031010110396891832, + 0.00863646063953638, + 0.005364030599594116, + 0.004800509661436081, + 0.006884302478283644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0006363751017488539 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily approaches a man standing in a dimly lit, luxurious room. He wears a perfectly tailored dark charcoal wool suit jacket over a crisp white dress shirt. A single, dramatic key light from the side sculpts his form, casting deep shadows and creating a powerful black-white contrast, highlighting the sharp line of his jaw and his focused, determined expression. His dark hair is neatly styled. As the camera advances, his hands rise with deliberate grace, resting on his neck. He carefully straightens the collar of his starched white shirt, his fingertips making tiny, precise adjustments to ensure perfect symmetry. The intense light captures the fine twill weave of the shirt fabric and the subtle texture of his skin, creating a tactile visual experience. The background remains a soft, dark blur, ensuring that all attention is focused on this ritual of preparation. The camera's slow, relentless push culminates in extreme close-up, filling the frame with the texture of the collar and the steady, controlled pressure of his fingers, creating an atmosphere of focused attention and quiet, powerful confidence.", + "seed": 7620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_15/index_9047_seed_7620_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": -0.013719825062435121, + "vlm_logprob": -0.013719825062435121, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.013719825062435121, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02018657512962818 + ], + "opposite_logprob_tokens": [ + -4.270186424255371 + ], + "logprob_reward": -0.02018657512962818, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02018657512962818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19232337176799774, + 0.06835076957941055, + 0.0692160427570343, + 0.11150376498699188, + 0.08439522981643677, + 0.2149333953857422, + 0.07543499767780304, + 0.09731423109769821, + 0.045279376208782196, + 0.041248869150877 + ], + "credit_map": [ + 0.0190606527030468, + 0.03322641924023628, + 0.02392961084842682, + 0.03781048581004143, + 0.016060639172792435, + 0.019949251785874367, + 0.011233494617044926, + 0.031052812933921814, + 0.0051146564073860645, + 0.008860856294631958, + 0.010021540336310863, + 0.013375614769756794, + 0.005870067980140448, + 0.006224878132343292, + 0.011122068390250206, + 0.007761088665574789, + 0.006830332800745964, + 0.011244216002523899, + 0.006514769978821278, + 0.014481760561466217, + 0.005688190460205078, + 0.005484404973685741, + 0.012032715603709221, + 0.006939656101167202, + 0.00956597551703453, + 0.02420099638402462, + 0.014124174602329731, + 0.02371986396610737, + 0.007258119061589241, + 0.011840102262794971, + 0.008750425651669502, + 0.012044111266732216, + 0.007504622917622328, + 0.01633039303123951, + 0.010035797022283077, + 0.015617400407791138, + 0.006452747620642185, + 0.008876580744981766, + 0.00593193294480443, + 0.013645751401782036, + 0.04868925362825394, + 0.019850514829158783, + 0.049191202968358994, + 0.01916637271642685, + 0.011546033434569836, + 0.02350568398833275, + 0.016502568498253822, + 0.02648177742958069, + 0.012632336467504501, + 0.011810937896370888, + 0.012209488078951836, + 0.006208600010722876, + 0.012445389293134212, + 0.010311643593013287, + 0.005729507189244032, + 0.004087097477167845, + 0.014957441948354244, + 0.01065252348780632, + 0.017932476475834846, + 0.011742684990167618, + 0.012501378543674946, + 0.009068995714187622, + 0.010531341657042503, + 0.009927384555339813, + 0.007465224713087082, + 0.005123203154653311, + 0.0074669308960437775, + 0.0040252357721328735, + 0.00364490645006299, + 0.003850102424621582, + 0.0034701647236943245, + 0.010233606211841106, + 0.006488010287284851, + 0.004081497434526682, + 0.00759631022810936, + 0.0038368673995137215, + 0.0029963550623506308, + 0.004724606405943632, + 0.0032782787457108498, + 0.008246945217251778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.02018657512962818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007645866833627224 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.007645866833627224, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645866833627224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08135316520929337, + 0.040442898869514465, + 0.04822930693626404, + 0.058340538293123245, + 0.09374600648880005, + 0.152289479970932, + 0.09553971886634827, + 0.12306036055088043, + 0.07180347293615341, + 0.23519505560398102 + ], + "credit_map": [ + 0.00889931432902813, + 0.011535289697349072, + 0.00974072702229023, + 0.014714237302541733, + 0.009388704784214497, + 0.0086545180529356, + 0.006958504673093557, + 0.011461869813501835, + 0.0030571832321584225, + 0.005943623371422291, + 0.006304506212472916, + 0.008694184944033623, + 0.0039865099824965, + 0.004584428854286671, + 0.00311027723364532, + 0.004762184806168079, + 0.003742043860256672, + 0.005810488015413284, + 0.00539231626316905, + 0.005370107479393482, + 0.007241810206323862, + 0.0033523908350616693, + 0.012543068267405033, + 0.004777082707732916, + 0.004830587655305862, + 0.010659012012183666, + 0.006935978773981333, + 0.009266248904168606, + 0.006171812303364277, + 0.005114289466291666, + 0.00955114047974348, + 0.005811469629406929, + 0.00983328465372324, + 0.010946530848741531, + 0.01391680259257555, + 0.015301598235964775, + 0.0070532034151256084, + 0.00832372810691595, + 0.01164760161191225, + 0.016723258420825005, + 0.0435459278523922, + 0.008880654349923134, + 0.04261395335197449, + 0.011881622485816479, + 0.008038144558668137, + 0.0154050188139081, + 0.009949474595487118, + 0.011974688619375229, + 0.013733281753957272, + 0.012302274815738201, + 0.011865166947245598, + 0.007041904143989086, + 0.018496282398700714, + 0.015055621974170208, + 0.007893050089478493, + 0.009152138605713844, + 0.019005829468369484, + 0.012876004911959171, + 0.020892225205898285, + 0.011475145816802979, + 0.020855434238910675, + 0.01121549028903246, + 0.013438346795737743, + 0.013301891274750233, + 0.009498827159404755, + 0.006430290639400482, + 0.015371216461062431, + 0.007954872213304043, + 0.006812024395912886, + 0.006004047580063343, + 0.009368783794343472, + 0.010363408364355564, + 0.03876454755663872, + 0.02726156823337078, + 0.04221738502383232, + 0.025158721953630447, + 0.023655785247683525, + 0.021775973960757256, + 0.02754235453903675, + 0.02881873957812786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.007645866833627224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357457041740417 + ], + "opposite_logprob_tokens": [ + -3.651357412338257 + ], + "logprob_reward": -0.026357457041740417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357457041740417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19806578755378723, + 0.058735765516757965, + 0.06151966005563736, + 0.0869448184967041, + 0.2189972996711731, + 0.24567952752113342, + 0.03962422162294388, + 0.02552940510213375, + 0.03913557529449463, + 0.02576790750026703 + ], + "credit_map": [ + 0.024277813732624054, + 0.040512166917324066, + 0.021383248269557953, + 0.03149939700961113, + 0.019191808998584747, + 0.01861625723540783, + 0.015192796476185322, + 0.027392297983169556, + 0.005940886680036783, + 0.011440915986895561, + 0.006863850634545088, + 0.012256531976163387, + 0.004443393554538488, + 0.005034197121858597, + 0.006144896615296602, + 0.006611097138375044, + 0.005238910671323538, + 0.007953302003443241, + 0.007160122971981764, + 0.010774358175694942, + 0.006618367042392492, + 0.005256031174212694, + 0.012224648147821426, + 0.006293925456702709, + 0.007837862707674503, + 0.01817464269697666, + 0.010253814980387688, + 0.02091377228498459, + 0.004917538724839687, + 0.00868899468332529, + 0.005646094214171171, + 0.010512103326618671, + 0.02252105623483658, + 0.0475495420396328, + 0.028289608657360077, + 0.044448018074035645, + 0.01141930092126131, + 0.023225175216794014, + 0.015643619000911713, + 0.025900984182953835, + 0.06385990977287292, + 0.01763768680393696, + 0.07609128952026367, + 0.024003583937883377, + 0.00933158490806818, + 0.023787790909409523, + 0.011600910685956478, + 0.0193667933344841, + 0.005837060511112213, + 0.00572499493137002, + 0.0069069103337824345, + 0.0030210199765861034, + 0.004708434920758009, + 0.00465006660670042, + 0.004311710596084595, + 0.004464027006179094, + 0.0038508495781570673, + 0.0046152109280228615, + 0.004694946575909853, + 0.002462630392983556, + 0.0034167503472417593, + 0.002115563489496708, + 0.0024271931033581495, + 0.0019462620839476585, + 0.00588518101722002, + 0.0035602052230387926, + 0.00971392821520567, + 0.00448719784617424, + 0.0025347580667585135, + 0.0038770618848502636, + 0.004605464171618223, + 0.004471782129257917, + 0.004068637732416391, + 0.002362783532589674, + 0.004736705217510462, + 0.0023750821128487587, + 0.002370300702750683, + 0.0022252595517784357, + 0.0038471678271889687, + 0.0037819722201675177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.026357457041740417 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197533018887043 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197533018887043, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197533018887043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08552929759025574, + 0.03457894176244736, + 0.05905978009104729, + 0.09274474531412125, + 0.06135961413383484, + 0.06202835962176323, + 0.13084562122821808, + 0.16561400890350342, + 0.26567307114601135, + 0.04256661236286163 + ], + "credit_map": [ + 0.008377066813409328, + 0.016342544928193092, + 0.012303736060857773, + 0.01588897965848446, + 0.007451794110238552, + 0.008429892361164093, + 0.0043456703424453735, + 0.012389599345624447, + 0.003939833026379347, + 0.0059981816448271275, + 0.005570049397647381, + 0.0066559044644236565, + 0.0023529850877821445, + 0.003809342859312892, + 0.0022167395800352097, + 0.004035904072225094, + 0.006214275024831295, + 0.009644512087106705, + 0.006989000830799341, + 0.00967307761311531, + 0.007693152874708176, + 0.00529828667640686, + 0.007194673176854849, + 0.006352797616273165, + 0.008001909591257572, + 0.010982311330735683, + 0.009380381554365158, + 0.01832481473684311, + 0.008965044282376766, + 0.006653845775872469, + 0.019000424072146416, + 0.01143600046634674, + 0.005963276140391827, + 0.00889354757964611, + 0.0061861248686909676, + 0.016915399581193924, + 0.004707193002104759, + 0.003834055969491601, + 0.006904982030391693, + 0.007955029606819153, + 0.010503990575671196, + 0.010347137227654457, + 0.010062258690595627, + 0.005540722515434027, + 0.003975677769631147, + 0.008158341981470585, + 0.005013194866478443, + 0.008427029475569725, + 0.017412081360816956, + 0.028993405401706696, + 0.020312175154685974, + 0.011954442597925663, + 0.011063610203564167, + 0.015835775062441826, + 0.014413594268262386, + 0.010860525071620941, + 0.03258856013417244, + 0.02266707271337509, + 0.04869117587804794, + 0.020031582564115524, + 0.012445288710296154, + 0.00851949118077755, + 0.012329377233982086, + 0.008341453038156033, + 0.04091395437717438, + 0.0367557555437088, + 0.07779939472675323, + 0.02316557615995407, + 0.01883443258702755, + 0.01505932305008173, + 0.025532923638820648, + 0.027611685916781425, + 0.006229996215552092, + 0.0037989832926541567, + 0.004624970257282257, + 0.006396953482180834, + 0.006799382623285055, + 0.004832395352423191, + 0.004233401734381914, + 0.005650525912642479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.014197533018887043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing a dark charcoal wool suit jacket over a crisp white dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the man raise his hands to his neck and straighten the collar of his shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single side key light creating deep shadows and high black-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily approach the subject, culminating in an extreme close-up of the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, dark blur that keeps the focus on the man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00021169328829273582 + ], + "opposite_logprob_tokens": [ + -8.500211715698242 + ], + "logprob_reward": -0.00021169328829273582, + "yes_no_margin_reward": 8.5, + "token_credit_reward": -0.00021169328829273582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24474315345287323, + 0.03634543716907501, + 0.04752098768949509, + 0.04093943536281586, + 0.12126034498214722, + 0.22506387531757355, + 0.06715639680624008, + 0.05325177311897278, + 0.1219693049788475, + 0.04174930229783058 + ], + "credit_map": [ + 0.024290237575769424, + 0.04468473792076111, + 0.037559300661087036, + 0.06712821871042252, + 0.018466025590896606, + 0.016608400270342827, + 0.009852126240730286, + 0.026154102757573128, + 0.004167086444795132, + 0.006346751935780048, + 0.004595677834004164, + 0.007190957199782133, + 0.003058885456994176, + 0.0042025260627269745, + 0.0024086700286716223, + 0.004374880809336901, + 0.0045586153864860535, + 0.009802794083952904, + 0.004500271286815405, + 0.008832416497170925, + 0.00677338195964694, + 0.0040915110148489475, + 0.004354668781161308, + 0.0046073272824287415, + 0.004383946303278208, + 0.005396751686930656, + 0.007052888162434101, + 0.005648170132189989, + 0.0062435404397547245, + 0.003524255007505417, + 0.0036550487857311964, + 0.0050348336808383465, + 0.011785444803535938, + 0.026700204238295555, + 0.014693054370582104, + 0.021224884316325188, + 0.007517077960073948, + 0.009501831606030464, + 0.014123553410172462, + 0.015714285895228386, + 0.04334310442209244, + 0.02707451768219471, + 0.03900975361466408, + 0.021981481462717056, + 0.014847188256680965, + 0.02863885834813118, + 0.018470844253897667, + 0.031698111444711685, + 0.00935145653784275, + 0.008517378941178322, + 0.010397764854133129, + 0.007683575619012117, + 0.006660462822765112, + 0.008823518641293049, + 0.007688592653721571, + 0.00803364347666502, + 0.010398408398032188, + 0.006627996452152729, + 0.015558131039142609, + 0.005429898388683796, + 0.004487426020205021, + 0.003113353392109275, + 0.0044220974668860435, + 0.003214461961761117, + 0.019439389929175377, + 0.013796864077448845, + 0.036929503083229065, + 0.011989954859018326, + 0.010528371669352055, + 0.0073114484548568726, + 0.012172883376479149, + 0.009800885803997517, + 0.005461862310767174, + 0.0043016718700528145, + 0.0059862323105335236, + 0.004337787162512541, + 0.00417061522603035, + 0.0038066490087658167, + 0.005237753503024578, + 0.008446730673313141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.00021169328829273582 + } + ], + "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/0000028/group_2/index_tensor([6483, 6483])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_2/index_tensor([6483, 6483])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9c5a032333bb4bf7f33fe34207b3e09ac052e661 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_2/index_tensor([6483, 6483])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium-long shot, captured from above, imbues the image with the texture of an ancient ink painting. The Monkey King, clad in dazzling golden chain mail, dominates the frame. Atop his head rests a phoenix-winged purple-gold crown, two long pheasant feathers trembling slightly with the force of his charged movement. His face is defiant, his fiery eyes gleaming with determination. He thrusts his body into a lunge, muscles flexing. He grips the heavy, jet-black Ruyi Jingu Club in both hands, raising it high above his head. With unstoppable force, he smashes it down upon a towering celestial mountain in the background, suspended above a sea of clouds. The instant the iron rod makes contact with the mountain's summit, the solid indigo-colored mass of the mountain cracks, sending the massive rock shattering into countless fragments. Most strikingly, the dust stirred up isn't just ordinary smoke, but countless splashing, splattering strokes of thick ink, condensing and flowing through the air with a powerful sense of movement and the rhythm of calligraphy. The main colors of the entire picture are composed of black, white and gray in varying shades. The only bright colors are the golden armor on the Monkey King and the fire in his eyes. The light spreads naturally on the rice paper, creating a unique visual effect of power and aesthetics. It is both violent and full of the unique freehand and ethereal beauty of ink painting. The atmosphere is solemn and extremely impactful.", + "seed": 856, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_2/index_6483_seed_856_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": -0.008675879053771495, + "vlm_logprob": -0.008675879053771495, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.008675879053771495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005201023072004318 + ], + "opposite_logprob_tokens": [ + -5.3802008628845215 + ], + "logprob_reward": -0.005201023072004318, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005201023072004318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05064457282423973, + 0.05328827723860741, + 0.054768092930316925, + 0.4677262604236603, + 0.057098835706710815, + 0.09746953845024109, + 0.03467165678739548, + 0.06465792655944824, + 0.0532662570476532, + 0.06640858948230743 + ], + "credit_map": [ + 0.005864319857209921, + 0.00835059117525816, + 0.005960854701697826, + 0.005431532394140959, + 0.005954831372946501, + 0.006010781973600388, + 0.007853048853576183, + 0.005218615289777517, + 0.006696544121950865, + 0.006456212606281042, + 0.007285044062882662, + 0.004167982842773199, + 0.006459630094468594, + 0.008623271249234676, + 0.006459689233452082, + 0.00713990768417716, + 0.007871134206652641, + 0.006856246385723352, + 0.005963381379842758, + 0.006390899885445833, + 0.008635720238089561, + 0.00579094747081399, + 0.007292312104254961, + 0.00596745265647769, + 0.04459263011813164, + 0.026178808882832527, + 0.1575351357460022, + 0.02292383648455143, + 0.021684274077415466, + 0.0890020802617073, + 0.03283543512225151, + 0.0729740634560585, + 0.006277394015341997, + 0.004794946871697903, + 0.01673051156103611, + 0.005518381949514151, + 0.004343406297266483, + 0.008342141285538673, + 0.004887930117547512, + 0.006204132456332445, + 0.01479104720056057, + 0.012509859167039394, + 0.016394805163145065, + 0.010084090754389763, + 0.00818796455860138, + 0.009896749630570412, + 0.009644143283367157, + 0.015960894525051117, + 0.00578478816896677, + 0.00402685534209013, + 0.005253055598586798, + 0.005802996456623077, + 0.004884067922830582, + 0.0037351984065026045, + 0.002645063679665327, + 0.002539635868743062, + 0.009956478141248226, + 0.006386667024344206, + 0.012566566467285156, + 0.007100274320691824, + 0.004457242786884308, + 0.007935885339975357, + 0.006793964188545942, + 0.009460857138037682, + 0.0066630966030061245, + 0.009187732823193073, + 0.005607542581856251, + 0.006750283297151327, + 0.009633880108594894, + 0.004882928915321827, + 0.005473509896546602, + 0.005067292135208845, + 0.010435515083372593, + 0.005866722669452429, + 0.01198545005172491, + 0.00586360739544034, + 0.007169533986598253, + 0.009748910553753376, + 0.006246217992156744, + 0.009092637337744236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.005201023072004318 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067454250529408455 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.0067454250529408455, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067454250529408455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2994838058948517, + 0.08662581443786621, + 0.033210232853889465, + 0.04021638631820679, + 0.04211847484111786, + 0.10787975788116455, + 0.05779500678181648, + 0.13773930072784424, + 0.09958051145076752, + 0.09535076469182968 + ], + "credit_map": [ + 0.0294878501445055, + 0.06699773669242859, + 0.028026901185512543, + 0.025283025577664375, + 0.045133598148822784, + 0.03159806504845619, + 0.0473279170691967, + 0.025628672912716866, + 0.011751470156013966, + 0.012522446922957897, + 0.012368956580758095, + 0.006394838914275169, + 0.009281559847295284, + 0.008970929309725761, + 0.013763254508376122, + 0.011572356335818768, + 0.004743020050227642, + 0.0037665204145014286, + 0.0058630406856536865, + 0.003051707986742258, + 0.004665686283260584, + 0.003806289052590728, + 0.004159474745392799, + 0.003154492937028408, + 0.003830832662060857, + 0.005017100367695093, + 0.008142305538058281, + 0.0030301378574222326, + 0.0046915519051253796, + 0.0056491270661354065, + 0.00378336850553751, + 0.006071958690881729, + 0.004982931073755026, + 0.006334118079394102, + 0.005080871284008026, + 0.003915733657777309, + 0.006083200220018625, + 0.005372823216021061, + 0.004980808589607477, + 0.005367991514503956, + 0.0180077962577343, + 0.018912602216005325, + 0.02187657169997692, + 0.011680868454277515, + 0.008688145317137241, + 0.01078404113650322, + 0.007206673268228769, + 0.010723056271672249, + 0.009364194236695766, + 0.007462922018021345, + 0.010423898696899414, + 0.010992021299898624, + 0.005128578282892704, + 0.005627588834613562, + 0.004685965832322836, + 0.004109837580472231, + 0.020853372290730476, + 0.011912146583199501, + 0.039981409907341, + 0.02577710710465908, + 0.010568076744675636, + 0.008068845607340336, + 0.01052121166139841, + 0.010057125240564346, + 0.01157867256551981, + 0.017901889979839325, + 0.010783135890960693, + 0.010051161050796509, + 0.017689291387796402, + 0.011349333450198174, + 0.009135697036981583, + 0.011091331951320171, + 0.01207326166331768, + 0.009388210251927376, + 0.02008402906358242, + 0.00794482696801424, + 0.01246553659439087, + 0.012951410375535488, + 0.007025078404694796, + 0.013418416492640972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0067454250529408455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638158440589905 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638158440589905, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638158440589905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05869037285447121, + 0.08645045757293701, + 0.08747873455286026, + 0.19411177933216095, + 0.06478139013051987, + 0.12339088320732117, + 0.034448590129613876, + 0.13939492404460907, + 0.06953534483909607, + 0.14171750843524933 + ], + "credit_map": [ + 0.005862653721123934, + 0.007903161458671093, + 0.008613119833171368, + 0.00592783885076642, + 0.00987257156521082, + 0.0076543898321688175, + 0.007460699416697025, + 0.00539593817666173, + 0.012935148552060127, + 0.012148471549153328, + 0.01678602024912834, + 0.00747802946716547, + 0.00783359445631504, + 0.012740221805870533, + 0.007192710414528847, + 0.009336262941360474, + 0.012552146799862385, + 0.010815251618623734, + 0.021367933601140976, + 0.005291970446705818, + 0.012502850033342838, + 0.00809407327324152, + 0.009547309949994087, + 0.007307202555239201, + 0.03298814222216606, + 0.020502721890807152, + 0.03474061191082001, + 0.01471224520355463, + 0.021954108029603958, + 0.029639117419719696, + 0.01430538296699524, + 0.02526944689452648, + 0.008438313379883766, + 0.008069475181400776, + 0.010083317756652832, + 0.005738004110753536, + 0.010258138179779053, + 0.008324316702783108, + 0.0071078757755458355, + 0.006761949043720961, + 0.02069944143295288, + 0.023242101073265076, + 0.022818343713879585, + 0.012495134025812149, + 0.012606792151927948, + 0.012031074613332748, + 0.006496154237538576, + 0.013001841492950916, + 0.0052637141197919846, + 0.0026482725515961647, + 0.004563836380839348, + 0.005658642388880253, + 0.005930610001087189, + 0.0039083026349544525, + 0.0030604263301938772, + 0.003414785722270608, + 0.02092057280242443, + 0.014430198818445206, + 0.037985481321811676, + 0.023771170526742935, + 0.01062889862805605, + 0.00844138115644455, + 0.011422044597566128, + 0.01179517526179552, + 0.00923800840973854, + 0.010884957388043404, + 0.005668295547366142, + 0.008835519663989544, + 0.014641745947301388, + 0.0066248276270926, + 0.006840473972260952, + 0.006801517680287361, + 0.016824547201395035, + 0.01676485314965248, + 0.02549969032406807, + 0.01223537977784872, + 0.022760862484574318, + 0.01784074865281582, + 0.010290469042956829, + 0.019500959664583206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.008638158440589905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061915084719658 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061915084719658, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061915084719658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044004619121551514, + 0.16450731456279755, + 0.028039192780852318, + 0.05503205582499504, + 0.06588968634605408, + 0.10622764378786087, + 0.057962823659181595, + 0.17785848677158356, + 0.09427516907453537, + 0.2062029093503952 + ], + "credit_map": [ + 0.004409563262015581, + 0.005077788140624762, + 0.005715425591915846, + 0.005661440547555685, + 0.004942373838275671, + 0.0047010877169668674, + 0.010246927849948406, + 0.0032500221859663725, + 0.023839443922042847, + 0.021503502503037453, + 0.029825551435351372, + 0.017786908894777298, + 0.015767212957143784, + 0.015135955065488815, + 0.021024154499173164, + 0.019624613225460052, + 0.003959388006478548, + 0.0028347496408969164, + 0.007236932870000601, + 0.002290119417011738, + 0.004284228663891554, + 0.002027951180934906, + 0.0032287989743053913, + 0.0021770307794213295, + 0.00724779162555933, + 0.004098960198462009, + 0.019950684159994125, + 0.00226351385936141, + 0.0020705261267721653, + 0.0067199598997831345, + 0.0030765209812670946, + 0.00960411038249731, + 0.007983940653502941, + 0.008273568004369736, + 0.00832896027714014, + 0.006458777002990246, + 0.009902223013341427, + 0.006827710662037134, + 0.009718178771436214, + 0.00839634332805872, + 0.018589049577713013, + 0.02207796275615692, + 0.019924316555261612, + 0.010915524326264858, + 0.010378929786384106, + 0.009763008914887905, + 0.0059667835012078285, + 0.008612091653048992, + 0.011401264928281307, + 0.004473146051168442, + 0.016047479584813118, + 0.009577127173542976, + 0.006826591212302446, + 0.00410045450553298, + 0.0032842159271240234, + 0.002252552891150117, + 0.027071500197052956, + 0.02239394746720791, + 0.04635801911354065, + 0.028500095009803772, + 0.01592567190527916, + 0.010837809182703495, + 0.012835975736379623, + 0.013935523107647896, + 0.012158608995378017, + 0.014894559048116207, + 0.010057272389531136, + 0.011006399989128113, + 0.01674707420170307, + 0.009482362307608128, + 0.010706083849072456, + 0.009222827851772308, + 0.025059448555111885, + 0.025451241061091423, + 0.033175595104694366, + 0.020936431363224983, + 0.03443950414657593, + 0.02308577671647072, + 0.017817508429288864, + 0.02623743750154972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.016061915084719658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732873618602753 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732873618602753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732873618602753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06023797392845154, + 0.09900821000337601, + 0.02841865085065365, + 0.07586408406496048, + 0.03752123564481735, + 0.09928368777036667, + 0.1413268744945526, + 0.16774247586727142, + 0.10241585224866867, + 0.18818098306655884 + ], + "credit_map": [ + 0.006375856697559357, + 0.008987209759652615, + 0.00805945135653019, + 0.00651263352483511, + 0.007990246638655663, + 0.005906644277274609, + 0.010145592503249645, + 0.0062603410333395, + 0.017361680045723915, + 0.013240509666502476, + 0.021322177723050117, + 0.01258738525211811, + 0.0065415301360189915, + 0.009621740318834782, + 0.009707202203571796, + 0.008625986985862255, + 0.004431520123034716, + 0.0022644337732344866, + 0.007390493527054787, + 0.002272701356559992, + 0.004065096378326416, + 0.0025100442580878735, + 0.003238854231312871, + 0.0022455055732280016, + 0.0072541618719697, + 0.005644618533551693, + 0.026782242581248283, + 0.0033831687178462744, + 0.004351507406681776, + 0.010969382710754871, + 0.0038194931112229824, + 0.013659512624144554, + 0.004756615962833166, + 0.00289828609675169, + 0.010241680778563023, + 0.0032255372498184443, + 0.002955770120024681, + 0.005425743758678436, + 0.00300695956684649, + 0.005010642576962709, + 0.017670467495918274, + 0.016899481415748596, + 0.021730808541178703, + 0.008959279395639896, + 0.007394275162369013, + 0.012113194912672043, + 0.00694860378280282, + 0.007567575667053461, + 0.02505837380886078, + 0.014249826781451702, + 0.026956096291542053, + 0.025851448997855186, + 0.015204905532300472, + 0.010387915186583996, + 0.015544893220067024, + 0.008073423989117146, + 0.024402597919106483, + 0.022891556844115257, + 0.02826198749244213, + 0.025249890983104706, + 0.016140535473823547, + 0.020344162359833717, + 0.017315521836280823, + 0.01313621923327446, + 0.011067678220570087, + 0.018020205199718475, + 0.015242787078022957, + 0.008974654600024223, + 0.01844784803688526, + 0.012387881055474281, + 0.00889566820114851, + 0.009379124268889427, + 0.030839018523693085, + 0.019190888851881027, + 0.03521411865949631, + 0.01712336763739586, + 0.027095921337604523, + 0.025103125721216202, + 0.012447264045476913, + 0.021167291328310966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.006732873618602753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-long shot, captured from above, imbues the image with the texture of an ancient ink painting. The Monkey King, clad in dazzling golden chain mail, dominates the frame. Atop his head rests a phoenix-winged purple-gold crown, two long pheasant feathers trembling slightly with the force of his charged movement. His face is defiant, his fiery eyes gleaming with determination. He thrusts his body into a lunge, muscles flexing. He grips the heavy, jet-black Ruyi Jingu Club in both hands, raising it high above his head. With unstoppable force, he smashes it down upon a towering celestial mountain in the background, suspended above a sea of clouds. The instant the iron rod makes contact with the mountain's summit, the solid indigo-colored mass of the mountain cracks, sending the massive rock shattering into countless fragments. Most strikingly, the dust stirred up isn't just ordinary smoke, but countless splashing, splattering strokes of thick ink, condensing and flowing through the air with a powerful sense of movement and the rhythm of calligraphy. The main colors of the entire picture are composed of black, white and gray in varying shades. The only bright colors are the golden armor on the Monkey King and the fire in his eyes. The light spreads naturally on the rice paper, creating a unique visual effect of power and aesthetics. It is both violent and full of the unique freehand and ethereal beauty of ink painting. The atmosphere is solemn and extremely impactful.", + "seed": 857, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_2/index_6483_seed_857_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": -0.0094373163767159, + "vlm_logprob": -0.0094373163767159, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.0094373163767159, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021299434825778008 + ], + "opposite_logprob_tokens": [ + -3.896299362182617 + ], + "logprob_reward": -0.021299434825778008, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021299434825778008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1769525557756424, + 0.1594070941209793, + 0.04306292161345482, + 0.0560026690363884, + 0.041626207530498505, + 0.0804300308227539, + 0.12106940895318985, + 0.19288823008537292, + 0.06986973434686661, + 0.058691058307886124 + ], + "credit_map": [ + 0.025911077857017517, + 0.03160469979047775, + 0.0213758647441864, + 0.020831910893321037, + 0.015316387638449669, + 0.020791495218873024, + 0.022016616538167, + 0.019104503095149994, + 0.024606602266430855, + 0.03477175533771515, + 0.015425866469740868, + 0.01596292480826378, + 0.020130867138504982, + 0.019214525818824768, + 0.014354677870869637, + 0.014939873479306698, + 0.007434759754687548, + 0.006960654631257057, + 0.004933919757604599, + 0.006133805029094219, + 0.004824418108910322, + 0.0051220012828707695, + 0.004190807230770588, + 0.0034625555854290724, + 0.00928721483796835, + 0.010881181806325912, + 0.006610431708395481, + 0.006766046397387981, + 0.005629702936857939, + 0.006458011455833912, + 0.00507351104170084, + 0.005296570248901844, + 0.007049342151731253, + 0.0075057754293084145, + 0.0048217750154435635, + 0.00571351358667016, + 0.004662373103201389, + 0.00394676998257637, + 0.00430663675069809, + 0.0036200250033289194, + 0.01259161252528429, + 0.012853218242526054, + 0.009096385911107063, + 0.010827993042767048, + 0.007800310384482145, + 0.011709907092154026, + 0.006959591526538134, + 0.008591015823185444, + 0.018148018047213554, + 0.008479008451104164, + 0.012376507744193077, + 0.013147306628525257, + 0.013437755405902863, + 0.027586542069911957, + 0.010899989865720272, + 0.016994278877973557, + 0.02754536084830761, + 0.019267013296484947, + 0.021113984286785126, + 0.019815538078546524, + 0.02919868752360344, + 0.03669881820678711, + 0.017495928332209587, + 0.02175290323793888, + 0.022769222036004066, + 0.005267021711915731, + 0.004237511660903692, + 0.004377071280032396, + 0.005740074440836906, + 0.014583022333681583, + 0.0036207051016390324, + 0.009275108575820923, + 0.010142598301172256, + 0.009182289242744446, + 0.005447680596262217, + 0.004433431196957827, + 0.008707494474947453, + 0.006773253437131643, + 0.004275562707334757, + 0.009728753007948399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.021299434825778008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746135652065277 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746135652065277, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746135652065277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16348694264888763, + 0.1066688597202301, + 0.07959982752799988, + 0.08647128939628601, + 0.0713147222995758, + 0.03986843302845955, + 0.13010826706886292, + 0.090666763484478, + 0.11482502520084381, + 0.11698990315198898 + ], + "credit_map": [ + 0.02945542149245739, + 0.021227296441793442, + 0.019251013174653053, + 0.015064365230500698, + 0.016814570873975754, + 0.02481216937303543, + 0.01678299345076084, + 0.020079102367162704, + 0.020606838166713715, + 0.011028767563402653, + 0.02316264994442463, + 0.009100410155951977, + 0.008816428482532501, + 0.013125387951731682, + 0.008024380542337894, + 0.012803997844457626, + 0.012175001204013824, + 0.015986984595656395, + 0.006759743671864271, + 0.00950776506215334, + 0.00973488762974739, + 0.01224811002612114, + 0.006592143792659044, + 0.006595193408429623, + 0.014393361285328865, + 0.01618083193898201, + 0.008989658206701279, + 0.011217836290597916, + 0.00792781449854374, + 0.010913974605500698, + 0.007007418666034937, + 0.009840397164225578, + 0.010983733460307121, + 0.013121378608047962, + 0.005982309579849243, + 0.009343394078314304, + 0.009645433165133, + 0.00980055145919323, + 0.007704001385718584, + 0.004733919631689787, + 0.0065680621191859245, + 0.0047391252592206, + 0.004837602376937866, + 0.00504068611189723, + 0.003356396919116378, + 0.007459437940269709, + 0.0036459381226450205, + 0.004221183713525534, + 0.02268074080348015, + 0.008410116657614708, + 0.010701156221330166, + 0.010026662610471249, + 0.013801888562738895, + 0.037870317697525024, + 0.009860515594482422, + 0.01675686612725258, + 0.010519266128540039, + 0.012396405450999737, + 0.009614059701561928, + 0.00847694929689169, + 0.012984177097678185, + 0.013020764105021954, + 0.0072959852404892445, + 0.01635916158556938, + 0.025266779586672783, + 0.012487993575632572, + 0.008297796361148357, + 0.008412450551986694, + 0.013455996289849281, + 0.019865894690155983, + 0.008146925829350948, + 0.018891192972660065, + 0.018499312922358513, + 0.021862322464585304, + 0.011734510771930218, + 0.012218128889799118, + 0.0127916494384408, + 0.01709567755460739, + 0.009591725654900074, + 0.013196577318012714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.006746135652065277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254624877125025 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254624877125025, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254624877125025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08113202452659607, + 0.08762636035680771, + 0.07115273177623749, + 0.05283327028155327, + 0.030888503417372704, + 0.04555905982851982, + 0.14696857333183289, + 0.08982240408658981, + 0.19083091616630554, + 0.20318613946437836 + ], + "credit_map": [ + 0.01104025449603796, + 0.010474060662090778, + 0.009783189743757248, + 0.008943254128098488, + 0.009014418348670006, + 0.010674471035599709, + 0.009160449728369713, + 0.012041923590004444, + 0.0162955429404974, + 0.007577001582831144, + 0.020387979224324226, + 0.009234311990439892, + 0.005577947478741407, + 0.009585179388523102, + 0.006928835529834032, + 0.012039566412568092, + 0.01167584303766489, + 0.01029238197952509, + 0.008478282019495964, + 0.010666333138942719, + 0.006515129003673792, + 0.010624747723340988, + 0.007285090163350105, + 0.005614925641566515, + 0.009474252350628376, + 0.006701847538352013, + 0.005577876232564449, + 0.005714298225939274, + 0.0050928834825754166, + 0.009339449927210808, + 0.004359706770628691, + 0.006572953891009092, + 0.00457652797922492, + 0.004332615528255701, + 0.0037168392445892096, + 0.003525194711983204, + 0.0033848981838673353, + 0.005074715707451105, + 0.0026482779067009687, + 0.0036294348537921906, + 0.007711993996053934, + 0.005044753663241863, + 0.006669568829238415, + 0.004743140656501055, + 0.0035970911849290133, + 0.00854529533535242, + 0.0032146903686225414, + 0.006032524164766073, + 0.025267168879508972, + 0.00913337804377079, + 0.013027658686041832, + 0.010851589031517506, + 0.01573568396270275, + 0.04149707779288292, + 0.01054530031979084, + 0.02091071382164955, + 0.012294224463403225, + 0.008199417963624, + 0.008117417804896832, + 0.006695926655083895, + 0.015720468014478683, + 0.01797974295914173, + 0.008182117715477943, + 0.012633088044822216, + 0.03946875035762787, + 0.023941384628415108, + 0.014049016870558262, + 0.014473719522356987, + 0.023766154423356056, + 0.029635505750775337, + 0.012556727975606918, + 0.03293965011835098, + 0.0308955367654562, + 0.038819361478090286, + 0.01900343969464302, + 0.022057194262742996, + 0.024435101076960564, + 0.029255583882331848, + 0.01287419255822897, + 0.02584572695195675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.005254624877125025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636267855763435 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636267855763435, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636267855763435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09520669281482697, + 0.18379148840904236, + 0.04206928610801697, + 0.02889852039515972, + 0.03351059928536415, + 0.05516017600893974, + 0.22206233441829681, + 0.12032297253608704, + 0.16056936979293823, + 0.05840851366519928 + ], + "credit_map": [ + 0.014622603543102741, + 0.015737980604171753, + 0.013501765206456184, + 0.012812454253435135, + 0.008683191612362862, + 0.00921240821480751, + 0.011444955132901669, + 0.009191329590976238, + 0.03341172635555267, + 0.017127400264143944, + 0.04410846158862114, + 0.016805006191134453, + 0.012312356382608414, + 0.018245607614517212, + 0.013365073129534721, + 0.028415868058800697, + 0.007757797371596098, + 0.00473039923235774, + 0.005032207351177931, + 0.004269949626177549, + 0.003792697796598077, + 0.008200148120522499, + 0.00335537432692945, + 0.00493071461096406, + 0.0048010079190135, + 0.0035594927612692118, + 0.0030500113498419523, + 0.0039050295017659664, + 0.0022683439310640097, + 0.004597888328135014, + 0.0027745780535042286, + 0.003942169714719057, + 0.005318344105035067, + 0.004828267730772495, + 0.0052138278260827065, + 0.0032992581836879253, + 0.0033848320599645376, + 0.004597666673362255, + 0.002770451596006751, + 0.004097952041774988, + 0.00924439076334238, + 0.004792994819581509, + 0.01095358282327652, + 0.007373095024377108, + 0.0038655323442071676, + 0.007702932693064213, + 0.004772965330630541, + 0.006454682908952236, + 0.039080481976270676, + 0.016695871949195862, + 0.02306828647851944, + 0.01868755929172039, + 0.02243036776781082, + 0.05627766251564026, + 0.018984220921993256, + 0.026837892830371857, + 0.017228838056325912, + 0.01013875100761652, + 0.009595249779522419, + 0.009051217697560787, + 0.01862020045518875, + 0.026911133900284767, + 0.00948298443108797, + 0.019294602796435356, + 0.04032327979803085, + 0.018121208995580673, + 0.011099174618721008, + 0.011281789280474186, + 0.018457558006048203, + 0.025466438382864, + 0.00979229249060154, + 0.026027629151940346, + 0.008993768133223057, + 0.011693269945681095, + 0.005961437243968248, + 0.005611622706055641, + 0.006692739203572273, + 0.008547498844563961, + 0.003933355212211609, + 0.006974824238568544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.008636267855763435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250118672847748 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005250118672847748, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250118672847748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07551651448011398, + 0.11326891928911209, + 0.09117989242076874, + 0.04429975897073746, + 0.0418228842318058, + 0.12833762168884277, + 0.16144873201847076, + 0.24674196541309357, + 0.06729307770729065, + 0.030090663582086563 + ], + "credit_map": [ + 0.012157846242189407, + 0.008144575171172619, + 0.014224894344806671, + 0.009586120024323463, + 0.005710827186703682, + 0.007919948548078537, + 0.009143073111772537, + 0.008629227057099342, + 0.019503986462950706, + 0.007554033771157265, + 0.02223285660147667, + 0.010858476161956787, + 0.006750038824975491, + 0.018524879589676857, + 0.010043005459010601, + 0.01780163124203682, + 0.01590612530708313, + 0.014589933678507805, + 0.011578626930713654, + 0.011420326307415962, + 0.00941321812570095, + 0.011819795705378056, + 0.007419494446367025, + 0.009032361209392548, + 0.007879790849983692, + 0.006045980378985405, + 0.005479532293975353, + 0.005162013228982687, + 0.00361272064037621, + 0.007760168984532356, + 0.0036179921589791775, + 0.004741554148495197, + 0.00657713133841753, + 0.0036850401666015387, + 0.0060048773884773254, + 0.004457306582480669, + 0.0035908750724047422, + 0.008108975365757942, + 0.0037901706527918577, + 0.005608501844108105, + 0.021077638491988182, + 0.014337478205561638, + 0.022533250972628593, + 0.016193443909287453, + 0.010004871524870396, + 0.016725152730941772, + 0.010219923220574856, + 0.017245851457118988, + 0.025116974487900734, + 0.015685372054576874, + 0.01691669411957264, + 0.014008019119501114, + 0.021718336269259453, + 0.024763934314250946, + 0.016223149374127388, + 0.027016233652830124, + 0.0351409949362278, + 0.018683018162846565, + 0.019139911979436874, + 0.018361322581768036, + 0.037872523069381714, + 0.05162070319056511, + 0.017523180693387985, + 0.04840027540922165, + 0.009504799731075764, + 0.008200805634260178, + 0.007989177480340004, + 0.0070077236741781235, + 0.008038838393986225, + 0.00918085128068924, + 0.004514405038207769, + 0.012856471352279186, + 0.005128195974975824, + 0.0036495572421699762, + 0.0035187057219445705, + 0.002819895511493087, + 0.003991666715592146, + 0.005008294712752104, + 0.002474261447787285, + 0.003500081831589341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.005250118672847748 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-long shot, captured from above, imbues the image with the texture of an ancient ink painting. The Monkey King, clad in dazzling golden chain mail, dominates the frame. Atop his head rests a phoenix-winged purple-gold crown, two long pheasant feathers trembling slightly with the force of his charged movement. His face is defiant, his fiery eyes gleaming with determination. He thrusts his body into a lunge, muscles flexing. He grips the heavy, jet-black Ruyi Jingu Club in both hands, raising it high above his head. With unstoppable force, he smashes it down upon a towering celestial mountain in the background, suspended above a sea of clouds. The instant the iron rod makes contact with the mountain's summit, the solid indigo-colored mass of the mountain cracks, sending the massive rock shattering into countless fragments. Most strikingly, the dust stirred up isn't just ordinary smoke, but countless splashing, splattering strokes of thick ink, condensing and flowing through the air with a powerful sense of movement and the rhythm of calligraphy. The main colors of the entire picture are composed of black, white and gray in varying shades. The only bright colors are the golden armor on the Monkey King and the fire in his eyes. The light spreads naturally on the rice paper, creating a unique visual effect of power and aesthetics. It is both violent and full of the unique freehand and ethereal beauty of ink painting. The atmosphere is solemn and extremely impactful.", + "seed": 858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_2/index_6483_seed_858_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": -0.012381019257009029, + "vlm_logprob": -0.012381019257009029, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.012381019257009029, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010839510709047318 + ], + "opposite_logprob_tokens": [ + -4.760839462280273 + ], + "logprob_reward": -0.010839510709047318, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010839510709047318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10445213317871094, + 0.10893980413675308, + 0.16451479494571686, + 0.09273884445428848, + 0.11011555045843124, + 0.18543480336666107, + 0.06181563809514046, + 0.06409937888383865, + 0.05567255988717079, + 0.052216529846191406 + ], + "credit_map": [ + 0.015695473179221153, + 0.016422245651483536, + 0.010529021732509136, + 0.01338169351220131, + 0.011918807402253151, + 0.012710513547062874, + 0.012788607738912106, + 0.01100577600300312, + 0.016360508278012276, + 0.018513374030590057, + 0.007804317865520716, + 0.010867509990930557, + 0.016455208882689476, + 0.014220131561160088, + 0.011247162707149982, + 0.013471594080328941, + 0.02390388585627079, + 0.02499580569565296, + 0.02082793228328228, + 0.010646033100783825, + 0.02287133038043976, + 0.024283643811941147, + 0.0199513491243124, + 0.017034821212291718, + 0.008713417686522007, + 0.009101569652557373, + 0.011513618752360344, + 0.008120754733681679, + 0.01460731029510498, + 0.013150099664926529, + 0.010549255646765232, + 0.01698281243443489, + 0.015073440968990326, + 0.012676752172410488, + 0.012572366744279861, + 0.014110976830124855, + 0.014970926567912102, + 0.014047338627278805, + 0.008335002698004246, + 0.01832873374223709, + 0.03416965901851654, + 0.023347392678260803, + 0.024531884118914604, + 0.014424671418964863, + 0.020714495331048965, + 0.023191101849079132, + 0.010589664801955223, + 0.03446591645479202, + 0.010984349064528942, + 0.00688924128189683, + 0.008190919645130634, + 0.004579724743962288, + 0.007353535387665033, + 0.007617081515491009, + 0.006072481162846088, + 0.010128302499651909, + 0.010149749927222729, + 0.006862784270197153, + 0.007939638569951057, + 0.005768631584942341, + 0.007221782114356756, + 0.007823647931218147, + 0.008678100071847439, + 0.00965504627674818, + 0.008557872846722603, + 0.006199805997312069, + 0.009741096757352352, + 0.005246609915047884, + 0.005233601201325655, + 0.006686692126095295, + 0.005109305493533611, + 0.008897579275071621, + 0.007719750050455332, + 0.007418676279485226, + 0.006180906668305397, + 0.0051367757841944695, + 0.005486026406288147, + 0.006792434491217136, + 0.0054217190481722355, + 0.00806023832410574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.010839510709047318 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979557354003191 + ], + "opposite_logprob_tokens": [ + -4.634795665740967 + ], + "logprob_reward": -0.00979557354003191, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979557354003191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19574417173862457, + 0.2073846310377121, + 0.05439486354589462, + 0.0655144676566124, + 0.08414202183485031, + 0.158200204372406, + 0.09541664272546768, + 0.0507948100566864, + 0.061334192752838135, + 0.027073929086327553 + ], + "credit_map": [ + 0.02140967920422554, + 0.045809078961610794, + 0.015524103306233883, + 0.025046924129128456, + 0.021645495668053627, + 0.030724534764885902, + 0.021221213042736053, + 0.014363170601427555, + 0.027359750121831894, + 0.061739012598991394, + 0.01578240655362606, + 0.016356119886040688, + 0.0301235169172287, + 0.02065921574831009, + 0.017133411020040512, + 0.018231205642223358, + 0.005408369470387697, + 0.010860641486942768, + 0.00445840647444129, + 0.004393736366182566, + 0.010492381639778614, + 0.007825554348528385, + 0.006684885360300541, + 0.0042708925902843475, + 0.013902062550187111, + 0.007702636998146772, + 0.01002892293035984, + 0.0046739885583519936, + 0.005103910341858864, + 0.00801822915673256, + 0.0033431455958634615, + 0.012741577811539173, + 0.010011526755988598, + 0.010455913841724396, + 0.010471728630363941, + 0.009169397875666618, + 0.017715975642204285, + 0.00807261560112238, + 0.0072325156070292, + 0.011012360453605652, + 0.030756771564483643, + 0.021817421540617943, + 0.01866183802485466, + 0.0130116306245327, + 0.016614001244306564, + 0.019572939723730087, + 0.010715569369494915, + 0.027050049975514412, + 0.01599077135324478, + 0.011357815936207771, + 0.011894050985574722, + 0.007451164070516825, + 0.010960743762552738, + 0.011792204342782497, + 0.008907286450266838, + 0.017062615603208542, + 0.00859767571091652, + 0.005401065107434988, + 0.007704014424234629, + 0.005162007175385952, + 0.005564835853874683, + 0.00611467519775033, + 0.004451811779290438, + 0.00779873039573431, + 0.010228254832327366, + 0.008302722126245499, + 0.008712787181138992, + 0.005066576413810253, + 0.006778204347938299, + 0.00695516262203455, + 0.00584506755694747, + 0.009445424191653728, + 0.004764792509377003, + 0.003114401362836361, + 0.004377241712063551, + 0.0025068456307053566, + 0.002681742887943983, + 0.003495629644021392, + 0.0023975353688001633, + 0.003735743463039398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.00979557354003191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639456048607826 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639456048607826, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639456048607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11521691083908081, + 0.08340077102184296, + 0.08567456901073456, + 0.1913437843322754, + 0.1372123807668686, + 0.09227929264307022, + 0.12373483926057816, + 0.07179265469312668, + 0.06677327305078506, + 0.032571546733379364 + ], + "credit_map": [ + 0.024784356355667114, + 0.012383817695081234, + 0.012246651574969292, + 0.010412270203232765, + 0.01009932067245245, + 0.016512714326381683, + 0.017614122480154037, + 0.011163654737174511, + 0.011191281490027905, + 0.01356751099228859, + 0.008390557020902634, + 0.007981576025485992, + 0.01301544439047575, + 0.010635048151016235, + 0.008086953312158585, + 0.010532398708164692, + 0.01749592460691929, + 0.010214348323643208, + 0.009928866289556026, + 0.00632889149710536, + 0.011262753047049046, + 0.013698744587600231, + 0.006992977112531662, + 0.009752069599926472, + 0.031247351318597794, + 0.01657150872051716, + 0.017596779391169548, + 0.011203479021787643, + 0.026610445231199265, + 0.02783365175127983, + 0.01699100248515606, + 0.04328956827521324, + 0.016380084678530693, + 0.022683050483465195, + 0.01057105977088213, + 0.010801518335938454, + 0.03331296890974045, + 0.013795379549264908, + 0.01114241685718298, + 0.018525898456573486, + 0.0156382005661726, + 0.01596747897565365, + 0.010364288464188576, + 0.006298947613686323, + 0.013620768673717976, + 0.01027526892721653, + 0.007767992559820414, + 0.012346349656581879, + 0.020866738632321358, + 0.0166932325810194, + 0.01166753750294447, + 0.008065205998718739, + 0.017446480691432953, + 0.015175397507846355, + 0.012011955492198467, + 0.021808290854096413, + 0.012524811550974846, + 0.00917449127882719, + 0.0068741156719625, + 0.00580793060362339, + 0.007931233383715153, + 0.008509315550327301, + 0.006583554670214653, + 0.014387198723852634, + 0.011986598372459412, + 0.00992525927722454, + 0.006833038758486509, + 0.00574887590482831, + 0.0073049645870924, + 0.007681078277528286, + 0.00614133570343256, + 0.011152125895023346, + 0.005380578804761171, + 0.003946525044739246, + 0.0039026655722409487, + 0.003473070915788412, + 0.0037061322946101427, + 0.004064866807311773, + 0.002931274939328432, + 0.005166433751583099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.004639456048607826 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.06993217766284943, + 0.06303340941667557, + 0.0731123760342598, + 0.051432251930236816, + 0.3795347213745117, + 0.09934937953948975, + 0.06380367279052734, + 0.10447614639997482, + 0.033131252974271774, + 0.06219465658068657 + ], + "credit_map": [ + 0.011120929382741451, + 0.009482353925704956, + 0.007671286817640066, + 0.008368671871721745, + 0.006143337581306696, + 0.00928921066224575, + 0.009892266243696213, + 0.007964113727211952, + 0.006052518729120493, + 0.009018138982355595, + 0.010109543800354004, + 0.006955199874937534, + 0.007756272330880165, + 0.007416264619678259, + 0.006616083439439535, + 0.00910938624292612, + 0.01344561018049717, + 0.009231246076524258, + 0.008417602628469467, + 0.006031690631061792, + 0.01038177590817213, + 0.01150091364979744, + 0.006562782917171717, + 0.007540747988969088, + 0.01002553477883339, + 0.005951534956693649, + 0.0071698264218866825, + 0.003889137879014015, + 0.006143964361399412, + 0.005947698373347521, + 0.004647829569876194, + 0.0076567200012505054, + 0.05435327813029289, + 0.05871501564979553, + 0.021908815950155258, + 0.026818804442882538, + 0.09517574310302734, + 0.04152529314160347, + 0.02701180800795555, + 0.054025936871767044, + 0.015336243435740471, + 0.0148435914888978, + 0.009152021259069443, + 0.010061589069664478, + 0.013772211968898773, + 0.010189422406256199, + 0.00757330609485507, + 0.018420979380607605, + 0.008019626140594482, + 0.008743184618651867, + 0.00676595838740468, + 0.006308638025075197, + 0.012206570245325565, + 0.006422138772904873, + 0.008794206194579601, + 0.006543345749378204, + 0.013102428987622261, + 0.013642698526382446, + 0.009950720705091953, + 0.009569208137691021, + 0.011815385892987251, + 0.012105227448046207, + 0.014126451686024666, + 0.020164024084806442, + 0.004906852263957262, + 0.004392971750348806, + 0.004928120411932468, + 0.003172658383846283, + 0.004487053956836462, + 0.0037326952442526817, + 0.0031385861802846193, + 0.004372309893369675, + 0.008052521385252476, + 0.01108565367758274, + 0.0066014183685183525, + 0.006944551598280668, + 0.009641429409384727, + 0.006406573578715324, + 0.005959631409496069, + 0.007502875290811062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.016060978174209595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205695778131485 + ], + "opposite_logprob_tokens": [ + -3.8955695629119873 + ], + "logprob_reward": -0.0205695778131485, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205695778131485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19385378062725067, + 0.07907973229885101, + 0.06645743548870087, + 0.07491692155599594, + 0.10386182367801666, + 0.1834968775510788, + 0.0607113316655159, + 0.08372905850410461, + 0.04581960663199425, + 0.10807345807552338 + ], + "credit_map": [ + 0.0436839796602726, + 0.024740593507885933, + 0.016684751957654953, + 0.023376675322651863, + 0.014628703705966473, + 0.02753276377916336, + 0.02072225697338581, + 0.02248404733836651, + 0.00823995377868414, + 0.014270555227994919, + 0.007723413407802582, + 0.007096298038959503, + 0.010146587155759335, + 0.011924214661121368, + 0.00828747358173132, + 0.011391236446797848, + 0.010724944993853569, + 0.005742968525737524, + 0.00803209375590086, + 0.005373787600547075, + 0.008835007436573505, + 0.013340210542082787, + 0.006903446279466152, + 0.007504972163587809, + 0.011203239671885967, + 0.0079239122569561, + 0.009089355356991291, + 0.004410380031913519, + 0.011459866538643837, + 0.01179070957005024, + 0.007126549258828163, + 0.011912906542420387, + 0.013998596929013729, + 0.015215054154396057, + 0.007502778433263302, + 0.008384470827877522, + 0.02271207794547081, + 0.011950800195336342, + 0.008448874577879906, + 0.01564917340874672, + 0.0331699475646019, + 0.026520997285842896, + 0.011310930363833904, + 0.013177932240068913, + 0.0236097052693367, + 0.026509666815400124, + 0.011027186177670956, + 0.038170523941516876, + 0.01161011029034853, + 0.005377134308218956, + 0.009796572849154472, + 0.005856488831341267, + 0.004908852744847536, + 0.007338675204664469, + 0.00527718011289835, + 0.010546315461397171, + 0.01070254947990179, + 0.009988255798816681, + 0.00958563294261694, + 0.009211424738168716, + 0.008869890123605728, + 0.009267494082450867, + 0.011074208654463291, + 0.0150296064093709, + 0.006247878540307283, + 0.006329359952360392, + 0.0035947251599282026, + 0.004317558836191893, + 0.006352625321596861, + 0.005480397958308458, + 0.004500843584537506, + 0.008996216580271721, + 0.0164320208132267, + 0.015459316782653332, + 0.009697486646473408, + 0.010415499098598957, + 0.01376877911388874, + 0.012743753381073475, + 0.009558352641761303, + 0.019998246803879738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0205695778131485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-long shot, captured from above, imbues the image with the texture of an ancient ink painting. The Monkey King, clad in dazzling golden chain mail, dominates the frame. Atop his head rests a phoenix-winged purple-gold crown, two long pheasant feathers trembling slightly with the force of his charged movement. His face is defiant, his fiery eyes gleaming with determination. He thrusts his body into a lunge, muscles flexing. He grips the heavy, jet-black Ruyi Jingu Club in both hands, raising it high above his head. With unstoppable force, he smashes it down upon a towering celestial mountain in the background, suspended above a sea of clouds. The instant the iron rod makes contact with the mountain's summit, the solid indigo-colored mass of the mountain cracks, sending the massive rock shattering into countless fragments. Most strikingly, the dust stirred up isn't just ordinary smoke, but countless splashing, splattering strokes of thick ink, condensing and flowing through the air with a powerful sense of movement and the rhythm of calligraphy. The main colors of the entire picture are composed of black, white and gray in varying shades. The only bright colors are the golden armor on the Monkey King and the fire in his eyes. The light spreads naturally on the rice paper, creating a unique visual effect of power and aesthetics. It is both violent and full of the unique freehand and ethereal beauty of ink painting. The atmosphere is solemn and extremely impactful.", + "seed": 859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_2/index_6483_seed_859_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": -0.007442759349942207, + "vlm_logprob": -0.007442759349942207, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.007442759349942207, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012119228951632977 + ], + "opposite_logprob_tokens": [ + -4.512119293212891 + ], + "logprob_reward": -0.012119228951632977, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012119228951632977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13891451060771942, + 0.08278866857290268, + 0.040681391954422, + 0.019108884036540985, + 0.04541865363717079, + 0.07823929935693741, + 0.13405172526836395, + 0.3924400210380554, + 0.04478910192847252, + 0.023567738011479378 + ], + "credit_map": [ + 0.013571169227361679, + 0.012655842117965221, + 0.012489397078752518, + 0.02320384979248047, + 0.024569816887378693, + 0.021683992817997932, + 0.014254184439778328, + 0.0164862722158432, + 0.00597381079569459, + 0.008767986670136452, + 0.004440920893102884, + 0.013041824102401733, + 0.016876928508281708, + 0.013766049407422543, + 0.007650616578757763, + 0.012270537205040455, + 0.0026919946540147066, + 0.0018415154190734029, + 0.0017151960637420416, + 0.0019310485804453492, + 0.004799434915184975, + 0.012265044264495373, + 0.0026658351998776197, + 0.01277132797986269, + 0.0016146778361871839, + 0.001467003719881177, + 0.0017276658909395337, + 0.0014135494129732251, + 0.0030672005377709866, + 0.005349242128431797, + 0.0019474896835163236, + 0.002522056456655264, + 0.004187725018709898, + 0.003075618762522936, + 0.0026163707952946424, + 0.004603683017194271, + 0.0052857971750199795, + 0.010670332238078117, + 0.0035237243864685297, + 0.011455406434834003, + 0.013554874807596207, + 0.007369589060544968, + 0.00985319446772337, + 0.008050135336816311, + 0.008042339235544205, + 0.009007226675748825, + 0.008322959765791893, + 0.014038985595107079, + 0.04129314795136452, + 0.007269310764968395, + 0.04610779881477356, + 0.009722338058054447, + 0.007217041216790676, + 0.007265531923621893, + 0.008172654546797276, + 0.0070039029233157635, + 0.1135811060667038, + 0.07564053684473038, + 0.07934895157814026, + 0.01841939240694046, + 0.029832568019628525, + 0.03578699007630348, + 0.02084418386220932, + 0.018986325711011887, + 0.0057836174964904785, + 0.005752035416662693, + 0.0049160984344780445, + 0.0040651122108101845, + 0.007547267712652683, + 0.005678627640008926, + 0.006355582736432552, + 0.004690761212259531, + 0.0037150403950363398, + 0.004101981874555349, + 0.0036441206466406584, + 0.0024116362910717726, + 0.003090134123340249, + 0.0018014076631516218, + 0.0023404741659760475, + 0.002462943783029914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.012119228951632977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958887282758951 + ], + "opposite_logprob_tokens": [ + -5.1309590339660645 + ], + "logprob_reward": -0.005958887282758951, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958887282758951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10522828996181488, + 0.05882672220468521, + 0.46539750695228577, + 0.012777485884726048, + 0.01803099922835827, + 0.03949910029768944, + 0.03858332708477974, + 0.20271927118301392, + 0.025520503520965576, + 0.033416736871004105 + ], + "credit_map": [ + 0.0065907626412808895, + 0.0069803521037101746, + 0.006521976087242365, + 0.007844449020922184, + 0.019480038434267044, + 0.03170869126915932, + 0.007867260836064816, + 0.01823476143181324, + 0.004098489414900541, + 0.0034783179871737957, + 0.003389692632481456, + 0.0066798957996070385, + 0.009262180887162685, + 0.01903289370238781, + 0.005343005992472172, + 0.007542248349636793, + 0.03051547333598137, + 0.03221307322382927, + 0.019799983128905296, + 0.024566175416111946, + 0.08749432861804962, + 0.1184665635228157, + 0.025404613465070724, + 0.12693732976913452, + 0.0012507708743214607, + 0.0012499336153268814, + 0.0007649015751667321, + 0.0014136445242911577, + 0.002994868205860257, + 0.002484995638951659, + 0.0010947788832709193, + 0.001523593789897859, + 0.0011873850598931313, + 0.0012893204111605883, + 0.0009946684585884213, + 0.0014019440859556198, + 0.003705488285049796, + 0.003916224464774132, + 0.0014293601270765066, + 0.004106611479073763, + 0.006075216457247734, + 0.00562904542312026, + 0.003999240230768919, + 0.0026761535555124283, + 0.0062431590631604195, + 0.005691345781087875, + 0.003604958299547434, + 0.005579986143857241, + 0.0064229038543999195, + 0.0031560822390019894, + 0.008436011150479317, + 0.006260812282562256, + 0.004190510138869286, + 0.003450696123763919, + 0.003696569474413991, + 0.0029697432182729244, + 0.05634617432951927, + 0.037561818957328796, + 0.040416307747364044, + 0.010882577858865261, + 0.015134843997657299, + 0.01944906637072563, + 0.012803914956748486, + 0.01012458000332117, + 0.0034368596971035004, + 0.0037807971239089966, + 0.0027763964608311653, + 0.002792028710246086, + 0.00351578276604414, + 0.002806618344038725, + 0.0036596436984837055, + 0.0027523792814463377, + 0.005388027522712946, + 0.0069841379299759865, + 0.0069168987683951855, + 0.002607806585729122, + 0.00351521885022521, + 0.002675932366400957, + 0.0027189787942916155, + 0.0026097409427165985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.005958887282758951 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252964794635773 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252964794635773, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252964794635773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0872996523976326, + 0.040625642985105515, + 0.06028757989406586, + 0.034851085394620895, + 0.06565438956022263, + 0.08609002828598022, + 0.24515433609485626, + 0.18436340987682343, + 0.13562530279159546, + 0.06004849448800087 + ], + "credit_map": [ + 0.005898767616599798, + 0.005350553430616856, + 0.004475550260394812, + 0.007211701478809118, + 0.011181405745446682, + 0.023009557276964188, + 0.005730460863560438, + 0.024441666901111603, + 0.0036813803017139435, + 0.003316283691674471, + 0.003131362609565258, + 0.0048434860073029995, + 0.0052977572195231915, + 0.008388918824493885, + 0.00399702088907361, + 0.007969439961016178, + 0.00444822758436203, + 0.0033241782803088427, + 0.0029233621899038553, + 0.003511663991957903, + 0.007301158271729946, + 0.01856222376227379, + 0.003543067490682006, + 0.016673702746629715, + 0.0025814538821578026, + 0.002787724370136857, + 0.0025769479107111692, + 0.0033679367043077946, + 0.006875154562294483, + 0.006397133693099022, + 0.0036196457222104073, + 0.00664509367197752, + 0.004020777065306902, + 0.004369975067675114, + 0.003027792787179351, + 0.0035437727347016335, + 0.013151685707271099, + 0.016056440770626068, + 0.0038403826765716076, + 0.01764357089996338, + 0.013276008889079094, + 0.010981494560837746, + 0.010047820396721363, + 0.007026059553027153, + 0.008722372353076935, + 0.007697360124439001, + 0.01053127832710743, + 0.017807645723223686, + 0.045950230211019516, + 0.01655845157802105, + 0.06156627833843231, + 0.026839835569262505, + 0.032765381038188934, + 0.020302273333072662, + 0.02214619144797325, + 0.01902572624385357, + 0.05603346973657608, + 0.03441101685166359, + 0.03957189992070198, + 0.011583690531551838, + 0.012326765805482864, + 0.012470971792936325, + 0.00954524241387844, + 0.008420383557677269, + 0.015872877091169357, + 0.012795361690223217, + 0.018890077248215675, + 0.012304785661399364, + 0.01768282987177372, + 0.02122870273888111, + 0.01458610687404871, + 0.022264575585722923, + 0.008278812281787395, + 0.013484460301697254, + 0.012891596183180809, + 0.005368518643081188, + 0.005469117779284716, + 0.00549711287021637, + 0.004777817986905575, + 0.004281067755073309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.005252964794635773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.10100561380386353, + 0.032749854028224945, + 0.10960766673088074, + 0.01976962201297283, + 0.027219044044613838, + 0.14275744557380676, + 0.2139516919851303, + 0.10325571149587631, + 0.08219581842422485, + 0.1674874722957611 + ], + "credit_map": [ + 0.008202267810702324, + 0.007345143705606461, + 0.006726948544383049, + 0.006438005715608597, + 0.017692705616354942, + 0.031032804399728775, + 0.006503689102828503, + 0.0170640517026186, + 0.002693026792258024, + 0.0020493189804255962, + 0.002929909620434046, + 0.0017190343933179975, + 0.004059800412505865, + 0.012738460674881935, + 0.002138966228812933, + 0.004421341232955456, + 0.007258876226842403, + 0.010255117900669575, + 0.006774405017495155, + 0.006993972696363926, + 0.033521927893161774, + 0.01994277723133564, + 0.008147758431732655, + 0.016712846234440804, + 0.0015696723712608218, + 0.0016358409775421023, + 0.0022202134132385254, + 0.002099229022860527, + 0.0028999210335314274, + 0.0037366074975579977, + 0.002043995074927807, + 0.0035641444846987724, + 0.001791127142496407, + 0.0015293501783162355, + 0.00127002305816859, + 0.0021932455711066723, + 0.0034514451399445534, + 0.007303352002054453, + 0.0017632580129429698, + 0.007917243055999279, + 0.026603519916534424, + 0.015282036736607552, + 0.01039907243102789, + 0.009500439278781414, + 0.023662803694605827, + 0.011759870685636997, + 0.01339706126600504, + 0.032152656465768814, + 0.04263148829340935, + 0.01561048999428749, + 0.05856422334909439, + 0.02225235477089882, + 0.018522968515753746, + 0.01749967597424984, + 0.021098867058753967, + 0.017771629616618156, + 0.024375280365347862, + 0.02099846489727497, + 0.01988872140645981, + 0.008822388015687466, + 0.007876239717006683, + 0.008548236452043056, + 0.007358180824667215, + 0.005388210061937571, + 0.009680368937551975, + 0.008578462526202202, + 0.008353933691978455, + 0.00716477120295167, + 0.01256076991558075, + 0.011872927658259869, + 0.010321522131562233, + 0.013663075864315033, + 0.022819513455033302, + 0.036830659955739975, + 0.034735746681690216, + 0.016335653141140938, + 0.014083434827625751, + 0.017583899199962616, + 0.015462430194020271, + 0.00963615532964468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "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 main subject the Monkey King wearing golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King wearing a purple-gold crown with two long pheasant feathers on his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King performing a lunge while holding a jet-black Ruyi Jingu Club above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the club smashing a celestial mountain, causing it to crack and shatter into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust and debris from the impact depicted as thick ink strokes flowing through the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248102825134993 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005248102825134993, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248102825134993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14594653248786926, + 0.06466065347194672, + 0.08867631107568741, + 0.01803721860051155, + 0.030307410284876823, + 0.07876908779144287, + 0.2819305956363678, + 0.09251327067613602, + 0.039688847959041595, + 0.15947003662586212 + ], + "credit_map": [ + 0.009770995937287807, + 0.009649822488427162, + 0.007954200729727745, + 0.00924323033541441, + 0.03085044026374817, + 0.04048221930861473, + 0.010102375410497189, + 0.027893247082829475, + 0.004668508656322956, + 0.0033534050453454256, + 0.003829263849183917, + 0.004422967787832022, + 0.00999569334089756, + 0.02466653473675251, + 0.0045231119729578495, + 0.009201166220009327, + 0.005118729081004858, + 0.0053351204842329025, + 0.0033070938661694527, + 0.005036324728280306, + 0.014936409890651703, + 0.02412087470293045, + 0.0043400549329817295, + 0.026481693610548973, + 0.002066004788503051, + 0.0010901434579864144, + 0.0013131151208654046, + 0.0015962865436449647, + 0.001891875988803804, + 0.005894304718822241, + 0.001570416148751974, + 0.0026150718331336975, + 0.0022977190092206, + 0.00228984490968287, + 0.001724647474475205, + 0.0016564343823119998, + 0.006926598958671093, + 0.006729234009981155, + 0.001984288450330496, + 0.006698643788695335, + 0.01503794826567173, + 0.009175962768495083, + 0.006301330402493477, + 0.00574121018871665, + 0.008811743929982185, + 0.007383838761597872, + 0.008019724860787392, + 0.018297327682375908, + 0.07046505808830261, + 0.016903232783079147, + 0.08589328080415726, + 0.02248048596084118, + 0.02252586930990219, + 0.02037006802856922, + 0.020109064877033234, + 0.023183532059192657, + 0.01366318017244339, + 0.011740276589989662, + 0.012292073108255863, + 0.01171031966805458, + 0.00999730546027422, + 0.01355666946619749, + 0.012560898438096046, + 0.006992544513195753, + 0.0058744619600474834, + 0.006451170425862074, + 0.0046583302319049835, + 0.005318469367921352, + 0.005114641971886158, + 0.004133838228881359, + 0.004543012008070946, + 0.003594923997297883, + 0.02305525727570057, + 0.03623637557029724, + 0.03040912188589573, + 0.013963648118078709, + 0.015032959170639515, + 0.017627105116844177, + 0.012960709631443024, + 0.010184849612414837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.005248102825134993 + } + ], + "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/0000028/group_3/index_tensor([2611, 2611])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_3/index_tensor([2611, 2611])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..00b7c6c0e79b6e6ff06b9c8c112a0d60512577b7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_3/index_tensor([2611, 2611])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From an intense first-person twin shot, simulating the perspective of a belligerent young rooster, the camera remains fixed as a brightly feathered opponent charges directly toward the camera. The entire explosive moment is captured using high-speed macro photography and rendered in stunning slow motion. The frame is filled with the opponent rooster, a majestic creature of its imposing size. Its head thrust forward, its neck feathers—a dazzling array of iridescent green and lustrous black—radiating aggressive intent. Its small eyes, tips of black obsidian, reflect the bright sun overhead, locked in a gaze of focused fury. As it thrusts forward, its powerful, scaly yellow legs extend, its sharp, curved claws gleaming in the light. It pushes off explosively from the dusty ground, creating a miniature cloud of dirt and straw particles in the surrounding air. The high-speed capture reveals every tiny detail: the way the air agitates and separates the individual barbs of its breast feathers, the subtle tension of the muscles beneath its skin, and the raw kinetic energy of the impending clash, creating a visceral and beautiful atmosphere of primal conflict.", + "seed": 30867, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_3/index_2611_seed_30867_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": -0.20864376053214073, + "vlm_logprob": -0.20864376053214073, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.20864376053214073, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.36519062519073486 + ], + "opposite_logprob_tokens": [ + -1.2401906251907349 + ], + "logprob_reward": -0.36519062519073486, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36519062519073486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09523871541023254, + 0.17815367877483368, + 0.1104804053902626, + 0.10781730711460114, + 0.07079111039638519, + 0.19625982642173767, + 0.10527299344539642, + 0.0730273649096489, + 0.04650786146521568, + 0.016450773924589157 + ], + "credit_map": [ + 0.006181030534207821, + 0.0097344433888793, + 0.010229203850030899, + 0.012138290330767632, + 0.011123040691018105, + 0.021015791222453117, + 0.012812022119760513, + 0.01200488768517971, + 0.017412446439266205, + 0.011453069746494293, + 0.02220771461725235, + 0.039752859622240067, + 0.0176071897149086, + 0.025795163586735725, + 0.017823444679379463, + 0.026101768016815186, + 0.02222096361219883, + 0.010334481485188007, + 0.020214693620800972, + 0.014599918387830257, + 0.009905444458127022, + 0.008848602883517742, + 0.014301063492894173, + 0.01005522720515728, + 0.009267589077353477, + 0.01941884495317936, + 0.010614655911922455, + 0.018697386607527733, + 0.008018705993890762, + 0.012115268968045712, + 0.012853719294071198, + 0.016831127926707268, + 0.006529235281050205, + 0.011515969410538673, + 0.00583972642198205, + 0.011658909730613232, + 0.008769329637289047, + 0.008237664587795734, + 0.010831880383193493, + 0.007408387493342161, + 0.01405722089111805, + 0.04066126421093941, + 0.011262591928243637, + 0.031929392367601395, + 0.02335035800933838, + 0.029165076091885567, + 0.03254105895757675, + 0.01329285278916359, + 0.011270890943706036, + 0.02381965145468712, + 0.006011173129081726, + 0.010388091206550598, + 0.029451478272676468, + 0.008660048246383667, + 0.010574470274150372, + 0.005097188986837864, + 0.007579837925732136, + 0.011184914037585258, + 0.0072915940545499325, + 0.00817285105586052, + 0.012456383556127548, + 0.006660172250121832, + 0.010286711156368256, + 0.009394891560077667, + 0.006886389572173357, + 0.006953167729079723, + 0.0039937845431268215, + 0.008134049363434315, + 0.004226597957313061, + 0.004851535893976688, + 0.005652249790728092, + 0.00581008568406105, + 0.002037910046055913, + 0.002214080886915326, + 0.0020996727980673313, + 0.002792480867356062, + 0.0019203106639906764, + 0.001774078467860818, + 0.0025582825765013695, + 0.0010539570357650518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.36519062519073486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.016102269291877747 + ], + "opposite_logprob_tokens": [ + -4.141102313995361 + ], + "logprob_reward": -0.016102269291877747, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016102269291877747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08216408640146255, + 0.2295285314321518, + 0.03556181117892265, + 0.09663224220275879, + 0.05489566549658775, + 0.28752562403678894, + 0.0992487370967865, + 0.05128899961709976, + 0.04533693194389343, + 0.017817329615354538 + ], + "credit_map": [ + 0.00656483881175518, + 0.009797807782888412, + 0.007554273586720228, + 0.009519042447209358, + 0.0078990263864398, + 0.02063535340130329, + 0.009165758267045021, + 0.011027987115085125, + 0.02397806942462921, + 0.013269138522446156, + 0.02463771216571331, + 0.04311405122280121, + 0.02519734390079975, + 0.03848909214138985, + 0.027111303061246872, + 0.033731818199157715, + 0.0034761421848088503, + 0.0025275894440710545, + 0.0028420391026884317, + 0.0053559537045657635, + 0.004573027137666941, + 0.005022245924919844, + 0.007085052318871021, + 0.004679759033024311, + 0.006772734224796295, + 0.015281367115676403, + 0.008258172310888767, + 0.021447516977787018, + 0.00698051555082202, + 0.00956121925264597, + 0.01484273187816143, + 0.01348799280822277, + 0.006384202279150486, + 0.012948974967002869, + 0.004900702275335789, + 0.009051110595464706, + 0.005110317841172218, + 0.00475146621465683, + 0.006227580364793539, + 0.005521311890333891, + 0.020808709785342216, + 0.0586373507976532, + 0.022801296785473824, + 0.0455932654440403, + 0.029928632080554962, + 0.03355598449707031, + 0.048982053995132446, + 0.027218323200941086, + 0.008427996188402176, + 0.020743651315569878, + 0.006192853674292564, + 0.009858335368335247, + 0.028332583606243134, + 0.01039040181785822, + 0.010018843226134777, + 0.005284070502966642, + 0.004298261366784573, + 0.0067461454309523106, + 0.004279333166778088, + 0.008589396253228188, + 0.009653322398662567, + 0.00581740727648139, + 0.006910564377903938, + 0.004994566086679697, + 0.0045972238294780254, + 0.0066305301152169704, + 0.004321326036006212, + 0.011119548231363297, + 0.003641722956672311, + 0.0039526023901999, + 0.005758242681622505, + 0.005315733607858419, + 0.0024902585428208113, + 0.0018916261615231633, + 0.0026534264907240868, + 0.00377761316485703, + 0.002766356337815523, + 0.0016660933615639806, + 0.001498539699241519, + 0.0010734157403931022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.016102269291877747 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.03819701448082924 + ], + "opposite_logprob_tokens": [ + -3.2881970405578613 + ], + "logprob_reward": -0.03819701448082924, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03819701448082924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06910732388496399, + 0.03820735588669777, + 0.06101069226861, + 0.10227696597576141, + 0.030254065990447998, + 0.22460518777370453, + 0.09272750467061996, + 0.2605811357498169, + 0.09346841275691986, + 0.027761416509747505 + ], + "credit_map": [ + 0.004509446211159229, + 0.007467646151781082, + 0.006952796597033739, + 0.006312218494713306, + 0.007650385145097971, + 0.020367112010717392, + 0.007183105684816837, + 0.008664604276418686, + 0.00433094659820199, + 0.0029306020587682724, + 0.004249867983162403, + 0.00656192097812891, + 0.0031874789856374264, + 0.00591818243265152, + 0.004747966770082712, + 0.0062803840264678, + 0.013133742846548557, + 0.005971488542854786, + 0.010838803835213184, + 0.008167355321347713, + 0.005443737376481295, + 0.006609395146369934, + 0.005178769119083881, + 0.00566739309579134, + 0.005459037143737078, + 0.010177996940910816, + 0.006725828628987074, + 0.021712100133299828, + 0.006786395329982042, + 0.014362781308591366, + 0.018261989578604698, + 0.01879083178937435, + 0.0026101181283593178, + 0.003541322657838464, + 0.006386465393006802, + 0.0037403996102511883, + 0.0030382066033780575, + 0.003729533636942506, + 0.0030773659236729145, + 0.0041306521743535995, + 0.017744876444339752, + 0.047331057488918304, + 0.020073236897587776, + 0.03330663964152336, + 0.024093613028526306, + 0.017428472638130188, + 0.046620164066553116, + 0.018007108941674232, + 0.008730036206543446, + 0.019382858648896217, + 0.006202685181051493, + 0.009952970780432224, + 0.024240445345640182, + 0.00819660909473896, + 0.011416581459343433, + 0.004605305381119251, + 0.022493286058306694, + 0.031746745109558105, + 0.024706987664103508, + 0.05276269465684891, + 0.02907097153365612, + 0.030683483928442, + 0.038411371409893036, + 0.030705559998750687, + 0.009640639647841454, + 0.014040167443454266, + 0.009651578031480312, + 0.020897720009088516, + 0.007865659892559052, + 0.009167943149805069, + 0.012114794924855232, + 0.01008989755064249, + 0.003473020391538739, + 0.0030259788036346436, + 0.0038523394614458084, + 0.006266937591135502, + 0.004257709253579378, + 0.0022249827161431313, + 0.0028516468591988087, + 0.001808797474950552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.03819701448082924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.10034116357564926 + ], + "opposite_logprob_tokens": [ + -2.3503410816192627 + ], + "logprob_reward": -0.10034116357564926, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034116357564926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04603172466158867, + 0.03553333133459091, + 0.035119399428367615, + 0.06104134023189545, + 0.055325210094451904, + 0.04468198120594025, + 0.16732674837112427, + 0.0863557681441307, + 0.4417250156402588, + 0.026859456673264503 + ], + "credit_map": [ + 0.002728011691942811, + 0.00619540037587285, + 0.004050533752888441, + 0.0076639167964458466, + 0.004473112989217043, + 0.009390982799232006, + 0.005495812278240919, + 0.00603395514190197, + 0.004402327351272106, + 0.0028219101950526237, + 0.004124338738620281, + 0.006585102993994951, + 0.0042970371432602406, + 0.005173369310796261, + 0.004693934228271246, + 0.0034353083465248346, + 0.0030441598501056433, + 0.002977584023028612, + 0.004536169581115246, + 0.006451350636780262, + 0.003019590163603425, + 0.003751112148165703, + 0.007491610012948513, + 0.0038478223141282797, + 0.0036774403415620327, + 0.006360527593642473, + 0.004891521297395229, + 0.010657897219061852, + 0.005619573872536421, + 0.006406272761523724, + 0.010775433853268623, + 0.012652670033276081, + 0.007718052715063095, + 0.007454302161931992, + 0.007261415012180805, + 0.007490396965295076, + 0.00625245226547122, + 0.00541613157838583, + 0.007255158852785826, + 0.006477301008999348, + 0.0042562298476696014, + 0.008935193531215191, + 0.0037477666046470404, + 0.007215338759124279, + 0.00516311451792717, + 0.005179548170417547, + 0.006874780170619488, + 0.0033100112341344357, + 0.019678035750985146, + 0.025929853320121765, + 0.02041308395564556, + 0.017488930374383926, + 0.031060591340065002, + 0.016808701679110527, + 0.021860243752598763, + 0.014087317511439323, + 0.007255017291754484, + 0.01243098359555006, + 0.008057622238993645, + 0.013194363564252853, + 0.012869704514741898, + 0.01088779792189598, + 0.013258961960673332, + 0.008401310071349144, + 0.04498370364308357, + 0.06392613053321838, + 0.04476135969161987, + 0.09359771013259888, + 0.041088759899139404, + 0.05298135429620743, + 0.052372902631759644, + 0.04801309481263161, + 0.003967454191297293, + 0.0034496851731091738, + 0.0037354151718318462, + 0.0049288454465568066, + 0.004143002443015575, + 0.0018903977470472455, + 0.0030925532337278128, + 0.001652101636864245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.10034116357564926 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.5233877301216125 + ], + "opposite_logprob_tokens": [ + -0.8983877301216125 + ], + "logprob_reward": -0.5233877301216125, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233877301216125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08142103254795074, + 0.08168070018291473, + 0.04061303287744522, + 0.09565109014511108, + 0.05786389112472534, + 0.4573386013507843, + 0.058910779654979706, + 0.04024428874254227, + 0.05724107474088669, + 0.029035525396466255 + ], + "credit_map": [ + 0.0057802703231573105, + 0.009128307923674583, + 0.008199824020266533, + 0.00866762176156044, + 0.009318886324763298, + 0.021270833909511566, + 0.009370465762913227, + 0.00968481507152319, + 0.010035195387899876, + 0.005771417636424303, + 0.011385913006961346, + 0.0159421656280756, + 0.009602414444088936, + 0.01029554009437561, + 0.0077401697635650635, + 0.010907880961894989, + 0.004621363244950771, + 0.003979030065238476, + 0.003172473283484578, + 0.006879144348204136, + 0.003468351438641548, + 0.004707108251750469, + 0.00930527038872242, + 0.004480292554944754, + 0.008608246222138405, + 0.014955664984881878, + 0.009088348597288132, + 0.018302293494343758, + 0.0066769798286259174, + 0.009232006967067719, + 0.012836309149861336, + 0.0159512460231781, + 0.007141850423067808, + 0.00680927000939846, + 0.01057557761669159, + 0.007771159056574106, + 0.005786081776022911, + 0.0059118010103702545, + 0.0067292433232069016, + 0.007138907443732023, + 0.03879055380821228, + 0.08268488943576813, + 0.04720476642251015, + 0.06816400587558746, + 0.044614505022764206, + 0.05004046857357025, + 0.08905276656150818, + 0.03678665682673454, + 0.004990886431187391, + 0.01278874184936285, + 0.005131132900714874, + 0.00499497028067708, + 0.01711893267929554, + 0.006098276935517788, + 0.003964104689657688, + 0.0038237327244132757, + 0.0030192004051059484, + 0.004243075381964445, + 0.004090758506208658, + 0.006608110386878252, + 0.0046984045766294, + 0.004562580958008766, + 0.005913278087973595, + 0.007108880206942558, + 0.005923733115196228, + 0.008831752464175224, + 0.005791844800114632, + 0.012762026861310005, + 0.004927849397063255, + 0.005473485216498375, + 0.007399467751383781, + 0.006130916997790337, + 0.0033016488887369633, + 0.003185677109286189, + 0.00439219968393445, + 0.006650634575635195, + 0.003924029879271984, + 0.002999428426846862, + 0.00276461080648005, + 0.0018172976560890675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.5233877301216125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From an intense first-person twin shot, simulating the perspective of a belligerent young rooster, the camera remains fixed as a brightly feathered opponent charges directly toward the camera. The entire explosive moment is captured using high-speed macro photography and rendered in stunning slow motion. The frame is filled with the opponent rooster, a majestic creature of its imposing size. Its head thrust forward, its neck feathers—a dazzling array of iridescent green and lustrous black—radiating aggressive intent. Its small eyes, tips of black obsidian, reflect the bright sun overhead, locked in a gaze of focused fury. As it thrusts forward, its powerful, scaly yellow legs extend, its sharp, curved claws gleaming in the light. It pushes off explosively from the dusty ground, creating a miniature cloud of dirt and straw particles in the surrounding air. The high-speed capture reveals every tiny detail: the way the air agitates and separates the individual barbs of its breast feathers, the subtle tension of the muscles beneath its skin, and the raw kinetic energy of the impending clash, creating a visceral and beautiful atmosphere of primal conflict.", + "seed": 30868, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_3/index_2611_seed_30868_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": -0.1859095560386777, + "vlm_logprob": -0.1859095560386777, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.1859095560386777, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.07090416550636292 + ], + "opposite_logprob_tokens": [ + -2.94590425491333 + ], + "logprob_reward": -0.07090416550636292, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07090416550636292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19223655760288239, + 0.10436451435089111, + 0.05286884680390358, + 0.03219957649707794, + 0.03948735073208809, + 0.032396841794252396, + 0.07377801090478897, + 0.07396959513425827, + 0.34955894947052, + 0.04913977533578873 + ], + "credit_map": [ + 0.01818138360977173, + 0.017878733575344086, + 0.012418177910149097, + 0.023928219452500343, + 0.03638802468776703, + 0.027562594041228294, + 0.038584765046834946, + 0.017294662073254585, + 0.012905431911349297, + 0.010210344567894936, + 0.011153003200888634, + 0.015261972323060036, + 0.013860730454325676, + 0.014067576266825199, + 0.011195997707545757, + 0.01570945419371128, + 0.0031005744822323322, + 0.004354987293481827, + 0.00631625996902585, + 0.009247428737580776, + 0.008351885713636875, + 0.005084235221147537, + 0.008699308149516582, + 0.007714166771620512, + 0.004745766520500183, + 0.0032592483330518007, + 0.006764980498701334, + 0.005950021091848612, + 0.002410182263702154, + 0.0034229191951453686, + 0.0029792923014611006, + 0.002667168155312538, + 0.0033235931769013405, + 0.006339119281619787, + 0.003854096168652177, + 0.003912644926458597, + 0.006618651561439037, + 0.004343010485172272, + 0.00445355661213398, + 0.006642679218202829, + 0.0040620374493300915, + 0.0032976798247545958, + 0.003208461683243513, + 0.003304851707071066, + 0.004576532170176506, + 0.0039145685732364655, + 0.00574543746188283, + 0.00428727176040411, + 0.0038212582003325224, + 0.011128819547593594, + 0.006822729948908091, + 0.011021086946129799, + 0.014785151928663254, + 0.01331262569874525, + 0.006465899758040905, + 0.006420433986932039, + 0.008919017389416695, + 0.011381841264665127, + 0.006428512278944254, + 0.009937634691596031, + 0.013561241328716278, + 0.009471074678003788, + 0.00743057718500495, + 0.006839691661298275, + 0.027839886024594307, + 0.13086487352848053, + 0.01678040809929371, + 0.021747969090938568, + 0.1129075214266777, + 0.014869392849504948, + 0.01112994272261858, + 0.013418965972959995, + 0.004871372599154711, + 0.0045334105379879475, + 0.00568847730755806, + 0.006964338943362236, + 0.006154303904622793, + 0.005599309224635363, + 0.00810975395143032, + 0.007218810264021158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.07090416550636292 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.007646103389561176 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.007646103389561176, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646103389561176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22468505799770355, + 0.130097895860672, + 0.042963381856679916, + 0.04610433056950569, + 0.05041845515370369, + 0.04864934831857681, + 0.03635290637612343, + 0.06994529068470001, + 0.28475987911224365, + 0.06602349132299423 + ], + "credit_map": [ + 0.018467599526047707, + 0.02118801698088646, + 0.014142570085823536, + 0.028020383790135384, + 0.045039426535367966, + 0.032796461135149, + 0.04668603464961052, + 0.018344560638070107, + 0.01341033075004816, + 0.00996325071901083, + 0.013553755357861519, + 0.01468653604388237, + 0.017627399414777756, + 0.017672166228294373, + 0.014975195750594139, + 0.028209269046783447, + 0.005529372952878475, + 0.0027854794170707464, + 0.009132315404713154, + 0.0046889446675777435, + 0.0044022249057888985, + 0.005142302252352238, + 0.0028785562608391047, + 0.008404186926782131, + 0.006321460474282503, + 0.006050402764230967, + 0.007213611155748367, + 0.007452267687767744, + 0.004842448979616165, + 0.005210735835134983, + 0.004573006648570299, + 0.004440397024154663, + 0.005174999590963125, + 0.005495025310665369, + 0.004994092509150505, + 0.007041831500828266, + 0.006999323610216379, + 0.005636817775666714, + 0.005996293388307095, + 0.009080071933567524, + 0.006703588180243969, + 0.005620830226689577, + 0.003910565748810768, + 0.006294798571616411, + 0.00743947783485055, + 0.007935981266200542, + 0.00504269078373909, + 0.005701412446796894, + 0.0028418272268027067, + 0.006471958477050066, + 0.003136218059808016, + 0.004874879494309425, + 0.006602200213819742, + 0.006620556116104126, + 0.003328562481328845, + 0.002476703142747283, + 0.008848781697452068, + 0.008336835540831089, + 0.006723048165440559, + 0.008483647368848324, + 0.00893485639244318, + 0.008353759534657001, + 0.012115845456719398, + 0.008148517459630966, + 0.0206026341766119, + 0.10947480797767639, + 0.012650317512452602, + 0.014619112946093082, + 0.09453842043876648, + 0.013047806918621063, + 0.009583299979567528, + 0.01024347823113203, + 0.007135068532079458, + 0.005535295233130455, + 0.009321130812168121, + 0.007288674358278513, + 0.007738129235804081, + 0.005922211799770594, + 0.014453575946390629, + 0.008629403077065945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.007646103389561176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.03818932920694351 + ], + "opposite_logprob_tokens": [ + -3.28818941116333 + ], + "logprob_reward": -0.03818932920694351, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03818932920694351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19015823304653168, + 0.0656542256474495, + 0.061314672231674194, + 0.0514054074883461, + 0.035915374755859375, + 0.03693044185638428, + 0.02978062629699707, + 0.20343604683876038, + 0.20661623775959015, + 0.11878866702318192 + ], + "credit_map": [ + 0.015576338395476341, + 0.018473127856850624, + 0.011602334678173065, + 0.0231565423309803, + 0.03854920715093613, + 0.026605656370520592, + 0.04020221158862114, + 0.015992823988199234, + 0.006189184263348579, + 0.00501300860196352, + 0.005854298826307058, + 0.005108162295073271, + 0.011500871740281582, + 0.009207131341099739, + 0.00900273397564888, + 0.013778827153146267, + 0.005865784827619791, + 0.005574276205152273, + 0.009221337735652924, + 0.009381918236613274, + 0.004415210802108049, + 0.005890838336199522, + 0.007015487644821405, + 0.013949818909168243, + 0.005143366288393736, + 0.0055995844304561615, + 0.008599926717579365, + 0.010718884877860546, + 0.005410374607890844, + 0.00543099083006382, + 0.0051743038929998875, + 0.005327975377440453, + 0.00489420723170042, + 0.004169465973973274, + 0.0032586336601525545, + 0.0038355712313205004, + 0.005750284995883703, + 0.00359736243262887, + 0.003930252976715565, + 0.006479594856500626, + 0.003877113340422511, + 0.0038377507589757442, + 0.0028700826223939657, + 0.00575811555609107, + 0.00849407259374857, + 0.004735083784908056, + 0.003743052249774337, + 0.003615171415731311, + 0.001985339680686593, + 0.0040823607705533504, + 0.0032257784623652697, + 0.004116919357329607, + 0.005326408427208662, + 0.004586941562592983, + 0.003286191029474139, + 0.003170687472447753, + 0.02112826332449913, + 0.032907385379076004, + 0.019609525799751282, + 0.018381362780928612, + 0.02679479867219925, + 0.02797231636941433, + 0.0335562601685524, + 0.023086147382855415, + 0.02413216419517994, + 0.04260842129588127, + 0.017296677455306053, + 0.016453558579087257, + 0.03375424072146416, + 0.013512921519577503, + 0.01218817476183176, + 0.0466700941324234, + 0.009906396269798279, + 0.00931235495954752, + 0.015875698998570442, + 0.017899148166179657, + 0.01437678188085556, + 0.010234443470835686, + 0.02342594414949417, + 0.01775789074599743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.03818932920694351 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.054965581744909286 + ], + "opposite_logprob_tokens": [ + -2.9299654960632324 + ], + "logprob_reward": -0.054965581744909286, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965581744909286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19357803463935852, + 0.24129840731620789, + 0.053108807653188705, + 0.06053028255701065, + 0.021999308839440346, + 0.020177703350782394, + 0.0333053320646286, + 0.09070205688476562, + 0.17271262407302856, + 0.11258743703365326 + ], + "credit_map": [ + 0.015038279816508293, + 0.018272893503308296, + 0.013452610932290554, + 0.028107719495892525, + 0.037514086812734604, + 0.027242837473750114, + 0.03860123082995415, + 0.015348391607403755, + 0.026540910825133324, + 0.019470535218715668, + 0.02296123281121254, + 0.023338256403803825, + 0.03586789220571518, + 0.03347203880548477, + 0.028961623087525368, + 0.05068593472242355, + 0.005837681703269482, + 0.006417377386242151, + 0.007011807058006525, + 0.008955721743404865, + 0.004729971755295992, + 0.006218759808689356, + 0.005466916598379612, + 0.008470567874610424, + 0.004846785683184862, + 0.003184507368132472, + 0.018026597797870636, + 0.017754947766661644, + 0.005105484277009964, + 0.00402477802708745, + 0.0035794684663414955, + 0.004007715731859207, + 0.002543521812185645, + 0.0020738323219120502, + 0.0022785100154578686, + 0.0027371456380933523, + 0.003201276296749711, + 0.0022342202719300985, + 0.0028927968814969063, + 0.004038005135953426, + 0.0026304428465664387, + 0.0022489966358989477, + 0.001596361631527543, + 0.00253555946983397, + 0.0026028542779386044, + 0.003085694508627057, + 0.0029557375237345695, + 0.0025220548268407583, + 0.00210394524037838, + 0.005583550781011581, + 0.0027045556344091892, + 0.0042306529358029366, + 0.007076415233314037, + 0.005288836546242237, + 0.003663283307105303, + 0.002654094947502017, + 0.012472901493310928, + 0.013237730599939823, + 0.011186245828866959, + 0.008232258260250092, + 0.01155796553939581, + 0.012550890445709229, + 0.009774532169103622, + 0.011689535342156887, + 0.013946466147899628, + 0.06618942320346832, + 0.007838210090994835, + 0.008275028318166733, + 0.057306401431560516, + 0.007698309142142534, + 0.005585646256804466, + 0.005873152054846287, + 0.008338131941854954, + 0.009201199747622013, + 0.01890391856431961, + 0.01327436976134777, + 0.012445535510778427, + 0.010952704586088657, + 0.024983983486890793, + 0.014487589709460735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.054965581744909286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578426003456116 + ], + "opposite_logprob_tokens": [ + -0.6328426003456116 + ], + "logprob_reward": -0.7578426003456116, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578426003456116, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.31430622935295105, + 0.05900038033723831, + 0.10129586607217789, + 0.058600980788469315, + 0.07552042603492737, + 0.032614849507808685, + 0.05817513167858124, + 0.05385777726769447, + 0.20685656368732452, + 0.039771806448698044 + ], + "credit_map": [ + 0.024096259847283363, + 0.03056366555392742, + 0.02026410587131977, + 0.039484161883592606, + 0.06326872110366821, + 0.04510927200317383, + 0.06485209614038467, + 0.026667941361665726, + 0.00528726214542985, + 0.004726087674498558, + 0.0066422284580767155, + 0.009194262325763702, + 0.00834824051707983, + 0.0073167565278708935, + 0.008577368222177029, + 0.008908172138035297, + 0.010112250223755836, + 0.010192496702075005, + 0.016137566417455673, + 0.015897365286946297, + 0.009395266883075237, + 0.011457822285592556, + 0.009393187239766121, + 0.01870991662144661, + 0.011050236411392689, + 0.007511242758482695, + 0.007584162522107363, + 0.008641477674245834, + 0.004851077683269978, + 0.006199787370860577, + 0.004940951243042946, + 0.007822046987712383, + 0.006088456138968468, + 0.008901916444301605, + 0.00815084483474493, + 0.01253292802721262, + 0.011212820187211037, + 0.006945970933884382, + 0.009274302050471306, + 0.01241319254040718, + 0.003981287591159344, + 0.0033951587975025177, + 0.002830301644280553, + 0.004086930770426989, + 0.004305319860577583, + 0.004379424266517162, + 0.005146494600921869, + 0.004489933140575886, + 0.0032915519550442696, + 0.008306511677801609, + 0.004983014892786741, + 0.008973905816674232, + 0.012968412600457668, + 0.0075079756788909435, + 0.0060909162275493145, + 0.006052838638424873, + 0.006585148628801107, + 0.007079101633280516, + 0.006158582866191864, + 0.0071621788665652275, + 0.00672649871557951, + 0.008322960697114468, + 0.006755893584340811, + 0.005067409016191959, + 0.014281176030635834, + 0.07951287180185318, + 0.008870436809957027, + 0.008302035741508007, + 0.06860402971506119, + 0.007734977174550295, + 0.008135530166327953, + 0.011415515094995499, + 0.004576832987368107, + 0.004789488390088081, + 0.0034304496366530657, + 0.005184379406273365, + 0.004713078029453754, + 0.005719142034649849, + 0.005137561354786158, + 0.006220877170562744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.7578426003456116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "From an intense first-person twin shot, simulating the perspective of a belligerent young rooster, the camera remains fixed as a brightly feathered opponent charges directly toward the camera. The entire explosive moment is captured using high-speed macro photography and rendered in stunning slow motion. The frame is filled with the opponent rooster, a majestic creature of its imposing size. Its head thrust forward, its neck feathers—a dazzling array of iridescent green and lustrous black—radiating aggressive intent. Its small eyes, tips of black obsidian, reflect the bright sun overhead, locked in a gaze of focused fury. As it thrusts forward, its powerful, scaly yellow legs extend, its sharp, curved claws gleaming in the light. It pushes off explosively from the dusty ground, creating a miniature cloud of dirt and straw particles in the surrounding air. The high-speed capture reveals every tiny detail: the way the air agitates and separates the individual barbs of its breast feathers, the subtle tension of the muscles beneath its skin, and the raw kinetic energy of the impending clash, creating a visceral and beautiful atmosphere of primal conflict.", + "seed": 30869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_3/index_2611_seed_30869_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": -0.17269758004695177, + "vlm_logprob": -0.17269758004695177, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.17269758004695177, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.08389110863208771 + ], + "opposite_logprob_tokens": [ + -2.7088911533355713 + ], + "logprob_reward": -0.08389110863208771, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08389110863208771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12132528424263, + 0.05590477213263512, + 0.08925820142030716, + 0.10702873766422272, + 0.03559468314051628, + 0.12013023346662521, + 0.23590858280658722, + 0.07483026385307312, + 0.10576062649488449, + 0.054258592426776886 + ], + "credit_map": [ + 0.03198781982064247, + 0.011617113836109638, + 0.012852491810917854, + 0.013286005705595016, + 0.009929061867296696, + 0.010158359073102474, + 0.017437640577554703, + 0.014056791551411152, + 0.009670698083937168, + 0.007631001062691212, + 0.010846161283552647, + 0.004514525178819895, + 0.007796641439199448, + 0.005170810502022505, + 0.0034965896047651768, + 0.00677834264934063, + 0.011809642426669598, + 0.006968249566853046, + 0.0115194758400321, + 0.008527331054210663, + 0.010463444516062737, + 0.015191888436675072, + 0.00864141620695591, + 0.016136756166815758, + 0.012975548394024372, + 0.012948816642165184, + 0.010327489115297794, + 0.010436556302011013, + 0.01522340252995491, + 0.012423903681337833, + 0.022104933857917786, + 0.010588091798126698, + 0.004793514497578144, + 0.004010094795376062, + 0.0036853847559541464, + 0.003282250137999654, + 0.0040907650254666805, + 0.007420409005135298, + 0.00389020680449903, + 0.004422055557370186, + 0.005740908440202475, + 0.004034238401800394, + 0.03486540541052818, + 0.004050404764711857, + 0.0032230461947619915, + 0.009606716223061085, + 0.006445284932851791, + 0.052164226770401, + 0.02340591885149479, + 0.04225391149520874, + 0.03348749130964279, + 0.028734425082802773, + 0.029098106548190117, + 0.029617130756378174, + 0.026697319000959396, + 0.02261427603662014, + 0.00723919365555048, + 0.00967938918620348, + 0.012549417093396187, + 0.009080814197659492, + 0.007852326147258282, + 0.005717392545193434, + 0.006342687178403139, + 0.01636904664337635, + 0.010632168501615524, + 0.02180875651538372, + 0.007964123040437698, + 0.007629745174199343, + 0.027049966156482697, + 0.008861172944307327, + 0.014243998564779758, + 0.007570697460323572, + 0.006803198251873255, + 0.007439238950610161, + 0.00781026016920805, + 0.00729190930724144, + 0.006163523532450199, + 0.007977975532412529, + 0.005902441684156656, + 0.004870045464485884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.08389110863208771 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.007647995837032795 + ], + "opposite_logprob_tokens": [ + -4.88264799118042 + ], + "logprob_reward": -0.007647995837032795, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647995837032795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1718009114265442, + 0.05674365907907486, + 0.08171696960926056, + 0.08879460394382477, + 0.10824286192655563, + 0.1631426215171814, + 0.046112313866615295, + 0.07821939885616302, + 0.16007143259048462, + 0.045155156403779984 + ], + "credit_map": [ + 0.049467675387859344, + 0.025993140414357185, + 0.018115416169166565, + 0.01393634919077158, + 0.011654925532639027, + 0.014523050747811794, + 0.01848480850458145, + 0.019625570625066757, + 0.009985161013901234, + 0.006525876000523567, + 0.009997990913689137, + 0.0055488478392362595, + 0.006602316163480282, + 0.0062458813190460205, + 0.004241922404617071, + 0.007595664821565151, + 0.011047668755054474, + 0.008636856451630592, + 0.010151263326406479, + 0.007618629839271307, + 0.009180571883916855, + 0.018609346821904182, + 0.007596182636916637, + 0.008876453153789043, + 0.00974817480891943, + 0.011422354727983475, + 0.0086973300203681, + 0.011063060723245144, + 0.010194034315645695, + 0.01443126704543829, + 0.014791404828429222, + 0.008446993306279182, + 0.012366794981062412, + 0.0170879065990448, + 0.011695965193212032, + 0.008850163780152798, + 0.012852496467530727, + 0.018750391900539398, + 0.011368912644684315, + 0.015270229429006577, + 0.008079798892140388, + 0.005619324278086424, + 0.0461067333817482, + 0.005271709989756346, + 0.005182271357625723, + 0.01310229767113924, + 0.008972256444394588, + 0.07080824673175812, + 0.005876478739082813, + 0.007307912223041058, + 0.004726008512079716, + 0.0054250736720860004, + 0.006666210945695639, + 0.006815908011049032, + 0.005395331420004368, + 0.003899393370375037, + 0.0067435880191624165, + 0.012221082113683224, + 0.015105316415429115, + 0.0072648050263524055, + 0.008988868445158005, + 0.005587717518210411, + 0.0051434896886348724, + 0.01716453954577446, + 0.010120978578925133, + 0.05006157234311104, + 0.011680272407829762, + 0.008375520817935467, + 0.04395419731736183, + 0.006693093106150627, + 0.007929747924208641, + 0.021256059408187866, + 0.004654936492443085, + 0.006783274933695793, + 0.005620361305773258, + 0.007963694632053375, + 0.0051710582338273525, + 0.005535667296499014, + 0.00546669214963913, + 0.003959474619477987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.007647995837032795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.048742204904556274 + ], + "opposite_logprob_tokens": [ + -3.0487422943115234 + ], + "logprob_reward": -0.048742204904556274, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048742204904556274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14110735058784485, + 0.046908970922231674, + 0.062140218913555145, + 0.06297624856233597, + 0.039270441979169846, + 0.04939819872379303, + 0.0451681949198246, + 0.07817911356687546, + 0.32154032588005066, + 0.1533108949661255 + ], + "credit_map": [ + 0.035705458372831345, + 0.020503191277384758, + 0.018135299906134605, + 0.009503737092018127, + 0.010332642123103142, + 0.013371664099395275, + 0.017574770376086235, + 0.015980590134859085, + 0.007149236276745796, + 0.004757424350827932, + 0.010269093327224255, + 0.0051017883233726025, + 0.004380826838314533, + 0.0050898827612400055, + 0.003149272408336401, + 0.007011449430137873, + 0.007951827719807625, + 0.004266745876520872, + 0.0076371366158127785, + 0.006037714425474405, + 0.006361260544508696, + 0.008168056607246399, + 0.0067710718140006065, + 0.014946410432457924, + 0.008885746821761131, + 0.007847225293517113, + 0.00674649840220809, + 0.007139170542359352, + 0.00853660050779581, + 0.005960407666862011, + 0.01213162112981081, + 0.00572898518294096, + 0.004111677873879671, + 0.00720038590952754, + 0.0036665445659309626, + 0.003815803909674287, + 0.004801133181899786, + 0.00698053976520896, + 0.003955577965825796, + 0.00473878113552928, + 0.004191981628537178, + 0.0033440743573009968, + 0.01067734882235527, + 0.002501053735613823, + 0.003990985453128815, + 0.005128898657858372, + 0.0030317751225084066, + 0.016532083973288536, + 0.004807840101420879, + 0.008217347785830498, + 0.004352242685854435, + 0.004207184072583914, + 0.009402930736541748, + 0.005521903280168772, + 0.00435111066326499, + 0.004307639319449663, + 0.006911312695592642, + 0.016053155064582825, + 0.008482512086629868, + 0.00874174665659666, + 0.011841847561299801, + 0.00909911934286356, + 0.007065251469612122, + 0.009984171018004417, + 0.022413773462176323, + 0.09292382001876831, + 0.024809181690216064, + 0.023555375635623932, + 0.07907609641551971, + 0.018850177526474, + 0.014506993815302849, + 0.04540490359067917, + 0.017232563346624374, + 0.023373207077383995, + 0.02229575254023075, + 0.025618713349103928, + 0.015246017836034298, + 0.018316585570573807, + 0.01686723157763481, + 0.01436082273721695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.048742204904556274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.029811512678861618 + ], + "opposite_logprob_tokens": [ + -3.5298116207122803 + ], + "logprob_reward": -0.029811512678861618, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029811512678861618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12345931679010391, + 0.2510739862918854, + 0.03197140246629715, + 0.027960920706391335, + 0.022617116570472717, + 0.26577284932136536, + 0.06224483624100685, + 0.08826534450054169, + 0.0961499735713005, + 0.03048422373831272 + ], + "credit_map": [ + 0.020988037809729576, + 0.014277617447078228, + 0.018162304535508156, + 0.012768032029271126, + 0.0098293237388134, + 0.016464130952954292, + 0.01696639321744442, + 0.014003490097820759, + 0.03420741483569145, + 0.020411603152751923, + 0.06743833422660828, + 0.01950828917324543, + 0.019341297447681427, + 0.025571687147021294, + 0.012952625751495361, + 0.05164274573326111, + 0.00439855270087719, + 0.003825789550319314, + 0.0037376212421804667, + 0.00524818804115057, + 0.003992959391325712, + 0.003440829925239086, + 0.004230530932545662, + 0.0030969292856752872, + 0.004330893978476524, + 0.0032390847336500883, + 0.0033654882572591305, + 0.0029256087727844715, + 0.002882347209379077, + 0.0037374289240688086, + 0.004527821205556393, + 0.002952245995402336, + 0.005485244560986757, + 0.0031054294668138027, + 0.003257900010794401, + 0.0015954466070979834, + 0.003038319293409586, + 0.0024343242403119802, + 0.001820355886593461, + 0.0018800980178639293, + 0.020065918564796448, + 0.012761074118316174, + 0.0612744577229023, + 0.010672995820641518, + 0.018266120925545692, + 0.026521336287260056, + 0.016195164993405342, + 0.1000157967209816, + 0.006053899880498648, + 0.012466702610254288, + 0.007645783014595509, + 0.007354292087256908, + 0.007946600206196308, + 0.008438631892204285, + 0.00703789247199893, + 0.005301032215356827, + 0.006688612513244152, + 0.011898607946932316, + 0.012833875603973866, + 0.011919232085347176, + 0.009854278527200222, + 0.010340129025280476, + 0.00942649133503437, + 0.01530411560088396, + 0.006463761441409588, + 0.01706084981560707, + 0.01206648163497448, + 0.01025608740746975, + 0.024106696248054504, + 0.008974676951766014, + 0.011479998007416725, + 0.005741415545344353, + 0.0038474248722195625, + 0.00389679241925478, + 0.0037508029490709305, + 0.004818309098482132, + 0.0037544250953942537, + 0.004276178311556578, + 0.003160342574119568, + 0.00297994795255363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.029811512678861618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933950781822205 + ], + "opposite_logprob_tokens": [ + -0.6933950781822205 + ], + "logprob_reward": -0.6933950781822205, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933950781822205, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17078545689582825, + 0.05669761821627617, + 0.07447325438261032, + 0.07580363005399704, + 0.058187369257211685, + 0.23723956942558289, + 0.04833506792783737, + 0.05705062672495842, + 0.15235956013202667, + 0.0690678060054779 + ], + "credit_map": [ + 0.025133181363344193, + 0.023565348237752914, + 0.021173475310206413, + 0.017127636820077896, + 0.01791507750749588, + 0.024462426081299782, + 0.02049332484602928, + 0.020914996042847633, + 0.008516793139278889, + 0.008090062066912651, + 0.009652133099734783, + 0.0057682772167027, + 0.0078122299164533615, + 0.005611984059214592, + 0.0036959790159016848, + 0.007550158072263002, + 0.010685513727366924, + 0.010079605504870415, + 0.00806872546672821, + 0.010637138038873672, + 0.007131250109523535, + 0.008054393343627453, + 0.012116010300815105, + 0.007700622547417879, + 0.02051519975066185, + 0.006744946353137493, + 0.010994985699653625, + 0.008083438500761986, + 0.008281377144157887, + 0.008220934309065342, + 0.0076431999914348125, + 0.005319550633430481, + 0.007242022082209587, + 0.00685414532199502, + 0.0063139270059764385, + 0.005782888270914555, + 0.008572662249207497, + 0.007863296195864677, + 0.0076884436421096325, + 0.007869983091950417, + 0.014287376776337624, + 0.011921831406652927, + 0.06234791874885559, + 0.00939197838306427, + 0.010871748439967632, + 0.020097361877560616, + 0.013185815885663033, + 0.09513555467128754, + 0.005618654657155275, + 0.00849894154816866, + 0.006623968947678804, + 0.004560356959700584, + 0.008320328779518604, + 0.006600663997232914, + 0.004538313485682011, + 0.0035738381557166576, + 0.007415889762341976, + 0.014239856041967869, + 0.004378076177090406, + 0.005676500964909792, + 0.010897736996412277, + 0.004976415075361729, + 0.004521251656115055, + 0.004944896325469017, + 0.00848692748695612, + 0.046486642211675644, + 0.010095461271703243, + 0.012022421695291996, + 0.03794938698410988, + 0.004680335521697998, + 0.007155097555369139, + 0.025483284145593643, + 0.0072892531752586365, + 0.010260180570185184, + 0.012295546010136604, + 0.013647506013512611, + 0.005052004940807819, + 0.0064071244560182095, + 0.00753497239202261, + 0.006581220775842667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.6933950781822205 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "From an intense first-person twin shot, simulating the perspective of a belligerent young rooster, the camera remains fixed as a brightly feathered opponent charges directly toward the camera. The entire explosive moment is captured using high-speed macro photography and rendered in stunning slow motion. The frame is filled with the opponent rooster, a majestic creature of its imposing size. Its head thrust forward, its neck feathers—a dazzling array of iridescent green and lustrous black—radiating aggressive intent. Its small eyes, tips of black obsidian, reflect the bright sun overhead, locked in a gaze of focused fury. As it thrusts forward, its powerful, scaly yellow legs extend, its sharp, curved claws gleaming in the light. It pushes off explosively from the dusty ground, creating a miniature cloud of dirt and straw particles in the surrounding air. The high-speed capture reveals every tiny detail: the way the air agitates and separates the individual barbs of its breast feathers, the subtle tension of the muscles beneath its skin, and the raw kinetic energy of the impending clash, creating a visceral and beautiful atmosphere of primal conflict.", + "seed": 30870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_3/index_2611_seed_30870_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": -0.14528646245598792, + "vlm_logprob": -0.14528646245598792, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.14528646245598792, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.05172527953982353 + ], + "opposite_logprob_tokens": [ + -3.176725387573242 + ], + "logprob_reward": -0.05172527953982353, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05172527953982353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06291335076093674, + 0.07765742391347885, + 0.09109082072973251, + 0.09872126579284668, + 0.07833172380924225, + 0.04693080112338066, + 0.06193246692419052, + 0.037171635776758194, + 0.4102214574813843, + 0.035029079765081406 + ], + "credit_map": [ + 0.004887757357209921, + 0.008986333385109901, + 0.006888275500386953, + 0.007264014799147844, + 0.013941153883934021, + 0.006060243118554354, + 0.009304998442530632, + 0.005580563563853502, + 0.004535915795713663, + 0.010371547192335129, + 0.003524984233081341, + 0.011767115443944931, + 0.022411681711673737, + 0.004267450887709856, + 0.011299759149551392, + 0.009478959254920483, + 0.008814515545964241, + 0.017956985160708427, + 0.01201894972473383, + 0.009626524522900581, + 0.009939105249941349, + 0.010264300741255283, + 0.006659070961177349, + 0.015811368823051453, + 0.008957638405263424, + 0.00822280440479517, + 0.017612416297197342, + 0.010047665797173977, + 0.007729847449809313, + 0.009764211252331734, + 0.01205209270119667, + 0.024334590882062912, + 0.009297329001128674, + 0.010422160848975182, + 0.00904255360364914, + 0.008921561762690544, + 0.010110422968864441, + 0.010776642709970474, + 0.011464823968708515, + 0.008296218700706959, + 0.00789839867502451, + 0.00804575439542532, + 0.006272595375776291, + 0.006811800878494978, + 0.004247548524290323, + 0.005693417973816395, + 0.004992155823856592, + 0.0029691243544220924, + 0.008073657751083374, + 0.011188067495822906, + 0.0055072614923119545, + 0.009046589955687523, + 0.014026296325027943, + 0.004263782408088446, + 0.006366436369717121, + 0.0034603718668222427, + 0.003929391968995333, + 0.00900644063949585, + 0.0038753822445869446, + 0.004466438665986061, + 0.005542595870792866, + 0.004023339599370956, + 0.002747895661741495, + 0.0035801490303128958, + 0.024079155176877975, + 0.08056022226810455, + 0.08920944482088089, + 0.01447770930826664, + 0.13468340039253235, + 0.022101059556007385, + 0.015823394060134888, + 0.02928704209625721, + 0.004373626317828894, + 0.0026483384426683187, + 0.0050918906927108765, + 0.005206844303756952, + 0.0053099035285413265, + 0.004350157920271158, + 0.0043630399741232395, + 0.003685274627059698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.05172527953982353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.011098556220531464 + ], + "opposite_logprob_tokens": [ + -4.511098384857178 + ], + "logprob_reward": -0.011098556220531464, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011098556220531464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06687460094690323, + 0.05542803928256035, + 0.17006564140319824, + 0.07697863131761551, + 0.08526761084794998, + 0.04855923727154732, + 0.06469981372356415, + 0.031014060601592064, + 0.3612130582332611, + 0.039899297058582306 + ], + "credit_map": [ + 0.00503510283306241, + 0.008069193921983242, + 0.005625118967145681, + 0.010220874100923538, + 0.015246279537677765, + 0.005557948257774115, + 0.01060977391898632, + 0.006510308012366295, + 0.005029137711971998, + 0.00726544251665473, + 0.0043677291832864285, + 0.00896998681128025, + 0.010784230194985867, + 0.003745371475815773, + 0.00785764865577221, + 0.007408499252051115, + 0.04249660298228264, + 0.019846422597765923, + 0.04239819571375847, + 0.011427054181694984, + 0.0170181542634964, + 0.011302621103823185, + 0.010677055455744267, + 0.014899535104632378, + 0.01549992710351944, + 0.009432261809706688, + 0.010603510774672031, + 0.007145254872739315, + 0.007419942412525415, + 0.008306633681058884, + 0.006289135664701462, + 0.012281960807740688, + 0.01198395062237978, + 0.00849915575236082, + 0.010855348780751228, + 0.013186979107558727, + 0.008925185538828373, + 0.012003922834992409, + 0.0131593719124794, + 0.006653694901615381, + 0.007326547987759113, + 0.008083859458565712, + 0.0057779173366725445, + 0.006524260155856609, + 0.006566562224179506, + 0.004974874202162027, + 0.005279581528156996, + 0.00402563763782382, + 0.008153874427080154, + 0.012169391848146915, + 0.004878184758126736, + 0.00869034044444561, + 0.01764291524887085, + 0.004596184939146042, + 0.005465887486934662, + 0.003103035967797041, + 0.0038947565481066704, + 0.005496778525412083, + 0.002790681319311261, + 0.003054134314879775, + 0.0041211131028831005, + 0.0036625575739890337, + 0.0030100021976977587, + 0.004984037484973669, + 0.023571284487843513, + 0.07858990132808685, + 0.06408815085887909, + 0.012097339145839214, + 0.1327497959136963, + 0.020351605489850044, + 0.014140532352030277, + 0.015624427236616611, + 0.004564322996884584, + 0.0027196980081498623, + 0.008458909578621387, + 0.0060198381543159485, + 0.004341470077633858, + 0.00435043266043067, + 0.005028539802879095, + 0.00441608764231205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.011098556220531464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.07016896456480026 + ], + "opposite_logprob_tokens": [ + -2.695168972015381 + ], + "logprob_reward": -0.07016896456480026, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07016896456480026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07683238387107849, + 0.027578040957450867, + 0.3047579526901245, + 0.10291460156440735, + 0.1887560933828354, + 0.03720778599381447, + 0.02123756892979145, + 0.023733874782919884, + 0.11462589353322983, + 0.10235586762428284 + ], + "credit_map": [ + 0.0063200234435498714, + 0.010999372228980064, + 0.006983471568673849, + 0.010148846544325352, + 0.015107751823961735, + 0.007107228506356478, + 0.01280566118657589, + 0.0073600150644779205, + 0.0022122429218143225, + 0.0040567005053162575, + 0.0025227635633200407, + 0.003480747574940324, + 0.005997590254992247, + 0.0018246958497911692, + 0.0038042040541768074, + 0.003679093671962619, + 0.05821574479341507, + 0.022599605843424797, + 0.08476299792528152, + 0.016968166455626488, + 0.02205130085349083, + 0.021132545545697212, + 0.018696868792176247, + 0.06033070385456085, + 0.011750158853828907, + 0.009894746355712414, + 0.013610371388494968, + 0.008530345745384693, + 0.012399423867464066, + 0.010921661742031574, + 0.013538012281060219, + 0.02226986736059189, + 0.019046403467655182, + 0.019240863621234894, + 0.025737138465046883, + 0.027691803872585297, + 0.021601570770144463, + 0.026079120114445686, + 0.031147362664341927, + 0.018211808055639267, + 0.005206466652452946, + 0.004010283388197422, + 0.0046288189478218555, + 0.005096225067973137, + 0.00398583197966218, + 0.003915068227797747, + 0.005473222117871046, + 0.00489186542108655, + 0.002790413098409772, + 0.003931564744561911, + 0.0016434412682428956, + 0.0026529051829129457, + 0.0035619691479951143, + 0.002317099366337061, + 0.0026391884312033653, + 0.0017009866423904896, + 0.0025816361885517836, + 0.0037015231791883707, + 0.002560713328421116, + 0.0029472538735717535, + 0.0043772184289991856, + 0.0021655773743987083, + 0.002296281512826681, + 0.003103667637333274, + 0.01054125651717186, + 0.022555453702807426, + 0.01898384653031826, + 0.0056947688572108746, + 0.038457825779914856, + 0.006559778470546007, + 0.005334717687219381, + 0.006498236209154129, + 0.015246856026351452, + 0.007218354847282171, + 0.02044980227947235, + 0.011403153650462627, + 0.007877458818256855, + 0.010447895154356956, + 0.018006231635808945, + 0.011706103570759296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.07016896456480026 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.07004671543836594 + ], + "opposite_logprob_tokens": [ + -2.6950466632843018 + ], + "logprob_reward": -0.07004671543836594, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004671543836594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0537697933614254, + 0.03807387501001358, + 0.18037928640842438, + 0.12988032400608063, + 0.22800157964229584, + 0.06548678874969482, + 0.03651810809969902, + 0.05154998600482941, + 0.0654204785823822, + 0.15091973543167114 + ], + "credit_map": [ + 0.006130461115390062, + 0.006258056499063969, + 0.00742671312764287, + 0.005386370699852705, + 0.010293114930391312, + 0.005157116334885359, + 0.007273795083165169, + 0.005844171158969402, + 0.005328514147549868, + 0.003897464834153652, + 0.004005142953246832, + 0.00471344031393528, + 0.00596489617601037, + 0.002199287759140134, + 0.007008813787251711, + 0.004956317134201527, + 0.028102949261665344, + 0.020244453102350235, + 0.04402921348810196, + 0.012163743376731873, + 0.014614272862672806, + 0.012377935461699963, + 0.01118350401520729, + 0.03766322508454323, + 0.02384866774082184, + 0.015109382569789886, + 0.01752353273332119, + 0.010324670001864433, + 0.014466219581663609, + 0.013285391964018345, + 0.010241677053272724, + 0.025080792605876923, + 0.023928910493850708, + 0.026023030281066895, + 0.028992854058742523, + 0.02650664560496807, + 0.032904841005802155, + 0.03245735168457031, + 0.035658128559589386, + 0.021529819816350937, + 0.00985728669911623, + 0.009328316897153854, + 0.009692729450762272, + 0.007620537653565407, + 0.007375091779977083, + 0.007356566842645407, + 0.008130034431815147, + 0.0061262319795787334, + 0.005159836262464523, + 0.00673060305416584, + 0.004892660304903984, + 0.004682266619056463, + 0.00632517272606492, + 0.0030639423057436943, + 0.0036909477785229683, + 0.0019726778846234083, + 0.00574876694008708, + 0.006131819915026426, + 0.0046838815324008465, + 0.005359575618058443, + 0.006366423796862364, + 0.006932928692549467, + 0.005423101596534252, + 0.010903488844633102, + 0.006504642311483622, + 0.010465585626661777, + 0.013294832780957222, + 0.003660329384729266, + 0.015288613736629486, + 0.0037264369893819094, + 0.003984185867011547, + 0.00849585048854351, + 0.023510517552495003, + 0.01145551260560751, + 0.02292459085583687, + 0.014725563116371632, + 0.021626291796565056, + 0.020982345566153526, + 0.019007721915841103, + 0.01668720319867134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.07004671543836594 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective fixed from a first-person viewpoint simulating a rooster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an opponent rooster with iridescent green and lustrous black neck feathers charging directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the opponent rooster's legs described as scaly yellow with sharp, curved claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rooster push off from a dusty ground creating a cloud of dirt and straw particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in high-speed macro photography with 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": [ + -0.5233927965164185 + ], + "opposite_logprob_tokens": [ + -0.8983927965164185 + ], + "logprob_reward": -0.5233927965164185, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233927965164185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07286754995584488, + 0.023661430925130844, + 0.09033232927322388, + 0.047508321702480316, + 0.06486386805772781, + 0.08934769034385681, + 0.039103247225284576, + 0.031461816281080246, + 0.4856397807598114, + 0.055213943123817444 + ], + "credit_map": [ + 0.007430823054164648, + 0.008180730976164341, + 0.009274407289922237, + 0.009636422619223595, + 0.012724049389362335, + 0.006592568941414356, + 0.01260833814740181, + 0.006420211400836706, + 0.002015758538618684, + 0.003284038743004203, + 0.0022613792680203915, + 0.004060681443661451, + 0.00427982909604907, + 0.001470352290198207, + 0.0033863282296806574, + 0.002903064014390111, + 0.00829119049012661, + 0.012122472748160362, + 0.016022926196455956, + 0.006756129674613476, + 0.007714204490184784, + 0.008779100142419338, + 0.007404080592095852, + 0.023242229595780373, + 0.006380632985383272, + 0.004511622712016106, + 0.004713206551969051, + 0.005383085925132036, + 0.005939831491559744, + 0.0058229221031069756, + 0.010216652415692806, + 0.004540370311588049, + 0.010338188149034977, + 0.00889565795660019, + 0.00709508266299963, + 0.008716335520148277, + 0.008573575876653194, + 0.008942588232457638, + 0.006079515907913446, + 0.006222923751920462, + 0.015966815873980522, + 0.014889180660247803, + 0.011532378382980824, + 0.014046644791960716, + 0.008417676202952862, + 0.009442618116736412, + 0.008381533436477184, + 0.006670850794762373, + 0.004606880713254213, + 0.006826556287705898, + 0.009013303555548191, + 0.0045284973457455635, + 0.0058653452433645725, + 0.002795232692733407, + 0.0035529949236661196, + 0.0019144387915730476, + 0.0034296882804483175, + 0.005008991342037916, + 0.004049757029861212, + 0.0037365390453487635, + 0.0051766932010650635, + 0.00374980759806931, + 0.0023862740490585566, + 0.003924064803868532, + 0.033196356147527695, + 0.09640369564294815, + 0.09439313411712646, + 0.01795004867017269, + 0.1719130128622055, + 0.026142889633774757, + 0.022111214697360992, + 0.023529449477791786, + 0.004275821149349213, + 0.00325262569822371, + 0.013087503612041473, + 0.007554900832474232, + 0.005219205282628536, + 0.008243906311690807, + 0.0054641421884298325, + 0.008115836419165134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.5233927965164185 + } + ], + "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/0000028/group_4/index_tensor([3913, 3913])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_4/index_tensor([3913, 3913])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cb6d36e57c1429c61e6c55fae3c4e8c2d23a5095 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_4/index_tensor([3913, 3913])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, shot from an upward angle just below the surface of the water, the camera sways with a barely perceptible oscillation, immersing the viewer in the water. At the center of the frame, a quaint wooden rowboat floats alone on the vast, calm expanse. Soft, golden light from dawn slants down, casting a warm patina on the weathered, dark wood of the boat, revealing the mottled paint and damp wood grain on the sides. As the silky, smooth waves roll in rhythmically and gently, the boat's bow gently rises and then slowly descends, carving an elegant arc across the shimmering water. The distant horizon is shrouded in a thin morning mist, and the sky shimmers in hues of pale pink and violet. The scene's palette is serene and subdued, with light and shadow constantly shifting across the water and the boat, creating a sense of profound tranquility, contemplation, and solitude.", + "seed": 10303, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_4/index_3913_seed_10303_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": -0.16233754027634859, + "vlm_logprob": -0.16233754027634859, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.16233754027634859, + "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 long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012263844721019268 + ], + "opposite_logprob_tokens": [ + -5.637263774871826 + ], + "logprob_reward": -0.012263844721019268, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.012263844721019268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13027043640613556, + 0.0769648626446724, + 0.041927214711904526, + 0.06479880213737488, + 0.03279070556163788, + 0.04935261607170105, + 0.05900415778160095, + 0.036034468561410904, + 0.08436159044504166, + 0.4244951605796814 + ], + "credit_map": [ + 0.03335137292742729, + 0.0074187093414366245, + 0.017832672223448753, + 0.00858442485332489, + 0.01118722464889288, + 0.015039423480629921, + 0.012322789989411831, + 0.024533802643418312, + 0.008173386566340923, + 0.004328442737460136, + 0.006895067170262337, + 0.0036379483062773943, + 0.005504640284925699, + 0.019063197076320648, + 0.004592717159539461, + 0.024769466370344162, + 0.008227313868701458, + 0.0028826415073126554, + 0.004486324731260538, + 0.0033521659206598997, + 0.004526349250227213, + 0.007893997244536877, + 0.0034113202709704638, + 0.00714710121974349, + 0.015671029686927795, + 0.005494848359376192, + 0.005810347851365805, + 0.0046357992105185986, + 0.006660530809313059, + 0.009823462925851345, + 0.006879049818962812, + 0.009823732078075409, + 0.00637004105374217, + 0.0034247785806655884, + 0.0041591571643948555, + 0.0023061339743435383, + 0.0036538972053676844, + 0.00470899511128664, + 0.0037418885622173548, + 0.004425815772265196, + 0.006910085678100586, + 0.004580119624733925, + 0.006777139380574226, + 0.0055122883059084415, + 0.00502575421705842, + 0.004754648078233004, + 0.010465684346854687, + 0.005326896905899048, + 0.014755770564079285, + 0.0046415491960942745, + 0.008129906840622425, + 0.004852014128118753, + 0.0074028498493134975, + 0.0071591660380363464, + 0.004468523431569338, + 0.007594377268105745, + 0.0054361638613045216, + 0.002694615162909031, + 0.006964321248233318, + 0.003401309484615922, + 0.003433874575421214, + 0.005439993925392628, + 0.003493496449664235, + 0.005170696880668402, + 0.012316271662712097, + 0.007366088684648275, + 0.009616606868803501, + 0.007290464360266924, + 0.012478743679821491, + 0.010862665250897408, + 0.00981269683688879, + 0.014618050307035446, + 0.040206193923950195, + 0.01757015846669674, + 0.13139894604682922, + 0.027747100219130516, + 0.02095053344964981, + 0.12640421092510223, + 0.023075135424733162, + 0.03714286908507347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012263844721019268 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578150629997253 + ], + "opposite_logprob_tokens": [ + -0.6328150629997253 + ], + "logprob_reward": -0.7578150629997253, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578150629997253, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21136264503002167, + 0.06304451078176498, + 0.0663318783044815, + 0.03381149098277092, + 0.01878298632800579, + 0.03148670122027397, + 0.06481281667947769, + 0.03269454836845398, + 0.13788118958473206, + 0.33979126811027527 + ], + "credit_map": [ + 0.0416594073176384, + 0.015279061160981655, + 0.02605573832988739, + 0.015135282650589943, + 0.01822999306023121, + 0.033260684460401535, + 0.02385442703962326, + 0.037888053804636, + 0.009891539812088013, + 0.002910095965489745, + 0.007580264937132597, + 0.002776698675006628, + 0.0023304405622184277, + 0.012942727655172348, + 0.0026303359773010015, + 0.021982409060001373, + 0.007228027563542128, + 0.002257531275972724, + 0.005847113206982613, + 0.0027851162012666464, + 0.0033846660517156124, + 0.017744245007634163, + 0.0028381962329149246, + 0.024246981367468834, + 0.00534775760024786, + 0.0022746026515960693, + 0.003792468924075365, + 0.002251646714285016, + 0.004825566429644823, + 0.004734968300908804, + 0.004348296206444502, + 0.006236185785382986, + 0.0030587119981646538, + 0.0015498986467719078, + 0.002706793136894703, + 0.0014555350644513965, + 0.0022247550077736378, + 0.002328235190361738, + 0.0016171495663002133, + 0.003841907950118184, + 0.004988422617316246, + 0.0038529112935066223, + 0.004440885502845049, + 0.0032738333102315664, + 0.004337126389145851, + 0.003842191072180867, + 0.003405304392799735, + 0.003346027573570609, + 0.012571291998028755, + 0.005035902373492718, + 0.007974411360919476, + 0.006697052624076605, + 0.007651814725250006, + 0.007530428469181061, + 0.0059389155358076096, + 0.011412996798753738, + 0.007621982600539923, + 0.0027687866240739822, + 0.0038962834514677525, + 0.0030492693185806274, + 0.004020106513053179, + 0.0034350850619375706, + 0.00341883790679276, + 0.0044841947965323925, + 0.02109597623348236, + 0.009382162243127823, + 0.009791516698896885, + 0.014179788529872894, + 0.013486746698617935, + 0.028729788959026337, + 0.013789664022624493, + 0.027425553649663925, + 0.049916598945856094, + 0.015524937771260738, + 0.08619227260351181, + 0.017197096720337868, + 0.018325325101614, + 0.1020561009645462, + 0.016563234850764275, + 0.03401567041873932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.7578150629997253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.19801397621631622, + 0.10519541800022125, + 0.047647785395383835, + 0.05675049498677254, + 0.055514756590127945, + 0.09550215303897858, + 0.043851934373378754, + 0.08201780915260315, + 0.09183856844902039, + 0.22366714477539062 + ], + "credit_map": [ + 0.030948052182793617, + 0.013979891315102577, + 0.027421483770012856, + 0.017755428329110146, + 0.014177260920405388, + 0.022191965952515602, + 0.021390242502093315, + 0.05014963075518608, + 0.016255969181656837, + 0.0056245834566652775, + 0.010761192999780178, + 0.0058357673697173595, + 0.007959380745887756, + 0.020122796297073364, + 0.006730490364134312, + 0.031905222684144974, + 0.010557479225099087, + 0.002502782503142953, + 0.005010727792978287, + 0.0034526775125414133, + 0.0059426166117191315, + 0.010042019188404083, + 0.004164934158325195, + 0.005974545609205961, + 0.011908373795449734, + 0.0037454834673553705, + 0.006896822713315487, + 0.0044786324724555016, + 0.006568242330104113, + 0.008988884277641773, + 0.006512407213449478, + 0.007651641499251127, + 0.012553910724818707, + 0.0057325223460793495, + 0.007573983632028103, + 0.0042449659667909145, + 0.006049978081136942, + 0.008374173194169998, + 0.0038569073658436537, + 0.0071283080615103245, + 0.011701579205691814, + 0.01210984867066145, + 0.017047584056854248, + 0.008765479549765587, + 0.013392078690230846, + 0.009555095806717873, + 0.010941758751869202, + 0.011988719925284386, + 0.0071210418827831745, + 0.00418102415278554, + 0.005122028756886721, + 0.006287833675742149, + 0.0055277650244534016, + 0.004617439117282629, + 0.0046349773183465, + 0.006359821185469627, + 0.014458861202001572, + 0.005073969718068838, + 0.011392745189368725, + 0.007534377742558718, + 0.01137610338628292, + 0.01168007031083107, + 0.0096536660566926, + 0.01084800437092781, + 0.01665055751800537, + 0.0061208936385810375, + 0.009420963935554028, + 0.006632374599575996, + 0.008268138393759727, + 0.020786188542842865, + 0.00828399509191513, + 0.01567544974386692, + 0.043477896600961685, + 0.015535306185483932, + 0.04317658394575119, + 0.012989314272999763, + 0.016187582165002823, + 0.05324656516313553, + 0.014344029128551483, + 0.02470986358821392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "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 shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633090019226074 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.007633090019226074, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633090019226074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15251582860946655, + 0.06381311267614365, + 0.06234467402100563, + 0.047050878405570984, + 0.02791658788919449, + 0.05465953052043915, + 0.035143814980983734, + 0.050563350319862366, + 0.29461026191711426, + 0.2113819420337677 + ], + "credit_map": [ + 0.02500474639236927, + 0.01322000753134489, + 0.01999782770872116, + 0.016576800495386124, + 0.014175471849739552, + 0.021731112152338028, + 0.018468739464879036, + 0.023341123014688492, + 0.010061873123049736, + 0.006494887638837099, + 0.00805745366960764, + 0.006036857608705759, + 0.009373406879603863, + 0.008826007135212421, + 0.007646070793271065, + 0.007316555827856064, + 0.008370030671358109, + 0.002995074260979891, + 0.005784503184258938, + 0.0029421767685562372, + 0.007140378002077341, + 0.015173929743468761, + 0.004702857229858637, + 0.015235723927617073, + 0.00759295467287302, + 0.003651009639725089, + 0.005207156762480736, + 0.003613719716668129, + 0.005356197711080313, + 0.006593076046556234, + 0.004483863245695829, + 0.010552902705967426, + 0.006056031212210655, + 0.0019673416391015053, + 0.0037676666397601366, + 0.0017075776122510433, + 0.0027098057325929403, + 0.0038225543685257435, + 0.002530464669689536, + 0.005355146713554859, + 0.009801666252315044, + 0.004089499358087778, + 0.006600312422960997, + 0.0029114929493516684, + 0.008339137770235538, + 0.007645166479051113, + 0.004560737404972315, + 0.010711519978940487, + 0.007433433085680008, + 0.003082987619563937, + 0.004792219027876854, + 0.003962075337767601, + 0.004201547242701054, + 0.004377143457531929, + 0.0028647601138800383, + 0.0044296481646597385, + 0.013158510439097881, + 0.002699128817766905, + 0.006049490068107843, + 0.003727802075445652, + 0.005033836700022221, + 0.006943536456674337, + 0.004298756830394268, + 0.008652286604046822, + 0.04928779602050781, + 0.01452846173197031, + 0.018818482756614685, + 0.01820726878941059, + 0.014755531214177608, + 0.08577516674995422, + 0.017304640263319016, + 0.07593290507793427, + 0.024528400972485542, + 0.01656370423734188, + 0.06003156304359436, + 0.018671665340662003, + 0.012524963356554508, + 0.03731175884604454, + 0.0183149091899395, + 0.023434961214661598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.007633090019226074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627766579389572 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627766579389572, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627766579389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2671233117580414, + 0.15924152731895447, + 0.07486838102340698, + 0.030302228406071663, + 0.0652998685836792, + 0.03288228064775467, + 0.057404838502407074, + 0.082368403673172, + 0.048823919147253036, + 0.18168526887893677 + ], + "credit_map": [ + 0.06938420981168747, + 0.013061032630503178, + 0.03538436070084572, + 0.016213467344641685, + 0.01821904629468918, + 0.03789878636598587, + 0.02124621532857418, + 0.055716175585985184, + 0.024303551763296127, + 0.007693823426961899, + 0.017103880643844604, + 0.006653402931988239, + 0.009114259853959084, + 0.031107759103178978, + 0.008854213170707226, + 0.054410647600889206, + 0.012188471853733063, + 0.002494171494618058, + 0.005035571753978729, + 0.0028787394985556602, + 0.003996152896434069, + 0.022387392818927765, + 0.0031514987349510193, + 0.022736383602023125, + 0.006425248924642801, + 0.002325626788660884, + 0.003241390222683549, + 0.0018811292247846723, + 0.003321234369650483, + 0.004552212543785572, + 0.0030105365440249443, + 0.005544847808778286, + 0.01364987250417471, + 0.0050690555945038795, + 0.01051755715161562, + 0.0038188996259123087, + 0.007835865020751953, + 0.007758029270917177, + 0.005845857784152031, + 0.01080473605543375, + 0.00602028239518404, + 0.0028192109894007444, + 0.0050725508481264114, + 0.003283440601080656, + 0.003934885840862989, + 0.004518659319728613, + 0.003927662502974272, + 0.003305588150396943, + 0.010829834267497063, + 0.0042194039560854435, + 0.0066413264721632, + 0.005516125354915857, + 0.0048753926530480385, + 0.007930138148367405, + 0.004906435031443834, + 0.012486185878515244, + 0.019116779789328575, + 0.0040026018396019936, + 0.010824555531144142, + 0.00826454721391201, + 0.006755200680345297, + 0.009123236872255802, + 0.008719703182578087, + 0.015561777167022228, + 0.005875153932720423, + 0.0025292031932622194, + 0.003473697230219841, + 0.003453608136624098, + 0.003920617047697306, + 0.013716766610741615, + 0.0035458144266158342, + 0.012309056706726551, + 0.03580819070339203, + 0.010437867604196072, + 0.036455538123846054, + 0.010101581923663616, + 0.00901877786964178, + 0.0506967194378376, + 0.009328469634056091, + 0.019838115200400352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.007627766579389572 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from an upward angle just below the surface of the water, the camera sways with a barely perceptible oscillation, immersing the viewer in the water. At the center of the frame, a quaint wooden rowboat floats alone on the vast, calm expanse. Soft, golden light from dawn slants down, casting a warm patina on the weathered, dark wood of the boat, revealing the mottled paint and damp wood grain on the sides. As the silky, smooth waves roll in rhythmically and gently, the boat's bow gently rises and then slowly descends, carving an elegant arc across the shimmering water. The distant horizon is shrouded in a thin morning mist, and the sky shimmers in hues of pale pink and violet. The scene's palette is serene and subdued, with light and shadow constantly shifting across the water and the boat, creating a sense of profound tranquility, contemplation, and solitude.", + "seed": 10304, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_4/index_3913_seed_10304_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": -0.12826697658747435, + "vlm_logprob": -0.12826697658747435, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.12826697658747435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010042497888207436 + ], + "opposite_logprob_tokens": [ + -5.510042667388916 + ], + "logprob_reward": -0.010042497888207436, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010042497888207436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08645733445882797, + 0.01952480338513851, + 0.015360433608293533, + 0.05042952671647072, + 0.03179032355546951, + 0.01596715860068798, + 0.6901047229766846, + 0.0470096729695797, + 0.016280341893434525, + 0.02707570046186447 + ], + "credit_map": [ + 0.009269094094634056, + 0.005788792856037617, + 0.010806812904775143, + 0.007545189466327429, + 0.01575843058526516, + 0.017805863171815872, + 0.008438936434686184, + 0.011044221930205822, + 0.0024520710576325655, + 0.001649569720029831, + 0.0026703050825744867, + 0.0012186485109850764, + 0.0038254933897405863, + 0.003224159125238657, + 0.0017961421981453896, + 0.0026884139515459538, + 0.0015937722055241466, + 0.0016245271544903517, + 0.0020638764835894108, + 0.0015928928041830659, + 0.002292314311489463, + 0.00213709007948637, + 0.0019951381254941225, + 0.0020608219783753157, + 0.005388279911130667, + 0.003083926858380437, + 0.008412126451730728, + 0.0036052230279892683, + 0.006177139468491077, + 0.014043410308659077, + 0.003121144836768508, + 0.006598277948796749, + 0.002982351230457425, + 0.00266235857270658, + 0.004623411223292351, + 0.004901557229459286, + 0.004285857547074556, + 0.0044180904515087605, + 0.00379166123457253, + 0.004125034902244806, + 0.0016394603298977017, + 0.0018989056115970016, + 0.002308174269273877, + 0.0017637896817177534, + 0.002595221158117056, + 0.002099528443068266, + 0.0018411040073260665, + 0.0018209763802587986, + 0.12055674195289612, + 0.028824185952544212, + 0.13276030123233795, + 0.02943304181098938, + 0.0661906898021698, + 0.18967080116271973, + 0.035554371774196625, + 0.0871146097779274, + 0.006954388227313757, + 0.002783948788419366, + 0.007922602817416191, + 0.0031409712973982096, + 0.004914605524390936, + 0.011722407303750515, + 0.003532088128849864, + 0.006038661114871502, + 0.0016012859996408224, + 0.0014301164774224162, + 0.002245463663712144, + 0.0027980695012956858, + 0.0018265829421579838, + 0.0017703602788969874, + 0.0021019724663347006, + 0.002506490796804428, + 0.0023549410980194807, + 0.002800848102197051, + 0.0035895886830985546, + 0.0042135464027523994, + 0.0025105399545282125, + 0.0042357416823506355, + 0.0035174647346138954, + 0.003853030502796173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.010042497888207436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761731266975403 + ], + "opposite_logprob_tokens": [ + -0.8261731266975403 + ], + "logprob_reward": -0.5761731266975403, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761731266975403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07870287448167801, + 0.03910207748413086, + 0.021800875663757324, + 0.015035195276141167, + 0.05677216872572899, + 0.019007258117198944, + 0.6981136798858643, + 0.009994699619710445, + 0.0362473763525486, + 0.02522382326424122 + ], + "credit_map": [ + 0.00733574666082859, + 0.007980694063007832, + 0.01142896804958582, + 0.009016459807753563, + 0.008276084437966347, + 0.012205516919493675, + 0.010254967026412487, + 0.012204433791339397, + 0.002719536889344454, + 0.0028581663500517607, + 0.005650753155350685, + 0.0028068730607628822, + 0.007986471988260746, + 0.005578694399446249, + 0.006088871043175459, + 0.0054127066396176815, + 0.0019478873582556844, + 0.0018824745202437043, + 0.002992610214278102, + 0.0015504530165344477, + 0.00580960838124156, + 0.004139202646911144, + 0.0016087674302980304, + 0.0018698700005188584, + 0.0016495269956067204, + 0.0012624174123629928, + 0.002331423107534647, + 0.002132816007360816, + 0.001312481937929988, + 0.002403972437605262, + 0.00158877728972584, + 0.002353778574615717, + 0.006960094440728426, + 0.004760341253131628, + 0.008459272794425488, + 0.004956259857863188, + 0.00887084100395441, + 0.010410042479634285, + 0.004578227177262306, + 0.007777079939842224, + 0.0027719889767467976, + 0.0014702572952955961, + 0.002731367014348507, + 0.002060798229649663, + 0.002332210773602128, + 0.0034202076494693756, + 0.002135501243174076, + 0.002084926003590226, + 0.1242537796497345, + 0.02938050404191017, + 0.1367686688899994, + 0.028492437675595284, + 0.06551932543516159, + 0.19020602107048035, + 0.03567919880151749, + 0.08781372755765915, + 0.001100779278203845, + 0.0014689007075503469, + 0.0014463237021118402, + 0.0013394317356869578, + 0.0010408980306237936, + 0.0009543693158775568, + 0.0013872722629457712, + 0.0012567240046337247, + 0.005500468425452709, + 0.002420596079900861, + 0.005587080493569374, + 0.0030863885767757893, + 0.004586141090840101, + 0.006082696374505758, + 0.004139179363846779, + 0.004844822455197573, + 0.0021541675087064505, + 0.0027928801719099283, + 0.0033119956497102976, + 0.004132757894694805, + 0.002286458620801568, + 0.0028464950155466795, + 0.00352921849116683, + 0.004169846419245005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.5761731266975403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038095004856586456 + ], + "opposite_logprob_tokens": [ + -3.288094997406006 + ], + "logprob_reward": -0.038095004856586456, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038095004856586456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03724907338619232, + 0.023461299017071724, + 0.015278179198503494, + 0.013310519978404045, + 0.014334707520902157, + 0.01345199067145586, + 0.7959184050559998, + 0.024956107139587402, + 0.03371668979525566, + 0.028323039412498474 + ], + "credit_map": [ + 0.005092427600175142, + 0.004174333531409502, + 0.006398024968802929, + 0.0031196391209959984, + 0.004940562881529331, + 0.005559803918004036, + 0.003935611806809902, + 0.0040286690928041935, + 0.0020444521214812994, + 0.002280554035678506, + 0.0030920577701181173, + 0.0019387883367016912, + 0.005851358640938997, + 0.003710557473823428, + 0.002146597485989332, + 0.002396932104602456, + 0.0015516571002081037, + 0.0012852015206590295, + 0.002083485946059227, + 0.0014237629948183894, + 0.0022849556989967823, + 0.003213822143152356, + 0.0015454088570550084, + 0.0018898847047239542, + 0.001784979598596692, + 0.0010456383461132646, + 0.002184301847591996, + 0.0012784524587914348, + 0.0027444574516266584, + 0.0016517093172296882, + 0.0013726226752623916, + 0.0012483573518693447, + 0.0012986529618501663, + 0.0011181272566318512, + 0.0020903085824102163, + 0.0020662618335336447, + 0.0020601253490895033, + 0.0019399981247261167, + 0.0015346214640885592, + 0.0022266111336648464, + 0.0015166696393862367, + 0.0010133128380402923, + 0.0020710639655590057, + 0.0012415883829817176, + 0.0022929003462195396, + 0.0020334310829639435, + 0.0012045549228787422, + 0.002078467747196555, + 0.13906733691692352, + 0.03536282852292061, + 0.15178914368152618, + 0.03414890915155411, + 0.07689239084720612, + 0.21922920644283295, + 0.04140263795852661, + 0.09802594035863876, + 0.003057150635868311, + 0.0023186393082141876, + 0.004115091636776924, + 0.0022304030135273933, + 0.002470201812684536, + 0.0054711708799004555, + 0.002614519326016307, + 0.0026789302937686443, + 0.004086853936314583, + 0.003258169861510396, + 0.00437561608850956, + 0.003432685974985361, + 0.006749020889401436, + 0.003430361161008477, + 0.004628983326256275, + 0.0037549955304712057, + 0.002491989405825734, + 0.004125326406210661, + 0.003398837987333536, + 0.004735782742500305, + 0.002553359605371952, + 0.004335861653089523, + 0.00354650616645813, + 0.003135372418910265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.038095004856586456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077453382313251 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077453382313251, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077453382313251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04381302744150162, + 0.02121572010219097, + 0.024042056873440742, + 0.05638914927840233, + 0.012348740361630917, + 0.0520816408097744, + 0.7155884504318237, + 0.011872517876327038, + 0.03742064908146858, + 0.0252279844135046 + ], + "credit_map": [ + 0.0048637837171554565, + 0.003652954241260886, + 0.006702256388962269, + 0.004473933018743992, + 0.004739368800073862, + 0.009564081206917763, + 0.0047914194874465466, + 0.005025227554142475, + 0.0015499242581427097, + 0.0022034079302102327, + 0.0028942537028342485, + 0.0017639180878177285, + 0.004793448373675346, + 0.003473380347713828, + 0.0019368254579603672, + 0.0026005618274211884, + 0.0015796574298292398, + 0.00155038689263165, + 0.0033391746692359447, + 0.001548800035379827, + 0.005491270683705807, + 0.0055953869596123695, + 0.002343954984098673, + 0.002593425801023841, + 0.006536352448165417, + 0.0038853876758366823, + 0.00915820524096489, + 0.004115723073482513, + 0.0059211598709225655, + 0.015499511733651161, + 0.004263536538928747, + 0.0070092747919261456, + 0.0011391942389309406, + 0.0010054000886157155, + 0.001956166699528694, + 0.00151717325206846, + 0.002058028243482113, + 0.0014510583132505417, + 0.0015192622086033225, + 0.001702457549981773, + 0.006101253442466259, + 0.004130967427045107, + 0.007080733310431242, + 0.005471366457641125, + 0.009014417417347431, + 0.008207939565181732, + 0.005585078615695238, + 0.006489885970950127, + 0.1276172399520874, + 0.027136890217661858, + 0.14452031254768372, + 0.027773460373282433, + 0.06490708142518997, + 0.19404225051403046, + 0.0355069600045681, + 0.09408432990312576, + 0.0018742483807727695, + 0.00077608204446733, + 0.0018179966136813164, + 0.0008079695980995893, + 0.001247405307367444, + 0.0027046443428844213, + 0.0007263878360390663, + 0.0019177845679223537, + 0.003323921235278249, + 0.003560020122677088, + 0.004305756650865078, + 0.0047658528201282024, + 0.005571150686591864, + 0.006649650167673826, + 0.004435619805008173, + 0.004808675032109022, + 0.0020586210303008556, + 0.003447884228080511, + 0.002992849098518491, + 0.004751981236040592, + 0.0023212896194308996, + 0.002442028373479843, + 0.0034875585697591305, + 0.0037257722578942776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.011077453382313251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946800112724304 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946800112724304, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946800112724304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08646717667579651, + 0.028060272336006165, + 0.02166900783777237, + 0.012427030131220818, + 0.012033860199153423, + 0.05250008404254913, + 0.6506843566894531, + 0.045285698026418686, + 0.02473524957895279, + 0.06613729894161224 + ], + "credit_map": [ + 0.011859231628477573, + 0.005842496175318956, + 0.012838151305913925, + 0.004858643747866154, + 0.009910108521580696, + 0.021421581506729126, + 0.007406429387629032, + 0.012330539524555206, + 0.0023682236205786467, + 0.0016900168266147375, + 0.003952660132199526, + 0.0027451589703559875, + 0.005296542774885893, + 0.00490832794457674, + 0.003250709269195795, + 0.003848635358735919, + 0.002400461118668318, + 0.0016594835324212909, + 0.0030214027501642704, + 0.0023565564770251513, + 0.005427889991551638, + 0.0028218363877385855, + 0.0022156338673084974, + 0.001765744760632515, + 0.0016078958287835121, + 0.001055757631547749, + 0.0018397819949313998, + 0.0015911245718598366, + 0.0013543418608605862, + 0.00225269328802824, + 0.0013054627925157547, + 0.0014199732104316354, + 0.0012990550603717566, + 0.0011800648644566536, + 0.001951786456629634, + 0.0015979547752067447, + 0.001997877610847354, + 0.0011792770819738507, + 0.0014744121581315994, + 0.0013534323079511523, + 0.0067515019327402115, + 0.005008989945054054, + 0.007122289855033159, + 0.008418750949203968, + 0.005872135050594807, + 0.006980038248002529, + 0.005897666793316603, + 0.006448710337281227, + 0.11453241109848022, + 0.026329156011343002, + 0.12537036836147308, + 0.027658462524414062, + 0.061265043914318085, + 0.18052396178245544, + 0.035625699907541275, + 0.07937928289175034, + 0.006349921692162752, + 0.0022475782316178083, + 0.007751320954412222, + 0.0022437700536102057, + 0.004926865920424461, + 0.012685559689998627, + 0.0030112951062619686, + 0.006069387309253216, + 0.002843171590939164, + 0.0019565955735743046, + 0.003092332510277629, + 0.0030013800133019686, + 0.0030698163900524378, + 0.004952737595885992, + 0.0023297597654163837, + 0.003489457769319415, + 0.008039239794015884, + 0.005753755569458008, + 0.009053077548742294, + 0.008190655149519444, + 0.006957098841667175, + 0.01106198225170374, + 0.006666096393018961, + 0.010415397584438324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.005946800112724304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from an upward angle just below the surface of the water, the camera sways with a barely perceptible oscillation, immersing the viewer in the water. At the center of the frame, a quaint wooden rowboat floats alone on the vast, calm expanse. Soft, golden light from dawn slants down, casting a warm patina on the weathered, dark wood of the boat, revealing the mottled paint and damp wood grain on the sides. As the silky, smooth waves roll in rhythmically and gently, the boat's bow gently rises and then slowly descends, carving an elegant arc across the shimmering water. The distant horizon is shrouded in a thin morning mist, and the sky shimmers in hues of pale pink and violet. The scene's palette is serene and subdued, with light and shadow constantly shifting across the water and the boat, creating a sense of profound tranquility, contemplation, and solitude.", + "seed": 10305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_4/index_3913_seed_10305_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": -0.03864262569695711, + "vlm_logprob": -0.03864262569695711, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03864262569695711, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07596861571073532 + ], + "opposite_logprob_tokens": [ + -2.8259685039520264 + ], + "logprob_reward": -0.07596861571073532, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07596861571073532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10301605612039566, + 0.02233322151005268, + 0.025761740282177925, + 0.008019832894206047, + 0.7678165435791016, + 0.005408602766692638, + 0.0051046875305473804, + 0.0016033912543207407, + 0.04954134672880173, + 0.011394470930099487 + ], + "credit_map": [ + 0.0054367887787520885, + 0.010972967371344566, + 0.0027459098491817713, + 0.019487883895635605, + 0.026013685390353203, + 0.005530817434191704, + 0.028570562601089478, + 0.004257462918758392, + 0.002586542861536145, + 0.0037399400025606155, + 0.0017847948474809527, + 0.002501599956303835, + 0.0033360819797962904, + 0.0024683470837771893, + 0.004047556314617395, + 0.0018683610251173377, + 0.0027767433784902096, + 0.004575017839670181, + 0.002960178069770336, + 0.0015695334877818823, + 0.00887413788586855, + 0.002440202981233597, + 0.0012288149446249008, + 0.0013371141394600272, + 0.0008252877159975469, + 0.0015135958092287183, + 0.0009211397264152765, + 0.0004363676125649363, + 0.0018171989358961582, + 0.001485892804339528, + 0.0004956864286214113, + 0.0005246649961918592, + 0.03451023995876312, + 0.056270670145750046, + 0.23673486709594727, + 0.03926336020231247, + 0.31325897574424744, + 0.04851828143000603, + 0.02024002932012081, + 0.01902027428150177, + 0.0007106985431164503, + 0.0007386807119473815, + 0.0007061074138619006, + 0.000499986344948411, + 0.0006506083882413805, + 0.0009539475431665778, + 0.0005361487856134772, + 0.0006124261999502778, + 0.00041203369619324803, + 0.0007865898660384119, + 0.000462899828562513, + 0.0004035596502944827, + 0.0006076970603317022, + 0.0014148835325613618, + 0.0005370294093154371, + 0.0004799951275344938, + 0.00021288359130267054, + 0.00014681604807265103, + 0.00020063076226506382, + 0.00015675950271543115, + 0.00018888118211179972, + 0.0002300571504747495, + 0.0002741967618931085, + 0.00019316650286782533, + 0.006182335317134857, + 0.007866601459681988, + 0.0026352438144385815, + 0.002167696598917246, + 0.0078760776668787, + 0.016535015776753426, + 0.0017913618357852101, + 0.004487020429223776, + 0.0008449418819509447, + 0.002246694639325142, + 0.0010684654116630554, + 0.0007137266802601516, + 0.001802359358407557, + 0.0028341966681182384, + 0.0008487500017508864, + 0.0010353383840993047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.07596861571073532 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700308233499527 + ], + "opposite_logprob_tokens": [ + -2.695030927658081 + ], + "logprob_reward": -0.0700308233499527, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700308233499527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07093116641044617, + 0.06655263155698776, + 0.04285048320889473, + 0.04542579501867294, + 0.5402005910873413, + 0.02666577510535717, + 0.015750577673316002, + 0.00552792614325881, + 0.13397224247455597, + 0.05212279409170151 + ], + "credit_map": [ + 0.006420650985091925, + 0.00995275191962719, + 0.0037539105396717787, + 0.012152714654803276, + 0.010378493927419186, + 0.00594174163416028, + 0.01769043132662773, + 0.004640479572117329, + 0.008463182486593723, + 0.010469391010701656, + 0.007193933241069317, + 0.006116242613643408, + 0.010889166034758091, + 0.00790057610720396, + 0.009730731137096882, + 0.005789405200630426, + 0.004589240066707134, + 0.006761109922081232, + 0.004626329056918621, + 0.002950509311631322, + 0.012167003937065601, + 0.004469121806323528, + 0.0022632330656051636, + 0.0050239358097314835, + 0.004127031192183495, + 0.006232120096683502, + 0.0064922356978058815, + 0.0021559500601142645, + 0.018331779167056084, + 0.003774842945858836, + 0.0021028865594416857, + 0.002208949998021126, + 0.024241778999567032, + 0.03935261815786362, + 0.16551272571086884, + 0.02773219905793667, + 0.21755287051200867, + 0.035153426229953766, + 0.01497887633740902, + 0.015676122158765793, + 0.0035139445681124926, + 0.0032856555189937353, + 0.0036555954720824957, + 0.0029607887845486403, + 0.0035587467718869448, + 0.002883513458073139, + 0.004218836780637503, + 0.0025886951480060816, + 0.0014947282616049051, + 0.0018915354739874601, + 0.0012710093287751079, + 0.002143687568604946, + 0.0018280957592651248, + 0.0027617295272648335, + 0.0023400799836963415, + 0.002019713167101145, + 0.0007571667083539069, + 0.0007742884336039424, + 0.0005978158442303538, + 0.0004345675406511873, + 0.0006670469883829355, + 0.000805889256298542, + 0.0007241051062010229, + 0.0007670463528484106, + 0.016517166048288345, + 0.021326914429664612, + 0.007067847065627575, + 0.005584155675023794, + 0.0216814037412405, + 0.045216478407382965, + 0.004664144944399595, + 0.01191413402557373, + 0.004093023017048836, + 0.01130661554634571, + 0.002387311775237322, + 0.0026649662759155035, + 0.011464699171483517, + 0.015530158765614033, + 0.0019758762791752815, + 0.0027001448906958103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.0700308233499527 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796238988637924 + ], + "opposite_logprob_tokens": [ + -3.5297961235046387 + ], + "logprob_reward": -0.029796238988637924, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796238988637924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08183371275663376, + 0.013498024083673954, + 0.013027744367718697, + 0.010954920202493668, + 0.7919785380363464, + 0.0069647617638111115, + 0.007794701494276524, + 0.00480806827545166, + 0.031175214797258377, + 0.03796429559588432 + ], + "credit_map": [ + 0.005010824650526047, + 0.014567424543201923, + 0.0018798448145389557, + 0.010603250004351139, + 0.02721085399389267, + 0.004583117086440325, + 0.015293058007955551, + 0.002685334999114275, + 0.0013875672593712807, + 0.0025433513801544905, + 0.0011174629908055067, + 0.001492466777563095, + 0.002210525795817375, + 0.0012475461699068546, + 0.0025545060634613037, + 0.0009445961331948638, + 0.0013318824348971248, + 0.0028876487631350756, + 0.0013338763965293765, + 0.0013098791241645813, + 0.0024552466347813606, + 0.0013548157876357436, + 0.0010160485981032252, + 0.001338345929980278, + 0.0010526771657168865, + 0.0014471921604126692, + 0.0015790382167324424, + 0.0006247546407394111, + 0.0037062836345285177, + 0.0009913422400131822, + 0.0007579182856716216, + 0.0007957129855640233, + 0.035515692085027695, + 0.05800985172390938, + 0.2440863847732544, + 0.04050500690937042, + 0.3227146863937378, + 0.05001478269696236, + 0.02119891159236431, + 0.019933195784687996, + 0.0006794958608224988, + 0.0010379087179899216, + 0.0008330735727213323, + 0.0007356764399446547, + 0.0007513911696150899, + 0.0011142534203827381, + 0.0012569313403218985, + 0.0005560305435210466, + 0.0008067505550570786, + 0.0011557043762877584, + 0.0006286846473813057, + 0.0007013401482254267, + 0.0009520311723463237, + 0.0018641703063622117, + 0.000942891463637352, + 0.0007431280100718141, + 0.0005590946529991925, + 0.000611065945122391, + 0.0005215191631577909, + 0.00048698042519390583, + 0.0005629808874800801, + 0.0005698936874978244, + 0.0006790116312913597, + 0.0008175215916708112, + 0.003843489335849881, + 0.005037790630012751, + 0.0016367308562621474, + 0.0014484170824289322, + 0.004888296592980623, + 0.010412084870040417, + 0.0011613067472353578, + 0.0027470970526337624, + 0.0028972909785807133, + 0.008781813085079193, + 0.0014216924319043756, + 0.002327271969988942, + 0.007150010205805302, + 0.01209266297519207, + 0.0013035015435889363, + 0.0019900468178093433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.029796238988637924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.0464121513068676, + 0.029994789510965347, + 0.032309018075466156, + 0.020873308181762695, + 0.7090240120887756, + 0.022274300456047058, + 0.015362695790827274, + 0.009814569726586342, + 0.05733751505613327, + 0.05659766122698784 + ], + "credit_map": [ + 0.0045714653097093105, + 0.011396394111216068, + 0.002823561429977417, + 0.005613917019218206, + 0.007761620916426182, + 0.0037186460103839636, + 0.008236439898610115, + 0.002290106611326337, + 0.003950208425521851, + 0.004681473597884178, + 0.0028853241819888353, + 0.0025082877837121487, + 0.005653166212141514, + 0.004010201431810856, + 0.0037317138630896807, + 0.002574414247646928, + 0.0028510428965091705, + 0.005233712028712034, + 0.0030590298119932413, + 0.002880383748561144, + 0.00886573363095522, + 0.0033698100596666336, + 0.0023826230317354202, + 0.003666685661301017, + 0.0017726992955431342, + 0.004272256046533585, + 0.0019091021968051791, + 0.0014002131065353751, + 0.002814735285937786, + 0.005080572795122862, + 0.0016072478611022234, + 0.00201648217625916, + 0.031643904745578766, + 0.052366871386766434, + 0.21887418627738953, + 0.036214716732501984, + 0.2896103262901306, + 0.04451457038521767, + 0.018610280007123947, + 0.01718916930258274, + 0.002456807531416416, + 0.0027625092770904303, + 0.0034163277596235275, + 0.002459587762132287, + 0.0028632094617933035, + 0.0026891285087913275, + 0.003456271253526211, + 0.002170459134504199, + 0.001606979756616056, + 0.0026946880389004946, + 0.0012392783537507057, + 0.0010066245449706912, + 0.0020740781910717487, + 0.003999648615717888, + 0.0013007988454774022, + 0.0014406002592295408, + 0.0012029390782117844, + 0.0014130671042948961, + 0.00133121304679662, + 0.0007754045072942972, + 0.0014860545052215457, + 0.0015054539544507861, + 0.0010835510911419988, + 0.00101688620634377, + 0.006970483809709549, + 0.009384040720760822, + 0.002870423486456275, + 0.0025838310830295086, + 0.009052718989551067, + 0.01947587914764881, + 0.0020631388761103153, + 0.004936997778713703, + 0.004678815603256226, + 0.01420860830694437, + 0.0018639053450897336, + 0.0029128652531653643, + 0.009438538923859596, + 0.018150167539715767, + 0.0017839003121480346, + 0.0035608571488410234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 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 shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783650748431683 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783650748431683, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783650748431683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04676264524459839, + 0.01764673739671707, + 0.014114804565906525, + 0.029450619593262672, + 0.8037686944007874, + 0.013635779730975628, + 0.009267931804060936, + 0.011392246931791306, + 0.015430566854774952, + 0.03853001073002815 + ], + "credit_map": [ + 0.0034301781561225653, + 0.00797977577894926, + 0.0014500408433377743, + 0.005456032697111368, + 0.015081807971000671, + 0.003041472053155303, + 0.008050930686295033, + 0.002272410085424781, + 0.002473666099831462, + 0.003110740100964904, + 0.0023215552791953087, + 0.0015548299998044968, + 0.0027068324852734804, + 0.0021394104696810246, + 0.0016844155034050345, + 0.0016552865272387862, + 0.0014506339794024825, + 0.0022819472942501307, + 0.001643811003305018, + 0.0008496444206684828, + 0.004406302701681852, + 0.001330879982560873, + 0.0009082963224500418, + 0.0012432883959263563, + 0.0031626636628061533, + 0.004055970814079046, + 0.0040205661207437515, + 0.0013128923019394279, + 0.010848444886505604, + 0.0034968398977071047, + 0.001048659672960639, + 0.0015045813051983714, + 0.036089878529310226, + 0.05904316529631615, + 0.24798732995986938, + 0.041165295988321304, + 0.3282720744609833, + 0.050381843000650406, + 0.021283233538269997, + 0.019545890390872955, + 0.0023297416046261787, + 0.0014544387813657522, + 0.0024894087109714746, + 0.0014974077930673957, + 0.0017064489657059312, + 0.0014493154594674706, + 0.0014114538207650185, + 0.0012975645950064063, + 0.001622540527023375, + 0.0012388628674671054, + 0.0010905949166044593, + 0.000717861985322088, + 0.00131007912568748, + 0.0017971326597034931, + 0.0007479817722924054, + 0.0007428784156218171, + 0.0018859230913221836, + 0.0020749650429934263, + 0.0014886660501360893, + 0.001049613580107689, + 0.0014649268705397844, + 0.0014116870006546378, + 0.0008593096281401813, + 0.001157155609689653, + 0.001820279285311699, + 0.0026673099491745234, + 0.0008874639170244336, + 0.0008123906445689499, + 0.002272194717079401, + 0.004947662819176912, + 0.0006631053984165192, + 0.0013601604150608182, + 0.003003301564604044, + 0.008814896456897259, + 0.002207257319241762, + 0.002278307918459177, + 0.008282024413347244, + 0.011028884910047054, + 0.0012525173369795084, + 0.0016628220910206437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.009783650748431683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from an upward angle just below the surface of the water, the camera sways with a barely perceptible oscillation, immersing the viewer in the water. At the center of the frame, a quaint wooden rowboat floats alone on the vast, calm expanse. Soft, golden light from dawn slants down, casting a warm patina on the weathered, dark wood of the boat, revealing the mottled paint and damp wood grain on the sides. As the silky, smooth waves roll in rhythmically and gently, the boat's bow gently rises and then slowly descends, carving an elegant arc across the shimmering water. The distant horizon is shrouded in a thin morning mist, and the sky shimmers in hues of pale pink and violet. The scene's palette is serene and subdued, with light and shadow constantly shifting across the water and the boat, creating a sense of profound tranquility, contemplation, and solitude.", + "seed": 10306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_4/index_3913_seed_10306_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": -0.1654866537079215, + "vlm_logprob": -0.1654866537079215, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.1654866537079215, + "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 long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015730109065771103 + ], + "opposite_logprob_tokens": [ + -5.140729904174805 + ], + "logprob_reward": -0.015730109065771103, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.015730109065771103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17512871325016022, + 0.07431644201278687, + 0.071739561855793, + 0.17930307984352112, + 0.06438352912664413, + 0.06102565675973892, + 0.05937083438038826, + 0.10681496560573578, + 0.11939790844917297, + 0.08851935714483261 + ], + "credit_map": [ + 0.01777263917028904, + 0.015730250626802444, + 0.024472041055560112, + 0.029653454199433327, + 0.010258500464260578, + 0.019649190828204155, + 0.02014702372252941, + 0.037445589900016785, + 0.011070634238421917, + 0.005551803391426802, + 0.010408653877675533, + 0.009200537577271461, + 0.007541840896010399, + 0.00603104941546917, + 0.014515215530991554, + 0.009996694512665272, + 0.010133703239262104, + 0.007180997636169195, + 0.009495408274233341, + 0.012311434373259544, + 0.006804448552429676, + 0.006662797648459673, + 0.008923778310418129, + 0.01022698450833559, + 0.025190941989421844, + 0.015199842862784863, + 0.02771659754216671, + 0.02446739934384823, + 0.0102899931371212, + 0.01603471301496029, + 0.01836869865655899, + 0.042034875601530075, + 0.014271148480474949, + 0.006267629098147154, + 0.010397091507911682, + 0.007916432805359364, + 0.005410005338490009, + 0.004952650982886553, + 0.004307943861931562, + 0.010860621929168701, + 0.013048102147877216, + 0.006204870529472828, + 0.008348087780177593, + 0.008226330392062664, + 0.004531909711658955, + 0.00433738436549902, + 0.006625168025493622, + 0.009703801944851875, + 0.008322169072926044, + 0.006400405894964933, + 0.00680053373798728, + 0.010485101491212845, + 0.004752468317747116, + 0.006330230738967657, + 0.006492583081126213, + 0.009787342511117458, + 0.015952279791235924, + 0.00942018162459135, + 0.013139376416802406, + 0.016931666061282158, + 0.00894408393651247, + 0.008577005006372929, + 0.014453714713454247, + 0.019396642223000526, + 0.01626950316131115, + 0.009577658958733082, + 0.01580987684428692, + 0.024141201749444008, + 0.006248073652386665, + 0.010837412439286709, + 0.017065858468413353, + 0.01944831758737564, + 0.01090759877115488, + 0.0066705504432320595, + 0.009715194813907146, + 0.017694441601634026, + 0.005682829301804304, + 0.006216024048626423, + 0.0172264464199543, + 0.014406261034309864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.015730109065771103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578293681144714 + ], + "opposite_logprob_tokens": [ + -0.6328293681144714 + ], + "logprob_reward": -0.7578293681144714, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578293681144714, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4209243059158325, + 0.07590062916278839, + 0.07260153442621231, + 0.08008954674005508, + 0.07275472581386566, + 0.06761918216943741, + 0.03904068470001221, + 0.04496684670448303, + 0.058298297226428986, + 0.067804254591465 + ], + "credit_map": [ + 0.046944789588451385, + 0.03806711733341217, + 0.05868229642510414, + 0.0798737034201622, + 0.023933157324790955, + 0.03507629781961441, + 0.044070176780223846, + 0.09427673369646072, + 0.013838531449437141, + 0.008374808356165886, + 0.012796512804925442, + 0.010530578903853893, + 0.004730449989438057, + 0.007833542302250862, + 0.005733728874474764, + 0.012062469497323036, + 0.011579648591578007, + 0.006756297312676907, + 0.011262264102697372, + 0.01149551197886467, + 0.004990282002836466, + 0.008881641551852226, + 0.006064829882234335, + 0.01157105527818203, + 0.009686185978353024, + 0.007616477087140083, + 0.012278768233954906, + 0.013923783786594868, + 0.007877577096223831, + 0.008784585632383823, + 0.006906087975949049, + 0.013016076758503914, + 0.014215999282896519, + 0.005015665665268898, + 0.010920747183263302, + 0.011615071445703506, + 0.005421299021691084, + 0.007156342267990112, + 0.005250512156635523, + 0.01315908133983612, + 0.01358122006058693, + 0.006462973542511463, + 0.008636733517050743, + 0.011764534749090672, + 0.005920745432376862, + 0.007375282235443592, + 0.005799722857773304, + 0.008077972568571568, + 0.0064693051390349865, + 0.00463902996852994, + 0.004218852147459984, + 0.005344454199075699, + 0.003420858643949032, + 0.0052356659434735775, + 0.0038140241522341967, + 0.005898491013795137, + 0.010285447351634502, + 0.005545736290514469, + 0.0066961380653083324, + 0.005589784123003483, + 0.0035609302576631308, + 0.004788750316947699, + 0.003677879925817251, + 0.004822179675102234, + 0.01188417337834835, + 0.005588094238191843, + 0.011479289270937443, + 0.008004098199307919, + 0.004836945794522762, + 0.005088577512651682, + 0.004840870853513479, + 0.006576246116310358, + 0.007202275563031435, + 0.005786452908068895, + 0.008146370761096478, + 0.014255024492740631, + 0.004974796902388334, + 0.004364912863820791, + 0.014580088667571545, + 0.00849433708935976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.7578293681144714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369971364736557 + ], + "opposite_logprob_tokens": [ + -3.4086997509002686 + ], + "logprob_reward": -0.03369971364736557, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369971364736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21977098286151886, + 0.09191204607486725, + 0.06872321665287018, + 0.11224100738763809, + 0.07162190973758698, + 0.07359037548303604, + 0.08201345801353455, + 0.09768591821193695, + 0.12544655799865723, + 0.05699444189667702 + ], + "credit_map": [ + 0.0313272550702095, + 0.030889812856912613, + 0.03091122768819332, + 0.03349416330456734, + 0.0227164588868618, + 0.022073829546570778, + 0.018950801342725754, + 0.029407447203993797, + 0.013782325200736523, + 0.010644444264471531, + 0.013753646984696388, + 0.009437679313123226, + 0.007373274303972721, + 0.009059179574251175, + 0.012839940376579762, + 0.015021557919681072, + 0.007252143230289221, + 0.006387660745531321, + 0.009318839758634567, + 0.012205500155687332, + 0.006201047915965319, + 0.005780734587460756, + 0.01124782394617796, + 0.010329467244446278, + 0.012937101535499096, + 0.011386193335056305, + 0.01887134462594986, + 0.01844911463558674, + 0.009982860647141933, + 0.008539590053260326, + 0.013286820612847805, + 0.01878798007965088, + 0.008862029761075974, + 0.007806533947587013, + 0.008877360261976719, + 0.010564393363893032, + 0.006702046375721693, + 0.007330271881073713, + 0.0063546509481966496, + 0.015124628320336342, + 0.01325538381934166, + 0.006077802740037441, + 0.008871478959918022, + 0.009885236620903015, + 0.006115240510553122, + 0.007181624881923199, + 0.010235581547021866, + 0.011968027800321579, + 0.010698138736188412, + 0.009798502549529076, + 0.009362941607832909, + 0.012652760371565819, + 0.007659207098186016, + 0.0082815270870924, + 0.011814809404313564, + 0.011745576746761799, + 0.013886774890124798, + 0.009479128755629063, + 0.010354410856962204, + 0.016805211082100868, + 0.009154909290373325, + 0.016666816547513008, + 0.010601558722555637, + 0.01073711272329092, + 0.023206433281302452, + 0.011805497109889984, + 0.01784188114106655, + 0.022554170340299606, + 0.012234949506819248, + 0.010832030326128006, + 0.011554131284356117, + 0.015417476184666157, + 0.010132206603884697, + 0.0041281795129179955, + 0.009185872040688992, + 0.009139218367636204, + 0.004748211242258549, + 0.005326847545802593, + 0.006924159359186888, + 0.007409750949591398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.03369971364736557 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254430040717125 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.01254430040717125, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254430040717125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22989147901535034, + 0.09245478361845016, + 0.14710989594459534, + 0.08708328008651733, + 0.07168977707624435, + 0.09203556180000305, + 0.07779988646507263, + 0.05055811256170273, + 0.12145806103944778, + 0.029919156804680824 + ], + "credit_map": [ + 0.024748604744672775, + 0.02650841698050499, + 0.028216619044542313, + 0.053362276405096054, + 0.017708204686641693, + 0.020011376589536667, + 0.02198495715856552, + 0.03735104948282242, + 0.02083011530339718, + 0.007377969566732645, + 0.018651578575372696, + 0.010491029359400272, + 0.00599945243448019, + 0.009802194312214851, + 0.008355927653610706, + 0.010946521535515785, + 0.03293503075838089, + 0.012580190785229206, + 0.02594583109021187, + 0.01555610354989767, + 0.010395736433565617, + 0.017092455178499222, + 0.012779554352164268, + 0.019825000315904617, + 0.010384215041995049, + 0.011708637699484825, + 0.012098954990506172, + 0.014976615086197853, + 0.009149156510829926, + 0.009124554693698883, + 0.006744937039911747, + 0.012896210886538029, + 0.01101752556860447, + 0.008890016935765743, + 0.010231554508209229, + 0.012283829972147942, + 0.00591261126101017, + 0.006981909740716219, + 0.0063721416518092155, + 0.010000189766287804, + 0.01757737807929516, + 0.01447636540979147, + 0.010387367568910122, + 0.011387985199689865, + 0.011281034909188747, + 0.009711923077702522, + 0.008451677858829498, + 0.008761840872466564, + 0.017186852172017097, + 0.007770421449095011, + 0.012182936072349548, + 0.01032568421214819, + 0.006645856890827417, + 0.007607846055179834, + 0.00705815339460969, + 0.00902213342487812, + 0.005685028620064259, + 0.007636505179107189, + 0.005834722425788641, + 0.008936997503042221, + 0.0043274289928376675, + 0.006911406293511391, + 0.004727283027023077, + 0.006498747039586306, + 0.02794739231467247, + 0.011637651361525059, + 0.020167306065559387, + 0.01380800548940897, + 0.012235651724040508, + 0.0106608672067523, + 0.010959500446915627, + 0.014041690155863762, + 0.0054951012134552, + 0.002258619526401162, + 0.0055227032862603664, + 0.0041498104110360146, + 0.0024119867011904716, + 0.002579113934189081, + 0.0037073383573442698, + 0.003794485703110695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.01254430040717125 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a wooden rowboat at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an upward angle just below the surface of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera sway with a barely perceptible oscillation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's bow rising and descending in a gentle, rhythmic arc due to silky smooth waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the distant horizon shrouded in a thin morning mist with a sky in hues of pale pink and violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2060466706752777, + 0.15970002114772797, + 0.0739889070391655, + 0.053319014608860016, + 0.05765953287482262, + 0.1670856773853302, + 0.036880262196063995, + 0.06657741218805313, + 0.035506680607795715, + 0.14323580265045166 + ], + "credit_map": [ + 0.04378430172801018, + 0.018149618059396744, + 0.03286890313029289, + 0.02473173663020134, + 0.012126355431973934, + 0.020300723612308502, + 0.016167255118489265, + 0.03791775926947594, + 0.0436234250664711, + 0.0120478430762887, + 0.029822422191500664, + 0.015688301995396614, + 0.009988028556108475, + 0.01506491657346487, + 0.010557148605585098, + 0.022907936945557594, + 0.016595713794231415, + 0.00674818130210042, + 0.011977393180131912, + 0.010486503131687641, + 0.00559218181297183, + 0.007063955068588257, + 0.0054984618909657, + 0.010026521049439907, + 0.010346731171011925, + 0.005266794469207525, + 0.008675072342157364, + 0.009462005458772182, + 0.0038517224602401257, + 0.004495859611779451, + 0.005015118047595024, + 0.006205713376402855, + 0.01483015064150095, + 0.004737403243780136, + 0.0113252904266119, + 0.00604213448241353, + 0.004418532829731703, + 0.00485195592045784, + 0.0031809306237846613, + 0.008273130282759666, + 0.03874098137021065, + 0.01685495674610138, + 0.024339746683835983, + 0.026040371507406235, + 0.01782057248055935, + 0.010826081037521362, + 0.013592707924544811, + 0.018870260566473007, + 0.006684344727545977, + 0.00380023499019444, + 0.005286834202706814, + 0.0051458789967000484, + 0.0033779144287109375, + 0.004264470189809799, + 0.004066201858222485, + 0.004254380706697702, + 0.01758929155766964, + 0.006364993751049042, + 0.01158083975315094, + 0.008717360906302929, + 0.0051021454855799675, + 0.0059948815032839775, + 0.00517716258764267, + 0.00605073943734169, + 0.005831524729728699, + 0.003249367233365774, + 0.004620390012860298, + 0.005157196428626776, + 0.002533447463065386, + 0.0027327376883476973, + 0.004382403567433357, + 0.006999614182859659, + 0.032379455864429474, + 0.01326955296099186, + 0.022256121039390564, + 0.022000666707754135, + 0.015363211743533611, + 0.009473824873566628, + 0.01602555625140667, + 0.012467416003346443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.007629777304828167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_5/index_tensor([5398, 5398])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_5/index_tensor([5398, 5398])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..86ee561d06b03c3012c0f479d51d0d98b1aad12b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_5/index_tensor([5398, 5398])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fixed low-angle medium shot captures a massive giant panda, its black and white fur languidly curled up deep in a bamboo forest. Its plump body rests gently on a thick, emerald green bamboo pole, its face a look of contentment and tranquility. Warm golden sunlight filters through the dense bamboo foliage overhead, casting countless swaying, mottled specks of light on its fluffy back and the surrounding leaf-strewn ground. Each beam is distinct, and tiny particles of light seem to drift slowly in the air. Surrounding it is an endless, towering sea of bamboo, its leaves a rich, emerald green, their veins each shimmering in the backlight. The entire scene is imbued with Makoto Shinkai's signature delicate textures and beautiful lighting, with warm, bright tones and sharp contrasts, creating a soothing atmosphere of profound tranquility, a feeling of almost frozen time.", + "seed": 11881, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_5/index_5398_seed_11881_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": -0.016883843252435325, + "vlm_logprob": -0.016883843252435325, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.016883843252435325, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025153683964163065 + ], + "opposite_logprob_tokens": [ + -6.7525153160095215 + ], + "logprob_reward": -0.0025153683964163065, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0025153683964163065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016371794044971466, + 0.008486970327794552, + 0.008411563001573086, + 0.8486640453338623, + 0.06866098940372467, + 0.014228477142751217, + 0.026515111327171326, + 0.00429430790245533, + 0.001982493558898568, + 0.0023842828813940287 + ], + "credit_map": [ + 0.002110274275764823, + 0.0019232617923989892, + 0.002110932022333145, + 0.0016221600817516446, + 0.0016971419099718332, + 0.0026253480464220047, + 0.0016361926682293415, + 0.0026464827824383974, + 0.0015224649105221033, + 0.0007912684814073145, + 0.001344606513157487, + 0.0009722971590235829, + 0.0007939105853438377, + 0.0016310294158756733, + 0.0005399274523369968, + 0.0008914657519198954, + 0.0009031239897012711, + 0.0006254995241761208, + 0.0007278417469933629, + 0.0008115306263789535, + 0.0005051164189353585, + 0.002339523984119296, + 0.0003983791102655232, + 0.0021005463786423206, + 0.09648529440164566, + 0.0331965871155262, + 0.037436578422784805, + 0.03924306109547615, + 0.04734143987298012, + 0.290627658367157, + 0.03262101858854294, + 0.27171239256858826, + 0.008225840516388416, + 0.0026337241288274527, + 0.003066894831135869, + 0.0034113607835024595, + 0.004145829472690821, + 0.023698801174759865, + 0.0027341267559677362, + 0.020744407549500465, + 0.0015029619680717587, + 0.0007881107158027589, + 0.000852195720653981, + 0.0012811297783628106, + 0.0008048508316278458, + 0.0042578307911753654, + 0.0007737173000350595, + 0.003967679105699062, + 0.002677787560969591, + 0.001085981260985136, + 0.002216682303696871, + 0.0016544526442885399, + 0.0014054819475859404, + 0.008351011201739311, + 0.0011286013759672642, + 0.007995110005140305, + 0.00045015430077910423, + 0.0002367253036936745, + 0.0002342437655897811, + 0.00034934436553157866, + 0.00031484029022976756, + 0.0013817129656672478, + 0.0002794739557430148, + 0.0010478125186637044, + 0.00026522777625359595, + 0.00018504213949199766, + 0.0002532648795749992, + 0.0002209551603300497, + 0.00021137518342584372, + 0.0003462083695922047, + 0.0002604660112410784, + 0.00023995382071007043, + 0.0002280588960275054, + 0.00031253648921847343, + 0.00029301768518052995, + 0.0002865178103093058, + 0.00030730446451343596, + 0.0003039204457309097, + 0.00036640375037677586, + 0.0002865232527256012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.0025153683964163065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589664578437805 + ], + "opposite_logprob_tokens": [ + -3.895589590072632 + ], + "logprob_reward": -0.020589664578437805, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589664578437805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016149194911122322, + 0.005333947017788887, + 0.009351109154522419, + 0.885496199131012, + 0.03430090472102165, + 0.0042964015156030655, + 0.017682652920484543, + 0.00662986608222127, + 0.00381123716942966, + 0.01694849506020546 + ], + "credit_map": [ + 0.0027319665532559156, + 0.0017095387447625399, + 0.002889816416427493, + 0.0017069161403924227, + 0.0014444496482610703, + 0.0025137339252978563, + 0.0011544673470780253, + 0.0019983076490461826, + 0.0008298360626213253, + 0.0003634617605712265, + 0.000955579336732626, + 0.000760757306125015, + 0.0004762858443427831, + 0.0009557350422255695, + 0.0003506842767819762, + 0.0006416075630113482, + 0.0010624055285006762, + 0.0006218263297341764, + 0.0005386149859987199, + 0.0005538976402021945, + 0.0005515695083886385, + 0.0028067384846508503, + 0.00046845804899930954, + 0.0027475988026708364, + 0.1004837229847908, + 0.03449154272675514, + 0.0391872376203537, + 0.04092750325798988, + 0.04927739500999451, + 0.3035621643066406, + 0.03389490768313408, + 0.2836717963218689, + 0.004094370640814304, + 0.0013218011008575559, + 0.0015002806903794408, + 0.0015021300641819835, + 0.0020190004725009203, + 0.012056231498718262, + 0.0013686862075701356, + 0.010438406839966774, + 0.0004796281282324344, + 0.0004626015143003315, + 0.0006138178869150579, + 0.0006192438886500895, + 0.00036028187605552375, + 0.0006947573274374008, + 0.0003597163304220885, + 0.0007063547964207828, + 0.0019039507023990154, + 0.0008403469691984355, + 0.0016722324071452022, + 0.0012399370316416025, + 0.0010130968876183033, + 0.005352185107767582, + 0.0007808768423274159, + 0.0048800278455019, + 0.0005752734141424298, + 0.0007958057103678584, + 0.0007398035377264023, + 0.0014140260173007846, + 0.0006444823811762035, + 0.0009508819784969091, + 0.0007848538225516677, + 0.0007247399189509451, + 0.0005042172851972282, + 0.00036196736618876457, + 0.0004877296451013535, + 0.0003393959777895361, + 0.00028683379059657454, + 0.0006115507567301393, + 0.0003646456461865455, + 0.0008548968471586704, + 0.001667633536271751, + 0.0021165197249501944, + 0.001547849620692432, + 0.0036169574595987797, + 0.0019428747473284602, + 0.002059740712866187, + 0.002357576508074999, + 0.0016393427504226565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.020589664578437805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021968772634863853 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021968772634863853, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021968772634863853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1307774931192398, + 0.05982406064867973, + 0.08786507695913315, + 0.3954893946647644, + 0.22239181399345398, + 0.03173022344708443, + 0.037353433668613434, + 0.013153032399713993, + 0.008615529164671898, + 0.012800008058547974 + ], + "credit_map": [ + 0.022349361330270767, + 0.010400219820439816, + 0.024210890755057335, + 0.013631305657327175, + 0.011500632390379906, + 0.02226702868938446, + 0.008703433908522129, + 0.017714612185955048, + 0.009923665784299374, + 0.004625512752681971, + 0.00944670382887125, + 0.007081423420459032, + 0.004331744741648436, + 0.011296126060187817, + 0.004482324235141277, + 0.008636548183858395, + 0.010688968002796173, + 0.004281582776457071, + 0.006891504395753145, + 0.004300881642848253, + 0.004818696063011885, + 0.026348793879151344, + 0.004033103585243225, + 0.026501532644033432, + 0.045004017651081085, + 0.01563270017504692, + 0.018081054091453552, + 0.018303068354725838, + 0.02233782783150673, + 0.13591845333576202, + 0.0151445297524333, + 0.12506768107414246, + 0.026012670248746872, + 0.008661170490086079, + 0.009082098491489887, + 0.009870489127933979, + 0.012835565954446793, + 0.07734712958335876, + 0.008973328396677971, + 0.06960932910442352, + 0.003695674706250429, + 0.0016617777291685343, + 0.0022814502008259296, + 0.0025255237706005573, + 0.0017313360003754497, + 0.010107273235917091, + 0.0016181181417778134, + 0.00810906384140253, + 0.0040282816626131535, + 0.002109373454004526, + 0.003447305876761675, + 0.002017076127231121, + 0.002171957166865468, + 0.01078846026211977, + 0.0016535315662622452, + 0.011137440800666809, + 0.0014866479905322194, + 0.0012081643799319863, + 0.001662178081460297, + 0.0010563372634351254, + 0.0012373978970572352, + 0.0029410654678940773, + 0.0012419559061527252, + 0.00231928494758904, + 0.0009749348973855376, + 0.0007984553230926394, + 0.0007602970581501722, + 0.001389108831062913, + 0.000762131588999182, + 0.0017827472183853388, + 0.0008551593055017292, + 0.0012926936615258455, + 0.0013362972531467676, + 0.0019161717500537634, + 0.0013855022843927145, + 0.002813305938616395, + 0.0013271122006699443, + 0.0016128403367474675, + 0.0013392532709985971, + 0.001069522462785244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.0021968772634863853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065552830696106 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065552830696106, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065552830696106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039637062698602676, + 0.011741125956177711, + 0.012847095727920532, + 0.804523229598999, + 0.095448337495327, + 0.006599804852157831, + 0.015142794698476791, + 0.005004788748919964, + 0.0038704401813447475, + 0.005185310263186693 + ], + "credit_map": [ + 0.005669423379004002, + 0.004054040648043156, + 0.00642492575570941, + 0.005103214178234339, + 0.0037577692419290543, + 0.006094482261687517, + 0.0032965224236249924, + 0.005236682016402483, + 0.0019997600466012955, + 0.0008528924663551152, + 0.0024363058619201183, + 0.0009262771345674992, + 0.0008049886091612279, + 0.002428477630019188, + 0.0005318472976796329, + 0.0017605756875127554, + 0.0016700030537322164, + 0.0006491206586360931, + 0.0008327565155923367, + 0.0005821824888698757, + 0.0007368977530859411, + 0.004035353194922209, + 0.0004559337394312024, + 0.0038848486728966236, + 0.09169236570596695, + 0.03168220445513725, + 0.035781219601631165, + 0.036988940089941025, + 0.04501327499747276, + 0.2761019468307495, + 0.030810508877038956, + 0.256452739238739, + 0.0113154835999012, + 0.003988961223512888, + 0.004382647108286619, + 0.004401359241455793, + 0.005577049218118191, + 0.03322863578796387, + 0.003944626543670893, + 0.02860957756638527, + 0.0007760969456285238, + 0.0006427625194191933, + 0.0008621008601039648, + 0.0006148996180854738, + 0.0005402666865848005, + 0.0013656070223078132, + 0.00047783690388314426, + 0.0013202342670410872, + 0.0016261866549029946, + 0.0006741721881553531, + 0.001300695352256298, + 0.0010328638600185513, + 0.0007353202090598643, + 0.004939489532262087, + 0.0005536887911148369, + 0.004280378110706806, + 0.0004948907298967242, + 0.0003572246932890266, + 0.0006881364970467985, + 0.000394239992601797, + 0.00045216712169349194, + 0.0011663157492876053, + 0.0004026723909191787, + 0.0010491417488083243, + 0.000482771749375388, + 0.00043264817213639617, + 0.0005513968062587082, + 0.0007384228520095348, + 0.0003557107411324978, + 0.0005642691976390779, + 0.00032014204771257937, + 0.0004250785568729043, + 0.0005754905287176371, + 0.0008481081458739936, + 0.0006762967095710337, + 0.0008450354216620326, + 0.0004964342224411666, + 0.0006834579980932176, + 0.0005116909742355347, + 0.0005487965536303818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.016065552830696106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305175319314003 + ], + "opposite_logprob_tokens": [ + -3.1680517196655273 + ], + "logprob_reward": -0.04305175319314003, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305175319314003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022804295644164085, + 0.021075690165162086, + 0.013093899004161358, + 0.7766326665878296, + 0.0705803707242012, + 0.00966070406138897, + 0.04315304011106491, + 0.024034613743424416, + 0.006327824201434851, + 0.012636993080377579 + ], + "credit_map": [ + 0.0026340025942772627, + 0.0021289647556841373, + 0.003436796134337783, + 0.002983557526022196, + 0.002116044284775853, + 0.004284857772290707, + 0.0021322970278561115, + 0.003087771125137806, + 0.003143016714602709, + 0.0014328699326142669, + 0.004860995337367058, + 0.0018375281943008304, + 0.0021518580615520477, + 0.0036995571572333574, + 0.0012900838628411293, + 0.002659775782376528, + 0.0013734146486967802, + 0.001057616202160716, + 0.0009523362386971712, + 0.0014868354192003608, + 0.0008975269156508148, + 0.003444720758125186, + 0.0007304079481400549, + 0.003151038195937872, + 0.08871389180421829, + 0.03004227578639984, + 0.03443475067615509, + 0.03614164516329765, + 0.04345904290676117, + 0.26580503582954407, + 0.029628602787852287, + 0.24840734899044037, + 0.008357038721442223, + 0.0027022366411983967, + 0.003032311797142029, + 0.0031281099654734135, + 0.00423514423891902, + 0.02475721761584282, + 0.0027498367708176374, + 0.021618464961647987, + 0.0010146336862817407, + 0.0011765483068302274, + 0.0013289368944242597, + 0.001972307451069355, + 0.0007197872619144619, + 0.001358939683996141, + 0.000922482751775533, + 0.0011670660460367799, + 0.004915921948850155, + 0.002633230295032263, + 0.003115891944617033, + 0.0030854984652251005, + 0.0023952850606292486, + 0.013461781665682793, + 0.0016529260901734233, + 0.011892497539520264, + 0.002023145789280534, + 0.0015184914227575064, + 0.0021462109871208668, + 0.0026968929450958967, + 0.0021065848413854837, + 0.005873252637684345, + 0.002587664872407913, + 0.005082367919385433, + 0.0008001623791642487, + 0.0006024350295774639, + 0.0006356642697937787, + 0.0007090019062161446, + 0.0008950546034611762, + 0.0008189200307242572, + 0.0012057940475642681, + 0.0006607905961573124, + 0.0017826847033575177, + 0.0014727524248883128, + 0.0014291690895333886, + 0.002083706669509411, + 0.0011042788391932845, + 0.0017929235473275185, + 0.0011761531932279468, + 0.0017953215865418315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.04305175319314003 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed low-angle medium shot captures a massive giant panda, its black and white fur languidly curled up deep in a bamboo forest. Its plump body rests gently on a thick, emerald green bamboo pole, its face a look of contentment and tranquility. Warm golden sunlight filters through the dense bamboo foliage overhead, casting countless swaying, mottled specks of light on its fluffy back and the surrounding leaf-strewn ground. Each beam is distinct, and tiny particles of light seem to drift slowly in the air. Surrounding it is an endless, towering sea of bamboo, its leaves a rich, emerald green, their veins each shimmering in the backlight. The entire scene is imbued with Makoto Shinkai's signature delicate textures and beautiful lighting, with warm, bright tones and sharp contrasts, creating a soothing atmosphere of profound tranquility, a feeling of almost frozen time.", + "seed": 11882, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_5/index_5398_seed_11882_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": -0.023440313851460816, + "vlm_logprob": -0.023440313851460816, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.023440313851460816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014057076536118984 + ], + "opposite_logprob_tokens": [ + -4.389057159423828 + ], + "logprob_reward": -0.014057076536118984, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014057076536118984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12101392447948456, + 0.172477588057518, + 0.031051838770508766, + 0.12897256016731262, + 0.030085016041994095, + 0.06901516020298004, + 0.19657529890537262, + 0.1332724243402481, + 0.02601219341158867, + 0.09152399003505707 + ], + "credit_map": [ + 0.010672803968191147, + 0.01157698780298233, + 0.02054397389292717, + 0.017386090010404587, + 0.013780349865555763, + 0.012758486904203892, + 0.014027392491698265, + 0.020267844200134277, + 0.02472657896578312, + 0.020358512178063393, + 0.04571656882762909, + 0.024902811273932457, + 0.01466172281652689, + 0.011766202747821808, + 0.01505069900304079, + 0.015294481068849564, + 0.004439531825482845, + 0.0037121560890227556, + 0.00826091319322586, + 0.003431357443332672, + 0.0024804992135614157, + 0.002475413493812084, + 0.0031291127670556307, + 0.00312285334803164, + 0.017987528815865517, + 0.01163734681904316, + 0.03342030569911003, + 0.0152969965711236, + 0.011805810034275055, + 0.015991918742656708, + 0.010792278684675694, + 0.012040365487337112, + 0.002783368341624737, + 0.003971965983510017, + 0.003629183629527688, + 0.002619830658659339, + 0.004506227560341358, + 0.0051395585760474205, + 0.0026544544380158186, + 0.004780425690114498, + 0.0057264030911028385, + 0.006320780608803034, + 0.008711153641343117, + 0.00854166317731142, + 0.014184967614710331, + 0.009514020755887032, + 0.007664523087441921, + 0.008351649157702923, + 0.01582339033484459, + 0.013218269683420658, + 0.03820459544658661, + 0.01974467933177948, + 0.018337897956371307, + 0.03311445191502571, + 0.025304920971393585, + 0.03282708674669266, + 0.007555647287517786, + 0.010513827204704285, + 0.018261225894093513, + 0.021840576082468033, + 0.016114648431539536, + 0.017180180177092552, + 0.01841902546584606, + 0.023387284949421883, + 0.0027980818413197994, + 0.0022468059323728085, + 0.004462269600480795, + 0.004587260074913502, + 0.0028014532290399075, + 0.0025020635221153498, + 0.002244645496830344, + 0.0043696132488548756, + 0.010785635560750961, + 0.008805869147181511, + 0.012412037700414658, + 0.01783614046871662, + 0.011265960521996021, + 0.009477175772190094, + 0.009753606282174587, + 0.011187571100890636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.014057076536118984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307607561349869 + ], + "opposite_logprob_tokens": [ + -3.1680760383605957 + ], + "logprob_reward": -0.04307607561349869, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307607561349869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11937123537063599, + 0.037985388189554214, + 0.3010941743850708, + 0.05859234556555748, + 0.06076351925730705, + 0.04366271570324898, + 0.1638234704732895, + 0.12899573147296906, + 0.0728849247097969, + 0.012826538644731045 + ], + "credit_map": [ + 0.01764068938791752, + 0.01000439003109932, + 0.027756310999393463, + 0.018113765865564346, + 0.01277216337621212, + 0.01139272004365921, + 0.009802358224987984, + 0.011888820677995682, + 0.0036641329061239958, + 0.00313218729570508, + 0.006971211638301611, + 0.004995399154722691, + 0.004925751592963934, + 0.005235581658780575, + 0.0035500547382980585, + 0.005511065945029259, + 0.04822024330496788, + 0.032922256737947464, + 0.0920267328619957, + 0.02982315421104431, + 0.03360181674361229, + 0.02105528488755226, + 0.028092358261346817, + 0.015352307818830013, + 0.008255908265709877, + 0.005630644503980875, + 0.017735766246914864, + 0.007307813968509436, + 0.005650420673191547, + 0.004782534204423428, + 0.005462439265102148, + 0.0037668144796043634, + 0.005975333973765373, + 0.007132742553949356, + 0.012815568596124649, + 0.008998790755867958, + 0.010324736125767231, + 0.00405358849093318, + 0.006401189602911472, + 0.005061561707407236, + 0.004035813268274069, + 0.004330864641815424, + 0.00534107955172658, + 0.005226536653935909, + 0.01070377230644226, + 0.004566091578453779, + 0.004510908387601376, + 0.004947645589709282, + 0.014093373902142048, + 0.010964002460241318, + 0.03329779952764511, + 0.01650230959057808, + 0.01625252142548561, + 0.02713775634765625, + 0.020583000034093857, + 0.024992691352963448, + 0.008188813924789429, + 0.010999542661011219, + 0.018914951011538506, + 0.02276243455708027, + 0.016171200200915337, + 0.015517395921051502, + 0.015216795727610588, + 0.02122459001839161, + 0.008634314872324467, + 0.007559587247669697, + 0.012691405601799488, + 0.012884832918643951, + 0.009833470918238163, + 0.0060668024234473705, + 0.006892642937600613, + 0.008321858942508698, + 0.0009594893781468272, + 0.0010491999564692378, + 0.0016853965353220701, + 0.001541081815958023, + 0.0018627466633915901, + 0.0012720043305307627, + 0.0019205929711461067, + 0.002536025596782565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.04307607561349869 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028183048125356436 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.0028183048125356436, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028183048125356436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3148285150527954, + 0.029813485220074654, + 0.0861557349562645, + 0.2086230367422104, + 0.1304008513689041, + 0.053310077637434006, + 0.06570763140916824, + 0.034406937658786774, + 0.016747379675507545, + 0.060006361454725266 + ], + "credit_map": [ + 0.05466984584927559, + 0.03556809574365616, + 0.09592128545045853, + 0.0284828320145607, + 0.031331900507211685, + 0.026725642383098602, + 0.026811867952346802, + 0.015317026525735855, + 0.0032952085603028536, + 0.0029595806263387203, + 0.0044393641874194145, + 0.0040132817812263966, + 0.006024589296430349, + 0.0028792291413992643, + 0.0033757248893380165, + 0.0028265088330954313, + 0.013595063239336014, + 0.009421267546713352, + 0.02618735283613205, + 0.008265812881290913, + 0.009436681866645813, + 0.006647059693932533, + 0.008344156667590141, + 0.004258340690284967, + 0.03545260801911354, + 0.023045135661959648, + 0.061392638832330704, + 0.01804070547223091, + 0.024577250704169273, + 0.017606277018785477, + 0.017200350761413574, + 0.011308060958981514, + 0.01610735058784485, + 0.019711561501026154, + 0.03111838735640049, + 0.015696799382567406, + 0.014378970488905907, + 0.010963956825435162, + 0.012220929376780987, + 0.010202895849943161, + 0.006850517820566893, + 0.007231058087199926, + 0.01157291978597641, + 0.00693640997633338, + 0.007544482592493296, + 0.004066058434545994, + 0.005348371807485819, + 0.0037602586671710014, + 0.00510094640776515, + 0.0053831725381314754, + 0.01129883248358965, + 0.008869487792253494, + 0.013525380752980709, + 0.007082473021000624, + 0.007334345951676369, + 0.007112993858754635, + 0.0033137400168925524, + 0.0027436392847448587, + 0.00489850016310811, + 0.00502992607653141, + 0.0045487359166145325, + 0.0042963409796357155, + 0.0037278051022440195, + 0.005848251283168793, + 0.002087887143716216, + 0.0019387624925002456, + 0.0030808618757873774, + 0.0026153011713176966, + 0.0025060202460736036, + 0.0012890679063275456, + 0.0014930744655430317, + 0.0017364034429192543, + 0.006373818498104811, + 0.005394316278398037, + 0.009775172919034958, + 0.010159785859286785, + 0.011113669723272324, + 0.0057601225562393665, + 0.005082443822175264, + 0.006347026210278273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0028183048125356436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419447734951973 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.01419447734951973, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419447734951973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19744454324245453, + 0.24072177708148956, + 0.0888669341802597, + 0.12157806009054184, + 0.047475919127464294, + 0.03173794597387314, + 0.06028405576944351, + 0.06108603999018669, + 0.06474735587835312, + 0.08605732768774033 + ], + "credit_map": [ + 0.032487452030181885, + 0.016116701066493988, + 0.045301128178834915, + 0.02511650137603283, + 0.028586450964212418, + 0.020010383799672127, + 0.0174034982919693, + 0.012422429397702217, + 0.027919696643948555, + 0.02679186686873436, + 0.032978251576423645, + 0.028192713856697083, + 0.054058726876974106, + 0.026808999478816986, + 0.022224005311727524, + 0.021747522056102753, + 0.012617756612598896, + 0.007671508472412825, + 0.01885942928493023, + 0.008923499844968319, + 0.01619681715965271, + 0.010138105601072311, + 0.0089877275750041, + 0.0054720803163945675, + 0.01935506984591484, + 0.012529509142041206, + 0.03574305772781372, + 0.011586541309952736, + 0.01164996437728405, + 0.01230000052601099, + 0.010057706385850906, + 0.008356215432286263, + 0.00564282201230526, + 0.007544415071606636, + 0.011045331135392189, + 0.006014187820255756, + 0.005398543551564217, + 0.003760907333344221, + 0.004363563377410173, + 0.0037061506882309914, + 0.0028385582845658064, + 0.0035132032353430986, + 0.0035227073822170496, + 0.003526123007759452, + 0.004354133736342192, + 0.00589342787861824, + 0.0032630162313580513, + 0.004826778545975685, + 0.003772200085222721, + 0.005277460906654596, + 0.009994630701839924, + 0.007650245912373066, + 0.009166594594717026, + 0.008268428035080433, + 0.007910914719104767, + 0.008243583142757416, + 0.005189431365579367, + 0.004770321771502495, + 0.01449313759803772, + 0.0082176448777318, + 0.006621789187192917, + 0.006675923243165016, + 0.005481326952576637, + 0.009636464528739452, + 0.009379444643855095, + 0.005706741474568844, + 0.01174632553011179, + 0.006575754843652248, + 0.011013006791472435, + 0.006984726991504431, + 0.006606223061680794, + 0.006735129747539759, + 0.012481479905545712, + 0.008000969886779785, + 0.01468612439930439, + 0.01225267443805933, + 0.017062408849596977, + 0.007844001054763794, + 0.006527406629174948, + 0.007202268112450838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01419447734951973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305563494563103 + ], + "opposite_logprob_tokens": [ + -3.168055534362793 + ], + "logprob_reward": -0.04305563494563103, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305563494563103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12444347888231277, + 0.37936991453170776, + 0.09510110318660736, + 0.04121502488851547, + 0.04962489381432533, + 0.031933967024087906, + 0.07177277654409409, + 0.09729375690221786, + 0.0748102143406868, + 0.034434836357831955 + ], + "credit_map": [ + 0.013975978828966618, + 0.012629237957298756, + 0.02595341205596924, + 0.015667660161852837, + 0.01628032512962818, + 0.009599296376109123, + 0.012082422152161598, + 0.018255146220326424, + 0.06322599202394485, + 0.043624166399240494, + 0.11771772801876068, + 0.041854023933410645, + 0.036310143768787384, + 0.025124117732048035, + 0.03177400305867195, + 0.01973976381123066, + 0.014695514924824238, + 0.01035025529563427, + 0.024825988337397575, + 0.007943861186504364, + 0.014213154092431068, + 0.006617128849029541, + 0.010039351880550385, + 0.006415841169655323, + 0.005014521535485983, + 0.00569488201290369, + 0.008795630186796188, + 0.004429729655385017, + 0.006114000454545021, + 0.004421053919941187, + 0.003423861926421523, + 0.003321343334391713, + 0.006285284645855427, + 0.007111482322216034, + 0.012113435193896294, + 0.005112725310027599, + 0.005237503908574581, + 0.004345903638750315, + 0.004759197123348713, + 0.0046593621373176575, + 0.0032566969748586416, + 0.004765319172292948, + 0.00380546017549932, + 0.0027452043723315, + 0.004931625910103321, + 0.00504827918484807, + 0.0029799870681017637, + 0.004401391837745905, + 0.011021224781870842, + 0.005755242425948381, + 0.01821255311369896, + 0.005726050119847059, + 0.011135566048324108, + 0.007524725049734116, + 0.007150127552449703, + 0.005247280467301607, + 0.008774793706834316, + 0.007239788305014372, + 0.01932670548558235, + 0.010403197258710861, + 0.008983563631772995, + 0.01554525550454855, + 0.0100226029753685, + 0.01699785329401493, + 0.010354381054639816, + 0.0056747253984212875, + 0.013060718774795532, + 0.008058705367147923, + 0.00894557498395443, + 0.009632911533117294, + 0.008538582362234592, + 0.010544616729021072, + 0.0049269115552306175, + 0.003548347158357501, + 0.006620823405683041, + 0.004084961488842964, + 0.006851378828287125, + 0.0027009265031665564, + 0.003190785413607955, + 0.002510703168809414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.04305563494563103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed low-angle medium shot captures a massive giant panda, its black and white fur languidly curled up deep in a bamboo forest. Its plump body rests gently on a thick, emerald green bamboo pole, its face a look of contentment and tranquility. Warm golden sunlight filters through the dense bamboo foliage overhead, casting countless swaying, mottled specks of light on its fluffy back and the surrounding leaf-strewn ground. Each beam is distinct, and tiny particles of light seem to drift slowly in the air. Surrounding it is an endless, towering sea of bamboo, its leaves a rich, emerald green, their veins each shimmering in the backlight. The entire scene is imbued with Makoto Shinkai's signature delicate textures and beautiful lighting, with warm, bright tones and sharp contrasts, creating a soothing atmosphere of profound tranquility, a feeling of almost frozen time.", + "seed": 11883, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_5/index_5398_seed_11883_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": -0.020208601374179125, + "vlm_logprob": -0.020208601374179125, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.020208601374179125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006592904217541218 + ], + "opposite_logprob_tokens": [ + -5.256592750549316 + ], + "logprob_reward": -0.006592904217541218, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006592904217541218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14403589069843292, + 0.16614407300949097, + 0.06619827449321747, + 0.08585312962532043, + 0.1422947496175766, + 0.05164821818470955, + 0.08607030659914017, + 0.0825420692563057, + 0.11412717401981354, + 0.061086125671863556 + ], + "credit_map": [ + 0.01889154128730297, + 0.022282030433416367, + 0.021221943199634552, + 0.01966543309390545, + 0.013067835941910744, + 0.010750135406851768, + 0.01881178840994835, + 0.019345181062817574, + 0.025038352236151695, + 0.028331421315670013, + 0.022029060870409012, + 0.01825571618974209, + 0.012744329869747162, + 0.013083512894809246, + 0.02563490718603134, + 0.02102677896618843, + 0.013135499320924282, + 0.010911647230386734, + 0.011616174131631851, + 0.005008946172893047, + 0.0053848568350076675, + 0.004878130741417408, + 0.0066379462368786335, + 0.00862507801502943, + 0.013609558343887329, + 0.009409288875758648, + 0.015417895279824734, + 0.007591355592012405, + 0.0069525656290352345, + 0.005411974620074034, + 0.01319102756679058, + 0.01426946371793747, + 0.023984132334589958, + 0.024140244349837303, + 0.017669692635536194, + 0.011799413710832596, + 0.013422679156064987, + 0.011757663451135159, + 0.0195113867521286, + 0.020009543746709824, + 0.006220205686986446, + 0.006399589125066996, + 0.007943812757730484, + 0.004059172235429287, + 0.004705190658569336, + 0.004115351475775242, + 0.008918304927647114, + 0.009286593645811081, + 0.016360443085432053, + 0.011955522000789642, + 0.017813028767704964, + 0.009929296560585499, + 0.00741954380646348, + 0.009472954086959362, + 0.005717355292290449, + 0.007402157876640558, + 0.013312521390616894, + 0.008242101408541203, + 0.018928363919258118, + 0.006500252056866884, + 0.007157996296882629, + 0.005751547869294882, + 0.01134380605071783, + 0.011305486783385277, + 0.02077743597328663, + 0.01302506122738123, + 0.014295884408056736, + 0.011903641745448112, + 0.012109026312828064, + 0.013168146833777428, + 0.01687949150800705, + 0.011968488804996014, + 0.010057755745947361, + 0.008818327449262142, + 0.010707343928515911, + 0.006341736763715744, + 0.004224803764373064, + 0.0031433552503585815, + 0.008107425644993782, + 0.009685375727713108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.006592904217541218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208580367267132 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208580367267132, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208580367267132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12130258232355118, + 0.0914551168680191, + 0.06712809950113297, + 0.23024430871009827, + 0.08631110936403275, + 0.08997707813978195, + 0.06885235011577606, + 0.11915337294340134, + 0.07383474707603455, + 0.051741208881139755 + ], + "credit_map": [ + 0.02088611200451851, + 0.01931741088628769, + 0.013951932080090046, + 0.009882904589176178, + 0.00972730852663517, + 0.01126060914248228, + 0.015687985345721245, + 0.02058832347393036, + 0.012735437601804733, + 0.016732219606637955, + 0.013998828828334808, + 0.009670370258390903, + 0.00930116232484579, + 0.0061947559006512165, + 0.01068748626857996, + 0.012134862132370472, + 0.008253408595919609, + 0.007312879431992769, + 0.009664571844041348, + 0.004691858775913715, + 0.007434191182255745, + 0.006523392628878355, + 0.013166280463337898, + 0.010081520304083824, + 0.02792254649102688, + 0.016047686338424683, + 0.04385709390044212, + 0.020713679492473602, + 0.0200954582542181, + 0.013467755168676376, + 0.04818160831928253, + 0.039958491921424866, + 0.013875295408070087, + 0.01285884715616703, + 0.012734946794807911, + 0.00820669811218977, + 0.007073978893458843, + 0.007211665622889996, + 0.013378974050283432, + 0.01097070425748825, + 0.01336334552615881, + 0.011908210813999176, + 0.011966314166784286, + 0.007629825733602047, + 0.007690689526498318, + 0.007372994441539049, + 0.013556722551584244, + 0.016488976776599884, + 0.010080202482640743, + 0.010037362575531006, + 0.007944728247821331, + 0.006954483222216368, + 0.006237931549549103, + 0.006148322951048613, + 0.010185284540057182, + 0.011264034546911716, + 0.017926020547747612, + 0.013605087995529175, + 0.023797880858182907, + 0.009701482951641083, + 0.007890842854976654, + 0.007478258106857538, + 0.01714632846415043, + 0.021607471629977226, + 0.014777977019548416, + 0.00994789507240057, + 0.01143032405525446, + 0.007874464616179466, + 0.007352547720074654, + 0.0070192692801356316, + 0.008300531655550003, + 0.007131739519536495, + 0.008765255101025105, + 0.005685950163751841, + 0.006643745116889477, + 0.004594290629029274, + 0.006127302069216967, + 0.005126671399921179, + 0.007194955833256245, + 0.007603038102388382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.014208580367267132 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252846051007509 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252846051007509, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252846051007509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25531724095344543, + 0.10082679986953735, + 0.11590409278869629, + 0.09841948747634888, + 0.09517492353916168, + 0.045767009258270264, + 0.07511519640684128, + 0.0481719896197319, + 0.14853325486183167, + 0.01677001640200615 + ], + "credit_map": [ + 0.03849653899669647, + 0.058504194021224976, + 0.034764889627695084, + 0.02081194333732128, + 0.019554996863007545, + 0.018226148560643196, + 0.0335693433880806, + 0.03138918802142143, + 0.014805723913013935, + 0.016228890046477318, + 0.014601130969822407, + 0.009314779192209244, + 0.007268641144037247, + 0.006586981005966663, + 0.01223826315253973, + 0.01978239230811596, + 0.016785848885774612, + 0.017041753977537155, + 0.018862519413232803, + 0.013726497069001198, + 0.0071718343533575535, + 0.008228834718465805, + 0.011904806829988956, + 0.022181998938322067, + 0.013708257116377354, + 0.009389207698404789, + 0.021119235083460808, + 0.008135550655424595, + 0.00815768912434578, + 0.008716562762856483, + 0.009018231183290482, + 0.02017475850880146, + 0.01572812721133232, + 0.020352758467197418, + 0.014569501392543316, + 0.008916654624044895, + 0.007880933582782745, + 0.005963339004665613, + 0.01063009537756443, + 0.011133512482047081, + 0.007229255978018045, + 0.0071991910226643085, + 0.008040591143071651, + 0.0053658452816307545, + 0.004402742255479097, + 0.0031985181849449873, + 0.004720547702163458, + 0.0056103188544511795, + 0.01531166210770607, + 0.013514753431081772, + 0.015579977072775364, + 0.007983839139342308, + 0.00550176901742816, + 0.004912875592708588, + 0.005951079539954662, + 0.0063592419028282166, + 0.009053898975253105, + 0.0077067408710718155, + 0.008933079428970814, + 0.005041979253292084, + 0.004844352602958679, + 0.003860844997689128, + 0.004243867006152868, + 0.004487224854528904, + 0.03334052488207817, + 0.016929389908909798, + 0.028507398441433907, + 0.017290811985731125, + 0.011878082528710365, + 0.013804534450173378, + 0.012621565721929073, + 0.014160951599478722, + 0.0028986327815800905, + 0.0014161474537104368, + 0.0039412882179021835, + 0.0016552044544368982, + 0.0019503221847116947, + 0.0017326726811006665, + 0.0014483784325420856, + 0.0017273692646995187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.005252846051007509 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634468488395214 + ], + "opposite_logprob_tokens": [ + -3.6513447761535645 + ], + "logprob_reward": -0.02634468488395214, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634468488395214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2001390904188156, + 0.07829728722572327, + 0.03225698322057724, + 0.10719523578882217, + 0.09546960145235062, + 0.1354585587978363, + 0.13226623833179474, + 0.05773551017045975, + 0.12853559851646423, + 0.032645903527736664 + ], + "credit_map": [ + 0.03733382746577263, + 0.037958912551403046, + 0.02820136770606041, + 0.021583156660199165, + 0.014329875819385052, + 0.016333336010575294, + 0.01802930422127247, + 0.02636929601430893, + 0.013676663860678673, + 0.008344745263457298, + 0.013946756720542908, + 0.007193923462182283, + 0.0072264596819877625, + 0.005424239207059145, + 0.010869157500565052, + 0.011615338735282421, + 0.005598553456366062, + 0.0035347819793969393, + 0.005208821501582861, + 0.003460121341049671, + 0.003217377234250307, + 0.0024979908484965563, + 0.003369190264493227, + 0.005370147991925478, + 0.016869574785232544, + 0.009850084781646729, + 0.017797809094190598, + 0.011941663920879364, + 0.009068483486771584, + 0.009663631208240986, + 0.01272914744913578, + 0.019274838268756866, + 0.01713828556239605, + 0.018496261909604073, + 0.017231546342372894, + 0.009169413708150387, + 0.007544237654656172, + 0.006554981227964163, + 0.009296689182519913, + 0.01003817655146122, + 0.015119930729269981, + 0.013885614462196827, + 0.01545287761837244, + 0.013418258167803288, + 0.014398767612874508, + 0.016712091863155365, + 0.02346726879477501, + 0.023003745824098587, + 0.01686125621199608, + 0.015359693206846714, + 0.019575346261262894, + 0.011771838180720806, + 0.012995854951441288, + 0.013379591517150402, + 0.02044086903333664, + 0.021881790831685066, + 0.010343736037611961, + 0.005782299675047398, + 0.01010194607079029, + 0.005194452591240406, + 0.005789289250969887, + 0.005349172279238701, + 0.008898419328033924, + 0.006276192143559456, + 0.02665296383202076, + 0.014601205475628376, + 0.020940084010362625, + 0.014564036391675472, + 0.011475379578769207, + 0.01235734298825264, + 0.01437459234148264, + 0.013569984585046768, + 0.006510031875222921, + 0.004335117060691118, + 0.004677996039390564, + 0.0034885616041719913, + 0.0030416445806622505, + 0.0029100999236106873, + 0.003364180214703083, + 0.004318274091929197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.02634468488395214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048643991351127625 + ], + "opposite_logprob_tokens": [ + -3.0486440658569336 + ], + "logprob_reward": -0.048643991351127625, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048643991351127625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27723559737205505, + 0.09291305392980576, + 0.045511092990636826, + 0.13946911692619324, + 0.07324416190385818, + 0.0733473002910614, + 0.0842195600271225, + 0.08763826638460159, + 0.09886124730110168, + 0.027560584247112274 + ], + "credit_map": [ + 0.04833335429430008, + 0.04844703897833824, + 0.04014626145362854, + 0.021911097690463066, + 0.022032955661416054, + 0.017408210784196854, + 0.028149602934718132, + 0.05080708488821983, + 0.016758309677243233, + 0.019695620983839035, + 0.011952295899391174, + 0.010800770483911037, + 0.00856688804924488, + 0.006357596721500158, + 0.007486728951334953, + 0.01129484549164772, + 0.006847900804132223, + 0.007866477593779564, + 0.005891998298466206, + 0.003845797386020422, + 0.005488446913659573, + 0.004191720858216286, + 0.005796908866614103, + 0.005581845063716173, + 0.024037400260567665, + 0.026283133774995804, + 0.01684078387916088, + 0.011229097843170166, + 0.011987483128905296, + 0.009579799138009548, + 0.014635642990469933, + 0.024875786155462265, + 0.010530379600822926, + 0.009363295510411263, + 0.010479286313056946, + 0.0070724328979849815, + 0.005990583449602127, + 0.006154597736895084, + 0.012917960993945599, + 0.010735632851719856, + 0.010509529151022434, + 0.00757775129750371, + 0.00946235191076994, + 0.009312384761869907, + 0.008193297311663628, + 0.00718297902494669, + 0.010746343992650509, + 0.010362666100263596, + 0.009015842340886593, + 0.007259662263095379, + 0.015160790644586086, + 0.009606425650417805, + 0.007160719018429518, + 0.008064132183790207, + 0.014859144575893879, + 0.013092848472297192, + 0.011491532437503338, + 0.012357525527477264, + 0.009909573011100292, + 0.00743896747007966, + 0.009437321685254574, + 0.010770651511847973, + 0.014381187967956066, + 0.011851509101688862, + 0.022132612764835358, + 0.015007361769676208, + 0.02006996050477028, + 0.011788474395871162, + 0.007689210120588541, + 0.006120022851973772, + 0.007153909653425217, + 0.008899702690541744, + 0.004813010338693857, + 0.0045080422423779964, + 0.004148958250880241, + 0.0025856990832835436, + 0.003559770528227091, + 0.0027715687174350023, + 0.0024242503568530083, + 0.002749286126345396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.048643991351127625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed low-angle medium shot captures a massive giant panda, its black and white fur languidly curled up deep in a bamboo forest. Its plump body rests gently on a thick, emerald green bamboo pole, its face a look of contentment and tranquility. Warm golden sunlight filters through the dense bamboo foliage overhead, casting countless swaying, mottled specks of light on its fluffy back and the surrounding leaf-strewn ground. Each beam is distinct, and tiny particles of light seem to drift slowly in the air. Surrounding it is an endless, towering sea of bamboo, its leaves a rich, emerald green, their veins each shimmering in the backlight. The entire scene is imbued with Makoto Shinkai's signature delicate textures and beautiful lighting, with warm, bright tones and sharp contrasts, creating a soothing atmosphere of profound tranquility, a feeling of almost frozen time.", + "seed": 11884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_5/index_5398_seed_11884_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": -0.016255797538906335, + "vlm_logprob": -0.016255797538906335, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.016255797538906335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025488125160336494 + ], + "opposite_logprob_tokens": [ + -3.7754881381988525 + ], + "logprob_reward": -0.025488125160336494, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025488125160336494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10357411950826645, + 0.12873150408267975, + 0.026215801015496254, + 0.02777680940926075, + 0.032006021589040756, + 0.3422439694404602, + 0.04370678961277008, + 0.09927146136760712, + 0.08583942800760269, + 0.11063411831855774 + ], + "credit_map": [ + 0.015844903886318207, + 0.009410090744495392, + 0.01033829990774393, + 0.016177518293261528, + 0.010417832992970943, + 0.017195941880345345, + 0.013227047398686409, + 0.010962482541799545, + 0.026206713169813156, + 0.011647062376141548, + 0.011549298651516438, + 0.009009405970573425, + 0.010904100723564625, + 0.039731696248054504, + 0.0077780624851584435, + 0.011905163526535034, + 0.002204902470111847, + 0.0031092248391360044, + 0.0023614794481545687, + 0.0032097394578158855, + 0.0033739348873496056, + 0.0049315765500068665, + 0.003697877051308751, + 0.0033270663116127253, + 0.0034688713494688272, + 0.004365155939012766, + 0.0018720513908192515, + 0.001958368578925729, + 0.00403986731544137, + 0.0039944127202034, + 0.0022726149763911963, + 0.005805468186736107, + 0.0030099437572062016, + 0.0035719608422368765, + 0.002557369414716959, + 0.00232725334353745, + 0.0036196683067828417, + 0.006211015861481428, + 0.002683482598513365, + 0.008025324903428555, + 0.09495683014392853, + 0.01775156520307064, + 0.015782639384269714, + 0.01652485318481922, + 0.018149057403206825, + 0.03284559026360512, + 0.0162186361849308, + 0.13001477718353271, + 0.003790120128542185, + 0.0037777493707835674, + 0.004264752380549908, + 0.005636232905089855, + 0.006135714240372181, + 0.00993574969470501, + 0.003764691762626171, + 0.006401776801794767, + 0.027883999049663544, + 0.006155502051115036, + 0.0050806691870093346, + 0.003736612619832158, + 0.007788869086652994, + 0.009535257704555988, + 0.004315581172704697, + 0.03477497771382332, + 0.007873251102864742, + 0.007697427179664373, + 0.008123531006276608, + 0.011945479549467564, + 0.01235316414386034, + 0.01727304607629776, + 0.010142444632947445, + 0.010431083850562572, + 0.03018081933259964, + 0.004977551754564047, + 0.004768665414303541, + 0.004311034455895424, + 0.042139261960983276, + 0.011839310638606548, + 0.004910656716674566, + 0.007506819907575846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.025488125160336494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610320806503296 + ], + "opposite_logprob_tokens": [ + -4.141103267669678 + ], + "logprob_reward": -0.01610320806503296, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610320806503296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08334106206893921, + 0.16426505148410797, + 0.036751583218574524, + 0.11133330315351486, + 0.05082196369767189, + 0.3006558120250702, + 0.07022462040185928, + 0.07766609638929367, + 0.029533356428146362, + 0.07540715485811234 + ], + "credit_map": [ + 0.011163875460624695, + 0.00746058439835906, + 0.008784766308963299, + 0.01171299908310175, + 0.008852987550199032, + 0.014264757744967937, + 0.011941364035010338, + 0.00915972888469696, + 0.03671305626630783, + 0.013323378749191761, + 0.015549900010228157, + 0.01316814124584198, + 0.013482936657965183, + 0.049228768795728683, + 0.009621985256671906, + 0.013176888227462769, + 0.006252506747841835, + 0.002911800052970648, + 0.0028523635119199753, + 0.002970250090584159, + 0.003645415650680661, + 0.01058207917958498, + 0.003112919395789504, + 0.004424250219017267, + 0.02181536704301834, + 0.013168087229132652, + 0.009411226026713848, + 0.006328899413347244, + 0.013022023253142834, + 0.012794203124940395, + 0.008179659023880959, + 0.026613842695951462, + 0.006910399068146944, + 0.003495490178465843, + 0.004024203401058912, + 0.004746771417558193, + 0.004817524924874306, + 0.00998716801404953, + 0.004334981087595224, + 0.012505428865551949, + 0.09142307937145233, + 0.01607341319322586, + 0.010449803434312344, + 0.008400027640163898, + 0.016432443633675575, + 0.026687676087021828, + 0.01270509697496891, + 0.1184842512011528, + 0.008454806171357632, + 0.00587113993242383, + 0.007196149323135614, + 0.008004270493984222, + 0.007696222048252821, + 0.011100990697741508, + 0.007751055061817169, + 0.014149988070130348, + 0.01907106302678585, + 0.004891338292509317, + 0.004119970835745335, + 0.004142815247178078, + 0.007129382807761431, + 0.0098545141518116, + 0.004903648514300585, + 0.02355336584150791, + 0.0033216977026313543, + 0.002540381159633398, + 0.0022340144496411085, + 0.0029674263205379248, + 0.003951932769268751, + 0.005796494428068399, + 0.0025304697919636965, + 0.0061909412033855915, + 0.004488769918680191, + 0.003437983337789774, + 0.0026427549310028553, + 0.0037786730099469423, + 0.025006698444485664, + 0.007286703679710627, + 0.0036057056859135628, + 0.025159871205687523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.01610320806503296 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638507030904293 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638507030904293, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638507030904293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12962983548641205, + 0.15837779641151428, + 0.061253923922777176, + 0.10860875993967056, + 0.06337722390890121, + 0.06244932860136032, + 0.12022525072097778, + 0.1537412852048874, + 0.05449951812624931, + 0.08783700317144394 + ], + "credit_map": [ + 0.024448275566101074, + 0.0135427201166749, + 0.011528910137712955, + 0.011919002048671246, + 0.015206157229840755, + 0.029855934903025627, + 0.011004799976944923, + 0.012124046683311462, + 0.033271655440330505, + 0.013839107006788254, + 0.014347802847623825, + 0.009138115681707859, + 0.012115148827433586, + 0.05025859922170639, + 0.008724414743483067, + 0.01668296568095684, + 0.011497993022203445, + 0.00820944458246231, + 0.005400935187935829, + 0.004782853648066521, + 0.006683554034680128, + 0.012942537665367126, + 0.004690343979746103, + 0.007046269718557596, + 0.009064543060958385, + 0.01678076572716236, + 0.009650594554841518, + 0.011588448658585548, + 0.016661137342453003, + 0.020769571885466576, + 0.010874390602111816, + 0.01321931928396225, + 0.0138516491279006, + 0.0034755233209580183, + 0.0037423146422952414, + 0.003872640198096633, + 0.0037641427479684353, + 0.0085970563814044, + 0.003237171098589897, + 0.022836729884147644, + 0.01829293929040432, + 0.0034773556981235743, + 0.0027935123071074486, + 0.002327587688341737, + 0.0035817979369312525, + 0.005860455799847841, + 0.0030858847312629223, + 0.02302980050444603, + 0.01666174829006195, + 0.007942034862935543, + 0.009442729875445366, + 0.011339387856423855, + 0.012438842095434666, + 0.024762598797678947, + 0.009373185224831104, + 0.028264736756682396, + 0.0384276881814003, + 0.010178612545132637, + 0.00804221909493208, + 0.00453192600980401, + 0.013819045387208462, + 0.022093702107667923, + 0.006413593888282776, + 0.050234515219926834, + 0.0051437863148748875, + 0.005335927940905094, + 0.00427298853173852, + 0.006588097661733627, + 0.00838007964193821, + 0.011457814835011959, + 0.004918329883366823, + 0.008402500301599503, + 0.027733318507671356, + 0.004200139082968235, + 0.003553651040419936, + 0.0023617821279913187, + 0.030128493905067444, + 0.00966272410005331, + 0.003910940606147051, + 0.006285957060754299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.004638507030904293 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255454685539007 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255454685539007, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255454685539007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0673401802778244, + 0.044316377490758896, + 0.029550231993198395, + 0.0861605554819107, + 0.02758990414440632, + 0.5283693671226501, + 0.026566511020064354, + 0.05443446338176727, + 0.06715092808008194, + 0.06852151453495026 + ], + "credit_map": [ + 0.008257239125669003, + 0.00770157715305686, + 0.008060131222009659, + 0.0065224068239331245, + 0.007672606501728296, + 0.010547516867518425, + 0.00863669253885746, + 0.009942002594470978, + 0.008647728711366653, + 0.0043655876070261, + 0.004155915696173906, + 0.003441141452640295, + 0.004971732385456562, + 0.010496309027075768, + 0.0034093854483217, + 0.004828571807593107, + 0.004858675412833691, + 0.0041375188156962395, + 0.0027265306562185287, + 0.0021369517780840397, + 0.003873585257679224, + 0.0046228100545704365, + 0.002671042690053582, + 0.00452311709523201, + 0.021534500643610954, + 0.004082036670297384, + 0.00596628850325942, + 0.004623400047421455, + 0.004102910403162241, + 0.010082933120429516, + 0.00408446928486228, + 0.03168400749564171, + 0.004838621709495783, + 0.0022942679934203625, + 0.0030014330986887217, + 0.003497348865494132, + 0.0025155639741569757, + 0.0041613830253481865, + 0.0024094553664326668, + 0.00487182941287756, + 0.15719744563102722, + 0.027506832033395767, + 0.019929958507418633, + 0.016965026035904884, + 0.027641354128718376, + 0.04852352291345596, + 0.02286779135465622, + 0.20773741602897644, + 0.0031885390635579824, + 0.002241062233224511, + 0.0026587899774312973, + 0.0032211465295404196, + 0.003076012246310711, + 0.003967350348830223, + 0.002881299704313278, + 0.005332307890057564, + 0.013738416135311127, + 0.003946844022721052, + 0.002761083422228694, + 0.002436063950881362, + 0.005630570463836193, + 0.006167348008602858, + 0.003364869626238942, + 0.0163892675191164, + 0.006639546249061823, + 0.0060038804076612, + 0.0061872960068285465, + 0.009199012070894241, + 0.009923545643687248, + 0.011641471646726131, + 0.008267474360764027, + 0.009288699366152287, + 0.014402295462787151, + 0.0041161393746733665, + 0.004534503910690546, + 0.0070924279280006886, + 0.00595843605697155, + 0.008276699110865593, + 0.00549896527081728, + 0.018642041832208633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.005255454685539007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 giant panda with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda resting its plump body on a thick, emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dense bamboo forest with towering, emerald green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does warm golden sunlight filter through the foliage, casting distinct mottled specks of light on the panda and ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a fixed low-angle medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793692752718925 + ], + "opposite_logprob_tokens": [ + -3.5297937393188477 + ], + "logprob_reward": -0.029793692752718925, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793692752718925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11937741190195084, + 0.14277294278144836, + 0.09049217402935028, + 0.09395056962966919, + 0.09087584167718887, + 0.10083404183387756, + 0.09258329123258591, + 0.10927730053663254, + 0.04528798162937164, + 0.1145484521985054 + ], + "credit_map": [ + 0.012563066557049751, + 0.00918828509747982, + 0.010152137838304043, + 0.00872522871941328, + 0.012503995560109615, + 0.04334614798426628, + 0.009727558121085167, + 0.013170992955565453, + 0.027959344908595085, + 0.012325157411396503, + 0.012537851929664612, + 0.008676234632730484, + 0.011947249062359333, + 0.04577306658029556, + 0.008639601059257984, + 0.0149144371971488, + 0.007783113047480583, + 0.010383869521319866, + 0.006911125499755144, + 0.006087043322622776, + 0.011296980082988739, + 0.02670353278517723, + 0.0062522380612790585, + 0.01507426518946886, + 0.008802833966910839, + 0.011113367974758148, + 0.012236448004841805, + 0.013727608136832714, + 0.010593565180897713, + 0.012575341388583183, + 0.012507569044828415, + 0.012393830344080925, + 0.021866802126169205, + 0.006390014197677374, + 0.008643257431685925, + 0.0059082889929413795, + 0.008077654987573624, + 0.016297802329063416, + 0.005972799379378557, + 0.017719214782118797, + 0.010372035205364227, + 0.010830610059201717, + 0.012340807355940342, + 0.013779925182461739, + 0.012715772725641727, + 0.011112532578408718, + 0.012350200675427914, + 0.01733214780688286, + 0.019274812191724777, + 0.006901262328028679, + 0.006617770530283451, + 0.0046511744149029255, + 0.009890114888548851, + 0.012446869164705276, + 0.006262683309614658, + 0.026538603007793427, + 0.02475527487695217, + 0.00751284696161747, + 0.006104386877268553, + 0.006158872507512569, + 0.010537405498325825, + 0.011195728555321693, + 0.007999056950211525, + 0.03501371294260025, + 0.004425198771059513, + 0.004102838225662708, + 0.003971554338932037, + 0.005054876208305359, + 0.007359922863543034, + 0.008924445137381554, + 0.005850415676832199, + 0.005598724819719791, + 0.011947771534323692, + 0.008181565441191196, + 0.004308837931603193, + 0.004887096118181944, + 0.028368603438138962, + 0.013911736197769642, + 0.007374645210802555, + 0.03556818142533302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.029793692752718925 + } + ], + "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/0000028/group_6/index_tensor([8124, 8124])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_6/index_tensor([8124, 8124])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1fd52054095f6bc9392e85d1530fea038d5ecd98 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_6/index_tensor([8124, 8124])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle close-up captures a modern, slim smartphone resting on a smooth, lightly textured wooden surface. Bright, direct midday sunlight streams in through an out-of-frame window, casting a sharp, warm, rectangular beam of light onto the table and partially enveloping the device. The phone's dark, polished glass screen is closed, framed by a matte graphite-colored frame, reflecting the glare as a brilliant white glare against the surface. As its internal vibration motor becomes active, the phone begins to move; it slowly and steadily rotates clockwise, almost imperceptibly spinning on its chassis while drifting in gradual, unpredictable arcs across the wooden plane. The constant, rhythmic humming from within transforms into a hypnotic, silent dance, as the phone's sleek form glides in the direct sunlight, creating an atmosphere of focus and tranquility, yet subtly tense.", + "seed": 15928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_6/index_8124_seed_15928_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": -0.06586977778933942, + "vlm_logprob": -0.06586977778933942, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.06586977778933942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018445389345288277 + ], + "opposite_logprob_tokens": [ + -4.6434454917907715 + ], + "logprob_reward": -0.018445389345288277, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.018445389345288277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060180049389600754, + 0.22408826649188995, + 0.36767804622650146, + 0.041925977915525436, + 0.019959600642323494, + 0.02729247510433197, + 0.02518974244594574, + 0.05538491904735565, + 0.03536807745695114, + 0.14293286204338074 + ], + "credit_map": [ + 0.004679122008383274, + 0.003065189579501748, + 0.014026961289346218, + 0.006088601425290108, + 0.005706869997084141, + 0.014280078001320362, + 0.004425023682415485, + 0.007908201776444912, + 0.020762506872415543, + 0.019268665462732315, + 0.03166805952787399, + 0.047205012291669846, + 0.02358217164874077, + 0.02345806546509266, + 0.02425912395119667, + 0.03388466313481331, + 0.023325128480792046, + 0.013739002868533134, + 0.15158748626708984, + 0.014372164383530617, + 0.005222339183092117, + 0.017438506707549095, + 0.011053814552724361, + 0.13093963265419006, + 0.006470851134508848, + 0.004205777309834957, + 0.00731995701789856, + 0.0033609645906835794, + 0.0032131962943822145, + 0.00818595476448536, + 0.003784656524658203, + 0.005384618882089853, + 0.002484353957697749, + 0.0017142207361757755, + 0.003254302777349949, + 0.0024961712770164013, + 0.0027212921995669603, + 0.002511019119992852, + 0.0023343884386122227, + 0.002443852601572871, + 0.0026215999387204647, + 0.0024143869522958994, + 0.0036719299387186766, + 0.0018660163041204214, + 0.0033855284564197063, + 0.007220572791993618, + 0.0014354756567627192, + 0.004676965530961752, + 0.0036674703005701303, + 0.0024051940999925137, + 0.005049084313213825, + 0.0019984752871096134, + 0.0030020587146282196, + 0.0031546796672046185, + 0.0026032086461782455, + 0.0033095735125243664, + 0.009026378393173218, + 0.005572096444666386, + 0.00750710116699338, + 0.007390480954200029, + 0.0067486888729035854, + 0.005554646253585815, + 0.008015916682779789, + 0.005569608882069588, + 0.005637874361127615, + 0.002026904374361038, + 0.009430013597011566, + 0.003254079958423972, + 0.003539999946951866, + 0.003458429127931595, + 0.00437266007065773, + 0.003648114390671253, + 0.008332552388310432, + 0.005594269838184118, + 0.008828943595290184, + 0.05491182208061218, + 0.0030799887608736753, + 0.005754487123340368, + 0.03287579491734505, + 0.023554988205432892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.018445389345288277 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022067499812692404 + ], + "opposite_logprob_tokens": [ + -6.127206802368164 + ], + "logprob_reward": -0.0022067499812692404, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022067499812692404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04751120135188103, + 0.10415173321962357, + 0.29899266362190247, + 0.03319409862160683, + 0.023331742733716965, + 0.35028043389320374, + 0.010015956126153469, + 0.02623404562473297, + 0.030291536822915077, + 0.07599649578332901 + ], + "credit_map": [ + 0.007680533453822136, + 0.0023822637740522623, + 0.00988423265516758, + 0.0029857060872018337, + 0.003734939033165574, + 0.009572400711476803, + 0.00401946809142828, + 0.007251659408211708, + 0.009092162363231182, + 0.008041681721806526, + 0.01959066465497017, + 0.01913168840110302, + 0.009302351623773575, + 0.008981965482234955, + 0.009605459868907928, + 0.020405761897563934, + 0.01969774439930916, + 0.011276904493570328, + 0.12184472382068634, + 0.01280145812779665, + 0.004833914805203676, + 0.01456559356302023, + 0.009830605238676071, + 0.10414175689220428, + 0.005065925884991884, + 0.0031131997238844633, + 0.006136900279670954, + 0.0031636126805096865, + 0.00304342876188457, + 0.005105396267026663, + 0.002270320663228631, + 0.005295314826071262, + 0.004021943546831608, + 0.001972770318388939, + 0.0027368240989744663, + 0.003249844303354621, + 0.002602003514766693, + 0.003512904280796647, + 0.0020154081284999847, + 0.0032200475689023733, + 0.024512512609362602, + 0.026765452697873116, + 0.02967410907149315, + 0.015647072345018387, + 0.0458376519382, + 0.1226014718413353, + 0.014864140190184116, + 0.07037808001041412, + 0.0017766905948519707, + 0.0011291627306491137, + 0.00152453628834337, + 0.0009526109788566828, + 0.001186214853078127, + 0.0009516009013168514, + 0.0011642460012808442, + 0.0013308955822139978, + 0.004495374392718077, + 0.0029607100877910852, + 0.00362226739525795, + 0.0033250178676098585, + 0.003315539797767997, + 0.0027014054358005524, + 0.003147826297208667, + 0.002665906446054578, + 0.005186113994568586, + 0.00182812730781734, + 0.007500376086682081, + 0.002536457497626543, + 0.0027579073794186115, + 0.003212232841178775, + 0.004257306922227144, + 0.0030130192171782255, + 0.006907104514539242, + 0.005793581251055002, + 0.0070951455272734165, + 0.022987831383943558, + 0.003926636651158333, + 0.007681775372475386, + 0.015306755900382996, + 0.006297665648162365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.0022067499812692404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100012592971325 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004100012592971325, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100012592971325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08303118497133255, + 0.27024003863334656, + 0.18062575161457062, + 0.1529456228017807, + 0.024163801223039627, + 0.20016291737556458, + 0.017779847607016563, + 0.022758524864912033, + 0.018252898007631302, + 0.030039463192224503 + ], + "credit_map": [ + 0.008117352612316608, + 0.004006615839898586, + 0.023209305480122566, + 0.0048832520842552185, + 0.0068442909978330135, + 0.020057300105690956, + 0.0069390153512358665, + 0.008974051102995872, + 0.024243444204330444, + 0.02059858851134777, + 0.04478231817483902, + 0.04958978667855263, + 0.02440943568944931, + 0.02780502289533615, + 0.026836929842829704, + 0.051974497735500336, + 0.011367592960596085, + 0.007043201941996813, + 0.07468420267105103, + 0.006669042631983757, + 0.0025850494857877493, + 0.009021430276334286, + 0.004749635234475136, + 0.06450559198856354, + 0.02176341600716114, + 0.03109106793999672, + 0.017137998715043068, + 0.014276964589953423, + 0.01844099722802639, + 0.020797230303287506, + 0.008977927267551422, + 0.020460018888115883, + 0.003349914215505123, + 0.00216899486258626, + 0.0034295236691832542, + 0.0035125601571053267, + 0.0026744205970317125, + 0.0030940892174839973, + 0.0027276522014290094, + 0.003206647699698806, + 0.014109577983617783, + 0.01567424274981022, + 0.016326062381267548, + 0.008528965525329113, + 0.026499787345528603, + 0.07033692300319672, + 0.008493876084685326, + 0.040193479508161545, + 0.0027344822883605957, + 0.001523675280623138, + 0.0016563504468649626, + 0.0018954150145873427, + 0.0023825354874134064, + 0.0028556135948747396, + 0.002200861694291234, + 0.0025309149641543627, + 0.0035685370676219463, + 0.0023156730458140373, + 0.003123535541817546, + 0.003297196701169014, + 0.0026620852295309305, + 0.0018374385545030236, + 0.0036415481008589268, + 0.0023125112056732178, + 0.0033028998877853155, + 0.0016086101531982422, + 0.0032351436093449593, + 0.0014482267433777452, + 0.0018350433092564344, + 0.002563845133408904, + 0.0026855848263949156, + 0.0015735446941107512, + 0.002558088395744562, + 0.001673756050877273, + 0.0027290540747344494, + 0.010305759496986866, + 0.0012500794837251306, + 0.0019367116037756205, + 0.007427772507071495, + 0.002158242277801037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.004100012592971325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813129723072052 + ], + "opposite_logprob_tokens": [ + -1.4063129425048828 + ], + "logprob_reward": -0.2813129723072052, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813129723072052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05841628834605217, + 0.2531912922859192, + 0.1846865862607956, + 0.07609505206346512, + 0.053556788712739944, + 0.14230667054653168, + 0.029942907392978668, + 0.05744338035583496, + 0.09355653822422028, + 0.05080443248152733 + ], + "credit_map": [ + 0.004344043787568808, + 0.004650189075618982, + 0.014131251722574234, + 0.004353698808699846, + 0.006554849911481142, + 0.012568114325404167, + 0.004553751088678837, + 0.007260394282639027, + 0.022214621305465698, + 0.01836235076189041, + 0.04762502759695053, + 0.04235479608178139, + 0.020459961146116257, + 0.025893816724419594, + 0.021866172552108765, + 0.05441456288099289, + 0.011126573197543621, + 0.007012925110757351, + 0.07323563098907471, + 0.009403176605701447, + 0.0036793763283640146, + 0.00810933392494917, + 0.007067463826388121, + 0.06505212187767029, + 0.02287399210035801, + 0.009413334541022778, + 0.009919011034071445, + 0.00504941726103425, + 0.004689260385930538, + 0.009889847598969936, + 0.004600178450345993, + 0.009660015814006329, + 0.015398598276078701, + 0.0043598683550953865, + 0.006668903399258852, + 0.004723582416772842, + 0.004707964137196541, + 0.005255189258605242, + 0.005287402309477329, + 0.007155281957238913, + 0.009643012657761574, + 0.010395867750048637, + 0.01305418647825718, + 0.006833466235548258, + 0.018487313762307167, + 0.049625273793935776, + 0.0058991629630327225, + 0.028368396684527397, + 0.0038799236062914133, + 0.0024190039839595556, + 0.003984637558460236, + 0.003366548800840974, + 0.0036114961840212345, + 0.004608698654919863, + 0.004029299598187208, + 0.004043299704790115, + 0.01056760549545288, + 0.00600970396772027, + 0.006909245625138283, + 0.007205985486507416, + 0.007145407609641552, + 0.00690001854673028, + 0.007088509388267994, + 0.005616904702037573, + 0.015071175061166286, + 0.00558516476303339, + 0.023638973012566566, + 0.007783178705722094, + 0.009485042653977871, + 0.009800245054066181, + 0.012527202256023884, + 0.009665563702583313, + 0.005290101747959852, + 0.0029700594022870064, + 0.005069421138614416, + 0.013479012995958328, + 0.002652459777891636, + 0.003467235015705228, + 0.009637389332056046, + 0.008238754235208035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.2813129723072052 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283764719963074 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283764719963074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283764719963074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13468243181705475, + 0.1708059459924698, + 0.17832060158252716, + 0.052361663430929184, + 0.03705267235636711, + 0.13266216218471527, + 0.023335127159953117, + 0.04276638105511665, + 0.15477409958839417, + 0.07323894649744034 + ], + "credit_map": [ + 0.011806659400463104, + 0.006761285476386547, + 0.03822096437215805, + 0.010006641037762165, + 0.011202682740986347, + 0.03482889384031296, + 0.010321573354303837, + 0.011533728800714016, + 0.016542091965675354, + 0.013209554366767406, + 0.023574844002723694, + 0.031564678996801376, + 0.015860524028539658, + 0.021265732124447823, + 0.01869008131325245, + 0.030098436400294304, + 0.011538496240973473, + 0.0064466288313269615, + 0.07031762599945068, + 0.007527719251811504, + 0.0031714136712253094, + 0.008864950388669968, + 0.007113759405910969, + 0.06334002315998077, + 0.008031422272324562, + 0.006158020347356796, + 0.006227095145732164, + 0.005844630766659975, + 0.00468846783041954, + 0.009973534382879734, + 0.00439938111230731, + 0.007039113901555538, + 0.008136400021612644, + 0.0026306635700166225, + 0.004001969937235117, + 0.004587946925312281, + 0.004912124015390873, + 0.0041961814276874065, + 0.003947772551327944, + 0.0046396153047680855, + 0.008757623843848705, + 0.011615782976150513, + 0.009277313016355038, + 0.005858555436134338, + 0.017806509509682655, + 0.0469525121152401, + 0.005582266021519899, + 0.026811614632606506, + 0.0037406496703624725, + 0.0023727756924927235, + 0.004358941689133644, + 0.002522918861359358, + 0.002867229050025344, + 0.0021889982745051384, + 0.0027578468434512615, + 0.002525769406929612, + 0.007519930135458708, + 0.004167619161307812, + 0.005754970945417881, + 0.005143911577761173, + 0.006036856677383184, + 0.004186561331152916, + 0.006050060968846083, + 0.0039064716547727585, + 0.023255299776792526, + 0.009275799617171288, + 0.03924445062875748, + 0.012499473989009857, + 0.01572776585817337, + 0.01399161946028471, + 0.024441681802272797, + 0.016338016837835312, + 0.006640532985329628, + 0.003324306569993496, + 0.005075251217931509, + 0.02344840206205845, + 0.002458699746057391, + 0.004274263046681881, + 0.014069868251681328, + 0.0139476228505373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.023283764719963074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a modern, slim smartphone resting on a smooth, lightly textured wooden surface. Bright, direct midday sunlight streams in through an out-of-frame window, casting a sharp, warm, rectangular beam of light onto the table and partially enveloping the device. The phone's dark, polished glass screen is closed, framed by a matte graphite-colored frame, reflecting the glare as a brilliant white glare against the surface. As its internal vibration motor becomes active, the phone begins to move; it slowly and steadily rotates clockwise, almost imperceptibly spinning on its chassis while drifting in gradual, unpredictable arcs across the wooden plane. The constant, rhythmic humming from within transforms into a hypnotic, silent dance, as the phone's sleek form glides in the direct sunlight, creating an atmosphere of focus and tranquility, yet subtly tense.", + "seed": 15929, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_6/index_8124_seed_15929_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": -0.032857080455869436, + "vlm_logprob": -0.032857080455869436, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.032857080455869436, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014611383900046349 + ], + "opposite_logprob_tokens": [ + -5.13961124420166 + ], + "logprob_reward": -0.014611383900046349, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.014611383900046349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04185330122709274, + 0.0215787123888731, + 0.016017194837331772, + 0.012512299232184887, + 0.029078876599669456, + 0.017793549224734306, + 0.046757813543081284, + 0.02632463350892067, + 0.770338773727417, + 0.01774490624666214 + ], + "credit_map": [ + 0.010716928169131279, + 0.004928207490593195, + 0.00539948046207428, + 0.00539354607462883, + 0.003134799422696233, + 0.0049524749629199505, + 0.0028961829375475645, + 0.004431683104485273, + 0.0024421303533017635, + 0.002346973866224289, + 0.0020784151274710894, + 0.0040052300319075584, + 0.0028859006706625223, + 0.0036863270215690136, + 0.0022559473291039467, + 0.0018777868244796991, + 0.00181620754301548, + 0.0018691286677494645, + 0.0019777456764131784, + 0.001289897132664919, + 0.001985330367460847, + 0.003695124527439475, + 0.0011498916428536177, + 0.002233868930488825, + 0.0014892482431605458, + 0.002401416888460517, + 0.002651633694767952, + 0.0011775020975619555, + 0.0011546688620001078, + 0.0012019900605082512, + 0.0009670405997894704, + 0.001468798378482461, + 0.0032174999359995127, + 0.006476297974586487, + 0.0062098149210214615, + 0.0027703833766281605, + 0.0027194982394576073, + 0.00202277977950871, + 0.0021732752211391926, + 0.0034893264528363943, + 0.0023718480952084064, + 0.0025197763461619616, + 0.0026290910318493843, + 0.0020796044263988733, + 0.001651251339353621, + 0.0014888239093124866, + 0.0018413262441754341, + 0.0032118281815201044, + 0.004892230965197086, + 0.005549026653170586, + 0.0051656863652169704, + 0.006532180588692427, + 0.009562601335346699, + 0.004824555944651365, + 0.005916448775678873, + 0.004315080586820841, + 0.0033476664684712887, + 0.003563067875802517, + 0.004791995510458946, + 0.0030768963042646646, + 0.0028585256077349186, + 0.002525004791095853, + 0.0029729888774454594, + 0.003188487607985735, + 0.06312534958124161, + 0.014696897938847542, + 0.33269187808036804, + 0.023962806910276413, + 0.012489979155361652, + 0.025799652561545372, + 0.017010828480124474, + 0.2805613875389099, + 0.0021281770896166563, + 0.001403701608069241, + 0.0034515804145485163, + 0.0026074517518281937, + 0.0012524576159194112, + 0.001292313332669437, + 0.002270330209285021, + 0.0033388929441571236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.014611383900046349 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002499315422028303 + ], + "opposite_logprob_tokens": [ + -6.002499103546143 + ], + "logprob_reward": -0.002499315422028303, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002499315422028303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021513136103749275, + 0.020564217120409012, + 0.01203977596014738, + 0.019254473969340324, + 0.02171710878610611, + 0.09044996649026871, + 0.020879270508885384, + 0.04868268966674805, + 0.7206425666809082, + 0.024256780743598938 + ], + "credit_map": [ + 0.0037265559658408165, + 0.002923026215285063, + 0.00366221833974123, + 0.002471436047926545, + 0.0021718384232372046, + 0.0028572429437190294, + 0.0014801784418523312, + 0.002220638794824481, + 0.0024178773164749146, + 0.00279758358374238, + 0.00326326466165483, + 0.002694615861400962, + 0.003120457986369729, + 0.002842406742274761, + 0.0016603524563834071, + 0.001767659094184637, + 0.0011464329436421394, + 0.0020160290878266096, + 0.0011652912944555283, + 0.0013975930633023381, + 0.0016617935616523027, + 0.0014565896708518267, + 0.0013640249380841851, + 0.0018320209346711636, + 0.003194481832906604, + 0.003800593549385667, + 0.0021075576078146696, + 0.0019670650362968445, + 0.002336343051865697, + 0.0017210585065186024, + 0.0016097520710900426, + 0.0025176224298775196, + 0.0022945033852010965, + 0.004899638704955578, + 0.004058057442307472, + 0.0022469402756541967, + 0.001995464088395238, + 0.001483940752223134, + 0.0021946558263152838, + 0.002543909242376685, + 0.010759874247014523, + 0.011676140129566193, + 0.012274788692593575, + 0.010551323182880878, + 0.011616142466664314, + 0.009971915744245052, + 0.009079327806830406, + 0.01452045701444149, + 0.00239824247546494, + 0.002102437661960721, + 0.0028746325988322496, + 0.0022567054256796837, + 0.00359636009670794, + 0.0021036616526544094, + 0.0021128037478774786, + 0.0034344277810305357, + 0.0031840328592807055, + 0.0038188279140740633, + 0.016287868842482567, + 0.0031235101632773876, + 0.0027682047802954912, + 0.002623987151309848, + 0.00303355953656137, + 0.013842696323990822, + 0.05851363018155098, + 0.014681716449558735, + 0.309521347284317, + 0.023305801674723625, + 0.013125951401889324, + 0.024592043831944466, + 0.017417941242456436, + 0.259484201669693, + 0.0037854358088225126, + 0.0021466510370373726, + 0.00398261658847332, + 0.0035686269402503967, + 0.0018912248779088259, + 0.002497363369911909, + 0.0029514438938349485, + 0.0034334196243435144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.002499315422028303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952251143753529 + ], + "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": [ + 0.010983257554471493, + 0.006899458356201649, + 0.029270729050040245, + 0.010526074096560478, + 0.04915906861424446, + 0.01573662832379341, + 0.0116205969825387, + 0.026397904381155968, + 0.8193824887275696, + 0.020023740828037262 + ], + "credit_map": [ + 0.0021816478110849857, + 0.002161191776394844, + 0.0013970775762572885, + 0.001208412810228765, + 0.0010244442382827401, + 0.0010844224598258734, + 0.0007483644294552505, + 0.0011776963947340846, + 0.0009404798620380461, + 0.0009853749070316553, + 0.000857804378028959, + 0.0011601102305576205, + 0.0008020667592063546, + 0.0010769099462777376, + 0.0005250952090136707, + 0.0005516170058399439, + 0.003317766822874546, + 0.0035228480119258165, + 0.004907481838017702, + 0.0022630549501627684, + 0.00403214804828167, + 0.004770169034600258, + 0.002253143582493067, + 0.004204119555652142, + 0.0013642790727317333, + 0.0019270353950560093, + 0.002217461820691824, + 0.0011602449230849743, + 0.00093477638438344, + 0.0007717757252976298, + 0.0006919376901350915, + 0.0014585629105567932, + 0.005806449335068464, + 0.011876787059009075, + 0.011776610277593136, + 0.0042028422467410564, + 0.004017514176666737, + 0.0031467198859900236, + 0.0034491801634430885, + 0.004882967099547386, + 0.001318508293479681, + 0.0023099796380847692, + 0.0015847637550905347, + 0.0026013669557869434, + 0.0020641698502004147, + 0.001931383740156889, + 0.0019138377392664552, + 0.0020126167219132185, + 0.0015812523197382689, + 0.001607405603863299, + 0.0015606275992468, + 0.001045574783347547, + 0.001217117765918374, + 0.0014381323708221316, + 0.0012394142104312778, + 0.0019310720963403583, + 0.0020222244784235954, + 0.0012216523755341768, + 0.009879075922071934, + 0.0014872848987579346, + 0.0010258505353704095, + 0.001519075594842434, + 0.0010220878757536411, + 0.008220651187002659, + 0.06882984936237335, + 0.01591680198907852, + 0.35431012511253357, + 0.02535545453429222, + 0.013867183588445187, + 0.02828732505440712, + 0.0181939247995615, + 0.29462185502052307, + 0.0033894293010234833, + 0.0016153431497514248, + 0.0031281562987715006, + 0.0022422978654503822, + 0.0016567993443459272, + 0.00232442794367671, + 0.0025024358183145523, + 0.0031648518051952124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "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 a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703220546245575 + ], + "opposite_logprob_tokens": [ + -2.1270322799682617 + ], + "logprob_reward": -0.12703220546245575, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703220546245575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.005421774927526712, + 0.006306307390332222, + 0.007244493346661329, + 0.005531046073883772, + 0.014211795292794704, + 0.019016969949007034, + 0.007381440605968237, + 0.03976815193891525, + 0.8888067603111267, + 0.006311262957751751 + ], + "credit_map": [ + 0.0011007982539013028, + 0.0008498847601003945, + 0.0006027462077327073, + 0.0004721175064332783, + 0.0004385957436170429, + 0.0005831936723552644, + 0.0005141054862178862, + 0.0008603330934420228, + 0.000622543680947274, + 0.0006737133371643722, + 0.0006773226195946336, + 0.0011564932065084577, + 0.0008194794063456357, + 0.00085067207692191, + 0.0007080130744725466, + 0.0007980699883773923, + 0.0009286236600019038, + 0.0006407831097021699, + 0.0006894905236549675, + 0.0016166430432349443, + 0.0006906895432621241, + 0.0008984726155176759, + 0.0005904835998080671, + 0.0011893074261024594, + 0.000759243848733604, + 0.0008646381902508438, + 0.0009881524601951241, + 0.000756666180677712, + 0.00048028150922618806, + 0.0004677984688896686, + 0.00048053078353405, + 0.0007337344577535987, + 0.001501463120803237, + 0.0018991176038980484, + 0.00302027422003448, + 0.0015239038038998842, + 0.001153380493633449, + 0.0012386178132146597, + 0.0014538217801600695, + 0.0024212170392274857, + 0.0016949065029621124, + 0.0027224686928093433, + 0.002137990202754736, + 0.003341980744153261, + 0.0026372987776994705, + 0.0023739056196063757, + 0.0021721916273236275, + 0.0019362288294360042, + 0.001074447063729167, + 0.0010667291935533285, + 0.0008583222515881062, + 0.0008915869402699172, + 0.00083836232079193, + 0.0009572167764417827, + 0.0008914992213249207, + 0.000803276663646102, + 0.003431203542277217, + 0.001416129874996841, + 0.0155866090208292, + 0.002188452985137701, + 0.0014529405161738396, + 0.0018950802041217685, + 0.0010743247112259269, + 0.012723415158689022, + 0.07428494095802307, + 0.016856322064995766, + 0.3830999732017517, + 0.028146930038928986, + 0.01606704108417034, + 0.030533507466316223, + 0.0205013919621706, + 0.3193166255950928, + 0.00046882685273885727, + 0.0006768490420654416, + 0.0009580669575370848, + 0.000995574169792235, + 0.0007325217593461275, + 0.0005612826789729297, + 0.0009494094992987812, + 0.0009687317069619894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.12703220546245575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190246351063251 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190246351063251, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190246351063251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0700252503156662, + 0.046089090406894684, + 0.05800168588757515, + 0.045509759336709976, + 0.24356576800346375, + 0.14045105874538422, + 0.037166714668273926, + 0.09060230851173401, + 0.20633472502231598, + 0.062253668904304504 + ], + "credit_map": [ + 0.014673522673547268, + 0.012394825927913189, + 0.01031392440199852, + 0.006640021223574877, + 0.0060491543263196945, + 0.008232134394347668, + 0.004536319058388472, + 0.007185355294495821, + 0.007395145483314991, + 0.005748141556978226, + 0.005583224818110466, + 0.007186897564679384, + 0.005259332712739706, + 0.005688583943992853, + 0.0037867226637899876, + 0.00544104166328907, + 0.006118185352534056, + 0.005803301930427551, + 0.005730637349188328, + 0.008179224096238613, + 0.006938851438462734, + 0.012150588445365429, + 0.005295026581734419, + 0.0077858720906078815, + 0.00613805279135704, + 0.007584634702652693, + 0.007592036854475737, + 0.005643623415380716, + 0.004196999594569206, + 0.004184179473668337, + 0.0038181678391993046, + 0.006352066528052092, + 0.02665182203054428, + 0.049422487616539, + 0.05485398322343826, + 0.023157281801104546, + 0.017590021714568138, + 0.01863761618733406, + 0.024998344480991364, + 0.02825421467423439, + 0.01710217073559761, + 0.01905757561326027, + 0.01636001467704773, + 0.015198846347630024, + 0.018590716645121574, + 0.01597019098699093, + 0.015503490343689919, + 0.02266804501414299, + 0.004032507073134184, + 0.004784667398780584, + 0.00469953054562211, + 0.0057515669614076614, + 0.004584755748510361, + 0.005248905159533024, + 0.0040931240655481815, + 0.00397165771573782, + 0.005967054050415754, + 0.01284803543239832, + 0.025632895529270172, + 0.005841980688273907, + 0.005526284221559763, + 0.007046251092106104, + 0.0047772107645869255, + 0.022962594404816628, + 0.015501121990382671, + 0.03756963089108467, + 0.0380035936832428, + 0.020352406427264214, + 0.02130999229848385, + 0.023954080417752266, + 0.017799239605665207, + 0.03184466436505318, + 0.008115817792713642, + 0.005073255859315395, + 0.01347061712294817, + 0.005501900799572468, + 0.004111272282898426, + 0.004894710611552, + 0.006246004719287157, + 0.014840093441307545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.014190246351063251 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a modern, slim smartphone resting on a smooth, lightly textured wooden surface. Bright, direct midday sunlight streams in through an out-of-frame window, casting a sharp, warm, rectangular beam of light onto the table and partially enveloping the device. The phone's dark, polished glass screen is closed, framed by a matte graphite-colored frame, reflecting the glare as a brilliant white glare against the surface. As its internal vibration motor becomes active, the phone begins to move; it slowly and steadily rotates clockwise, almost imperceptibly spinning on its chassis while drifting in gradual, unpredictable arcs across the wooden plane. The constant, rhythmic humming from within transforms into a hypnotic, silent dance, as the phone's sleek form glides in the direct sunlight, creating an atmosphere of focus and tranquility, yet subtly tense.", + "seed": 15930, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_6/index_8124_seed_15930_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": -0.029334725951775907, + "vlm_logprob": -0.029334725951775907, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.029334725951775907, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01461150124669075 + ], + "opposite_logprob_tokens": [ + -4.764611721038818 + ], + "logprob_reward": -0.01461150124669075, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01461150124669075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14140576124191284, + 0.19960357248783112, + 0.13243450224399567, + 0.08758262544870377, + 0.07924234867095947, + 0.06203583627939224, + 0.12991824746131897, + 0.08299344033002853, + 0.054594580084085464, + 0.030188996344804764 + ], + "credit_map": [ + 0.019818631932139397, + 0.021334333345294, + 0.015568665228784084, + 0.01665988191962242, + 0.014050981029868126, + 0.01869702711701393, + 0.015125996433198452, + 0.020150260999798775, + 0.026660416275262833, + 0.017965584993362427, + 0.03750140219926834, + 0.02102276124060154, + 0.015657475218176842, + 0.03117448277771473, + 0.022388186305761337, + 0.027233298867940903, + 0.024021996185183525, + 0.017213061451911926, + 0.01797165907919407, + 0.014291602186858654, + 0.015038451179862022, + 0.017585953697562218, + 0.010005831718444824, + 0.0163059551268816, + 0.012124743312597275, + 0.010794120840728283, + 0.011117285117506981, + 0.007258611731231213, + 0.014998880214989185, + 0.012561126612126827, + 0.006872786208987236, + 0.01185508631169796, + 0.007955662906169891, + 0.009879199787974358, + 0.012192332185804844, + 0.008287415839731693, + 0.00870784092694521, + 0.012522563338279724, + 0.009847821667790413, + 0.009849515743553638, + 0.008583267219364643, + 0.005770040210336447, + 0.009237384423613548, + 0.00663943262770772, + 0.0035568950697779655, + 0.00972285121679306, + 0.011442276649177074, + 0.007083692587912083, + 0.007667106576263905, + 0.015423920936882496, + 0.02010529488325119, + 0.022994766011834145, + 0.013150978833436966, + 0.018931658938527107, + 0.016755755990743637, + 0.014888784848153591, + 0.010364610701799393, + 0.007367302663624287, + 0.012070872820913792, + 0.008218037895858288, + 0.008418051525950432, + 0.015718121081590652, + 0.008329684846103191, + 0.012506771832704544, + 0.00658232718706131, + 0.004931819159537554, + 0.009154594503343105, + 0.005855550058186054, + 0.007034990470856428, + 0.007486657705157995, + 0.005774660501629114, + 0.007773988414555788, + 0.003992898855358362, + 0.002971006790176034, + 0.004813574254512787, + 0.0037933422718197107, + 0.004116757307201624, + 0.0027125035412609577, + 0.00398772768676281, + 0.003801188198849559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01461150124669075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002492894185706973 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.002492894185706973, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002492894185706973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3153197467327118, + 0.10701845586299896, + 0.0818445160984993, + 0.07285857200622559, + 0.09183799475431442, + 0.11989448964595795, + 0.05238128453493118, + 0.05276033654808998, + 0.07090725004673004, + 0.035177331417798996 + ], + "credit_map": [ + 0.06130087003111839, + 0.04234480485320091, + 0.047568537294864655, + 0.022458823397755623, + 0.025547275319695473, + 0.0461130365729332, + 0.021919814869761467, + 0.04806660860776901, + 0.01805991306900978, + 0.012181180529296398, + 0.01774018444120884, + 0.012941270135343075, + 0.008722684346139431, + 0.01363667193800211, + 0.009787212125957012, + 0.01394932996481657, + 0.01368874404579401, + 0.008669854141771793, + 0.014593849889934063, + 0.007963750511407852, + 0.010226880200207233, + 0.009256735444068909, + 0.007214892655611038, + 0.010229806415736675, + 0.009125304408371449, + 0.00778823671862483, + 0.015138523653149605, + 0.00615524360910058, + 0.009965768083930016, + 0.0083400197327137, + 0.007861390709877014, + 0.008484083227813244, + 0.01090738270431757, + 0.010395701974630356, + 0.012834261171519756, + 0.013989992439746857, + 0.008110446855425835, + 0.012582714669406414, + 0.012670282274484634, + 0.0103472163900733, + 0.024141419678926468, + 0.008429729379713535, + 0.021967975422739983, + 0.008548850193619728, + 0.00642821379005909, + 0.024167397990822792, + 0.012513555586338043, + 0.013697342947125435, + 0.005040914285928011, + 0.007074378430843353, + 0.005812611896544695, + 0.0077047813683748245, + 0.006990252994000912, + 0.006214142311364412, + 0.0062537891790270805, + 0.007290418725460768, + 0.007022930774837732, + 0.005810101516544819, + 0.008687674067914486, + 0.005209655966609716, + 0.005678748246282339, + 0.007848542183637619, + 0.005860565695911646, + 0.006642120890319347, + 0.00852140225470066, + 0.006297485437244177, + 0.013499281369149685, + 0.006329034920781851, + 0.008887432515621185, + 0.009879220277071, + 0.007527358364313841, + 0.00996603537350893, + 0.0055123972706496716, + 0.0024999442975968122, + 0.007158758584409952, + 0.003268314292654395, + 0.0038780292961746454, + 0.0041916645132005215, + 0.0043710325844585896, + 0.004297191742807627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.002492894185706973 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350490748882294 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350490748882294, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350490748882294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28895771503448486, + 0.16863670945167542, + 0.1685364544391632, + 0.05590512603521347, + 0.12406454235315323, + 0.048046912997961044, + 0.05659853667020798, + 0.03497910872101784, + 0.030266650021076202, + 0.02400832623243332 + ], + "credit_map": [ + 0.0569772832095623, + 0.04127344861626625, + 0.04380197450518608, + 0.02263467200100422, + 0.020387671887874603, + 0.03805965930223465, + 0.02114693634212017, + 0.04467606917023659, + 0.031177472323179245, + 0.020734230056405067, + 0.02808675356209278, + 0.015942946076393127, + 0.013842699117958546, + 0.021108150482177734, + 0.0157383494079113, + 0.02200610749423504, + 0.029040507972240448, + 0.022071203216910362, + 0.022340985015034676, + 0.013208945281803608, + 0.01536371186375618, + 0.02524818293750286, + 0.018169691786170006, + 0.02309323474764824, + 0.007183569483458996, + 0.008106025867164135, + 0.008968736976385117, + 0.0044646309688687325, + 0.008159564808011055, + 0.007446311879903078, + 0.00436789495870471, + 0.007208388764411211, + 0.015899688005447388, + 0.015170830301940441, + 0.019311800599098206, + 0.01558363065123558, + 0.010943042114377022, + 0.017506897449493408, + 0.016459353268146515, + 0.01318931020796299, + 0.006308726500719786, + 0.004596539307385683, + 0.009495495818555355, + 0.004413094371557236, + 0.001936631160788238, + 0.009014655835926533, + 0.0065758065320551395, + 0.005705964285880327, + 0.004219328984618187, + 0.007222512736916542, + 0.006574283819645643, + 0.010115680284798145, + 0.005516808480024338, + 0.007050564046949148, + 0.008929013274610043, + 0.006970344111323357, + 0.004028216935694218, + 0.0031726628076285124, + 0.005537038668990135, + 0.003365150187164545, + 0.0032148147001862526, + 0.006990931928157806, + 0.003268344094976783, + 0.005401952192187309, + 0.002923844614997506, + 0.002750788116827607, + 0.00438011484220624, + 0.003961816430091858, + 0.002762041287496686, + 0.005488046444952488, + 0.003964872099459171, + 0.0040351226925849915, + 0.0034423915203660727, + 0.001680644229054451, + 0.004228964913636446, + 0.002954000374302268, + 0.0028578918427228928, + 0.002168386010453105, + 0.0028281293343752623, + 0.0038479191716760397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026350490748882294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902779221534729 + ], + "opposite_logprob_tokens": [ + -2.4640278816223145 + ], + "logprob_reward": -0.08902779221534729, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902779221534729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07573888450860977, + 0.2777892053127289, + 0.06972945481538773, + 0.03787492588162422, + 0.28769490122795105, + 0.057346004992723465, + 0.04124167934060097, + 0.06508176028728485, + 0.051985569298267365, + 0.035517629235982895 + ], + "credit_map": [ + 0.013521503657102585, + 0.006609056610614061, + 0.009514817036688328, + 0.007424081210047007, + 0.00566869555041194, + 0.010975399985909462, + 0.007333527784794569, + 0.014691799879074097, + 0.03787903115153313, + 0.0366506390273571, + 0.04445510730147362, + 0.036871228367090225, + 0.024687716737389565, + 0.03580664470791817, + 0.026905369013547897, + 0.03453344479203224, + 0.01407722756266594, + 0.006710642948746681, + 0.012784535996615887, + 0.007138701155781746, + 0.006858753506094217, + 0.0059789614751935005, + 0.006635832600295544, + 0.0095447963103652, + 0.004361833445727825, + 0.004496630746871233, + 0.006792258005589247, + 0.004031942691653967, + 0.004938176833093166, + 0.004140298347920179, + 0.004631878808140755, + 0.004481906536966562, + 0.02319415844976902, + 0.051398273557424545, + 0.03323984891176224, + 0.035511113703250885, + 0.02740270458161831, + 0.03331082686781883, + 0.04553358629345894, + 0.038104403764009476, + 0.010446000844240189, + 0.0041422671638429165, + 0.008180704899132252, + 0.004583870526403189, + 0.004201539792120457, + 0.01054389588534832, + 0.007730628829449415, + 0.0075170984491705894, + 0.0033749090507626534, + 0.004079736303538084, + 0.004683015868067741, + 0.007056158035993576, + 0.004231583792716265, + 0.004902718588709831, + 0.007454529404640198, + 0.005459029693156481, + 0.005658167414367199, + 0.0059920116327703, + 0.011960781179368496, + 0.00877697765827179, + 0.005843911319971085, + 0.010294913314282894, + 0.006631221156567335, + 0.009923775680363178, + 0.006344616413116455, + 0.005154240410774946, + 0.007676202803850174, + 0.005571802146732807, + 0.008246107958257198, + 0.0045666322112083435, + 0.007307121064513922, + 0.007118840701878071, + 0.0051516033709049225, + 0.0025780426803976297, + 0.008378534577786922, + 0.0037111197598278522, + 0.004135263618081808, + 0.004111589398235083, + 0.002940800040960312, + 0.004510677885264158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.08902779221534729 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190951362252235 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014190951362252235, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190951362252235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17915000021457672, + 0.07753179967403412, + 0.084328792989254, + 0.051109861582517624, + 0.11442262679338455, + 0.09648657590150833, + 0.0507647804915905, + 0.07907762378454208, + 0.23586343228816986, + 0.0312645360827446 + ], + "credit_map": [ + 0.024267448112368584, + 0.02512417361140251, + 0.02447224035859108, + 0.017671309411525726, + 0.01453272346407175, + 0.02366621047258377, + 0.018974611535668373, + 0.030441273003816605, + 0.010917766951024532, + 0.007429923862218857, + 0.015097982250154018, + 0.006998810917139053, + 0.005984121467918158, + 0.012473639100790024, + 0.007215157616883516, + 0.011414384469389915, + 0.015313736163079739, + 0.008625692687928677, + 0.013433101586997509, + 0.008379662409424782, + 0.008872727863490582, + 0.011803573928773403, + 0.007003189530223608, + 0.01089710183441639, + 0.007444108370691538, + 0.005805118940770626, + 0.006443245802074671, + 0.004486648365855217, + 0.006367265712469816, + 0.006979038007557392, + 0.004717772826552391, + 0.008866657502949238, + 0.00890022236853838, + 0.016211779788136482, + 0.016388148069381714, + 0.014504349790513515, + 0.009902535937726498, + 0.015835314989089966, + 0.01667090877890587, + 0.01600935496389866, + 0.011836002580821514, + 0.01044570654630661, + 0.015995917841792107, + 0.011313571594655514, + 0.005230700597167015, + 0.015220286324620247, + 0.014598321169614792, + 0.011846061795949936, + 0.005322270095348358, + 0.004942494444549084, + 0.008344640955328941, + 0.00774349644780159, + 0.004158749710768461, + 0.005173760466277599, + 0.008686794899404049, + 0.0063925692811608315, + 0.01094795111566782, + 0.009028117172420025, + 0.013616024516522884, + 0.007810467388480902, + 0.007790539413690567, + 0.01135737169533968, + 0.007472787983715534, + 0.01105434913188219, + 0.024291211739182472, + 0.021383093670010567, + 0.03425602987408638, + 0.031336650252342224, + 0.031216831877827644, + 0.02541503868997097, + 0.03690466284751892, + 0.03105989284813404, + 0.00367654743604362, + 0.0025911780539900064, + 0.00428563030436635, + 0.0033332149032503366, + 0.005279204808175564, + 0.004323235712945461, + 0.0037531335838139057, + 0.004022388253360987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.014190951362252235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a modern, slim smartphone resting on a smooth, lightly textured wooden surface. Bright, direct midday sunlight streams in through an out-of-frame window, casting a sharp, warm, rectangular beam of light onto the table and partially enveloping the device. The phone's dark, polished glass screen is closed, framed by a matte graphite-colored frame, reflecting the glare as a brilliant white glare against the surface. As its internal vibration motor becomes active, the phone begins to move; it slowly and steadily rotates clockwise, almost imperceptibly spinning on its chassis while drifting in gradual, unpredictable arcs across the wooden plane. The constant, rhythmic humming from within transforms into a hypnotic, silent dance, as the phone's sleek form glides in the direct sunlight, creating an atmosphere of focus and tranquility, yet subtly tense.", + "seed": 15931, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_6/index_8124_seed_15931_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": -0.04323669178411364, + "vlm_logprob": -0.04323669178411364, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.04323669178411364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025865618139505386 + ], + "opposite_logprob_tokens": [ + -4.15086555480957 + ], + "logprob_reward": -0.025865618139505386, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.025865618139505386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3785923421382904, + 0.06943301111459732, + 0.15854816138744354, + 0.04797549545764923, + 0.05901150032877922, + 0.06765544414520264, + 0.05571094900369644, + 0.05525810271501541, + 0.07202461361885071, + 0.03579040244221687 + ], + "credit_map": [ + 0.03737872093915939, + 0.06794650107622147, + 0.04119942709803581, + 0.032033760100603104, + 0.07201584428548813, + 0.05194764584302902, + 0.03542407974600792, + 0.04064638912677765, + 0.0042080627754330635, + 0.00503005925565958, + 0.010921435430645943, + 0.006380661856383085, + 0.006812190171331167, + 0.012863361276686192, + 0.01059112697839737, + 0.01262611709535122, + 0.008866971358656883, + 0.005636414512991905, + 0.022222431376576424, + 0.012430752627551556, + 0.008829250931739807, + 0.05308344215154648, + 0.005601170938462019, + 0.04187772050499916, + 0.005583673249930143, + 0.0050485264509916306, + 0.005295682232826948, + 0.006832275073975325, + 0.00787425972521305, + 0.006983757019042969, + 0.005189479794353247, + 0.005167840979993343, + 0.013025159947574139, + 0.006325485184788704, + 0.006646997295320034, + 0.00430102925747633, + 0.004865686874836683, + 0.008511900901794434, + 0.003932351712137461, + 0.011402887292206287, + 0.007506292778998613, + 0.007458704058080912, + 0.01268673874437809, + 0.007017579860985279, + 0.0046655405312776566, + 0.00717318756505847, + 0.011645562946796417, + 0.009501834399998188, + 0.005222890991717577, + 0.0028345403261482716, + 0.00885575357824564, + 0.00873151421546936, + 0.003971773199737072, + 0.007718001026660204, + 0.004106169566512108, + 0.014270303770899773, + 0.007213903125375509, + 0.005063725169748068, + 0.01276865229010582, + 0.005185738205909729, + 0.007126800715923309, + 0.006036474369466305, + 0.004322738852351904, + 0.007540068589150906, + 0.013393638655543327, + 0.008847939781844616, + 0.005878536030650139, + 0.004038828890770674, + 0.007570517249405384, + 0.016399318352341652, + 0.006056620739400387, + 0.00983921717852354, + 0.010700018145143986, + 0.010682974010705948, + 0.0027535315603017807, + 0.002589288167655468, + 0.0027857765089720488, + 0.001807920983992517, + 0.002065027365460992, + 0.002405863953754306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.025865618139505386 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032055205665528774 + ], + "opposite_logprob_tokens": [ + -5.753205299377441 + ], + "logprob_reward": -0.0032055205665528774, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032055205665528774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37611010670661926, + 0.0638774186372757, + 0.1964402049779892, + 0.0330948531627655, + 0.019435551017522812, + 0.020014911890029907, + 0.11376476287841797, + 0.052828747779130936, + 0.09301050752401352, + 0.03142290934920311 + ], + "credit_map": [ + 0.0378476157784462, + 0.07251735776662827, + 0.03621833026409149, + 0.032644208520650864, + 0.07487739622592926, + 0.04925984889268875, + 0.03561115264892578, + 0.03713420033454895, + 0.004274910781532526, + 0.004558823071420193, + 0.010490470565855503, + 0.005886582657694817, + 0.005385878495872021, + 0.013374460861086845, + 0.006585914641618729, + 0.013320383615791798, + 0.00905311293900013, + 0.0069261169992387295, + 0.016685282811522484, + 0.011835534125566483, + 0.009292430244386196, + 0.0748782530426979, + 0.0077584004029631615, + 0.060011085122823715, + 0.0067321001552045345, + 0.003671065904200077, + 0.00405085040256381, + 0.0032212126534432173, + 0.0031871898099780083, + 0.004513680003583431, + 0.0030806148424744606, + 0.004638142883777618, + 0.0030789345037192106, + 0.002582506276667118, + 0.0021825674921274185, + 0.0017926578875631094, + 0.002836513100191951, + 0.002321572508662939, + 0.0019923332147300243, + 0.0026484669651836157, + 0.0028230485040694475, + 0.002282496076077223, + 0.0029781865887343884, + 0.0018422581488266587, + 0.0033853533677756786, + 0.002152704633772373, + 0.0018892361549660563, + 0.0026616279501467943, + 0.014180628582835197, + 0.007412270177155733, + 0.018059827387332916, + 0.018047990277409554, + 0.00865243747830391, + 0.02131473459303379, + 0.008491329848766327, + 0.017605548724532127, + 0.007647640071809292, + 0.006492467597126961, + 0.010250138118863106, + 0.004751032683998346, + 0.004364022985100746, + 0.007495495490729809, + 0.004498656839132309, + 0.007329296786338091, + 0.018160585314035416, + 0.016438545659184456, + 0.010069269686937332, + 0.004871685057878494, + 0.00837417971342802, + 0.015421091578900814, + 0.007316594012081623, + 0.012358564883470535, + 0.007882941514253616, + 0.007663468364626169, + 0.0027812167536467314, + 0.002718524308875203, + 0.003120900597423315, + 0.002024710876867175, + 0.0026135535445064306, + 0.002617595950141549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0032055205665528774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356644928455353 + ], + "opposite_logprob_tokens": [ + -3.6513566970825195 + ], + "logprob_reward": -0.026356644928455353, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356644928455353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5173260569572449, + 0.06685465574264526, + 0.2357582449913025, + 0.02167486771941185, + 0.04415018856525421, + 0.020078668370842934, + 0.010017170570790768, + 0.020239856094121933, + 0.04051029682159424, + 0.023389944806694984 + ], + "credit_map": [ + 0.041963983327150345, + 0.10544819384813309, + 0.052658841013908386, + 0.04202883318066597, + 0.11119605600833893, + 0.06513334065675735, + 0.043920256197452545, + 0.054976608604192734, + 0.0043890527449548244, + 0.004707917105406523, + 0.011861027218401432, + 0.005990171805024147, + 0.005705818068236113, + 0.013602348044514656, + 0.008222031407058239, + 0.012376293540000916, + 0.011840587481856346, + 0.008840612135827541, + 0.02457222156226635, + 0.017015788704156876, + 0.011855807155370712, + 0.08439675718545914, + 0.009873438626527786, + 0.06736303120851517, + 0.004172112327069044, + 0.0026288400404155254, + 0.0026180564891546965, + 0.0026825431268662214, + 0.0022525552194565535, + 0.0026600288692861795, + 0.001942659611813724, + 0.0027180733159184456, + 0.009221176616847515, + 0.0052362713031470776, + 0.0051851412281394005, + 0.003566028317436576, + 0.004538416862487793, + 0.005851970054209232, + 0.004008940886706114, + 0.00654224818572402, + 0.0026838060002774, + 0.0021826764568686485, + 0.003436540486291051, + 0.001603388343937695, + 0.001961409579962492, + 0.0030913716182112694, + 0.0017217417480424047, + 0.0033977359998971224, + 0.0015660622157156467, + 0.0009505279595032334, + 0.0013495918828994036, + 0.0013179263332858682, + 0.001257469761185348, + 0.0012611160054802895, + 0.0008872587932273746, + 0.0014272195985540748, + 0.0027105119079351425, + 0.0028835118282586336, + 0.004244403447955847, + 0.001372354687191546, + 0.0022530844435095787, + 0.002504311501979828, + 0.0015089656226336956, + 0.0027627150993794203, + 0.0052438280545175076, + 0.005353007465600967, + 0.004000850021839142, + 0.0017246983479708433, + 0.0037948491517454386, + 0.012340779416263103, + 0.0025833388790488243, + 0.005468946415930986, + 0.006005285773426294, + 0.006184272933751345, + 0.0016377328429371119, + 0.0014688780065625906, + 0.0031514784786850214, + 0.0016951239667832851, + 0.0017014916520565748, + 0.0015456825494766235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.026356644928455353 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705688178539276 + ], + "opposite_logprob_tokens": [ + -2.127056837081909 + ], + "logprob_reward": -0.12705688178539276, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705688178539276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5257713198661804, + 0.0380081906914711, + 0.08973474055528641, + 0.037965841591358185, + 0.03813503310084343, + 0.058632154017686844, + 0.07845371961593628, + 0.035313092172145844, + 0.04976101592183113, + 0.04822487756609917 + ], + "credit_map": [ + 0.047793883830308914, + 0.1097063198685646, + 0.05112004280090332, + 0.04368055239319801, + 0.1144370287656784, + 0.06277167052030563, + 0.047240905463695526, + 0.04902093484997749, + 0.0030814511701464653, + 0.0022674808278679848, + 0.005248854402452707, + 0.0042809732258319855, + 0.005490469746291637, + 0.005709032528102398, + 0.006369459442794323, + 0.0055604674853384495, + 0.006254844833165407, + 0.00462708156555891, + 0.00815336499363184, + 0.006529979407787323, + 0.004971486050635576, + 0.03337535262107849, + 0.0031069449614733458, + 0.02271568775177002, + 0.008386455476284027, + 0.005392069462686777, + 0.004638012498617172, + 0.005581910256296396, + 0.0035409138072282076, + 0.004621005151420832, + 0.002739585703238845, + 0.003065886441618204, + 0.007057589013129473, + 0.004883825778961182, + 0.004754053894430399, + 0.004108950030058622, + 0.005516895558685064, + 0.004087503999471664, + 0.004170455038547516, + 0.003555756527930498, + 0.006607000716030598, + 0.013151406310498714, + 0.007762146182358265, + 0.006620624102652073, + 0.004441641736775637, + 0.006582868751138449, + 0.007009825669229031, + 0.006456643808633089, + 0.00848536565899849, + 0.006614113692194223, + 0.009009718894958496, + 0.018722759559750557, + 0.007360868621617556, + 0.008517512120306492, + 0.011345550417900085, + 0.00839783251285553, + 0.0074417502619326115, + 0.003160507185384631, + 0.005078278481960297, + 0.0038357973098754883, + 0.0030426797457039356, + 0.005801528226584196, + 0.0026839873753488064, + 0.0042685638181865215, + 0.007478519808501005, + 0.008316743187606335, + 0.008204205892980099, + 0.0033026712480932474, + 0.0053541348315775394, + 0.006986852735280991, + 0.0036005915608257055, + 0.00651729479432106, + 0.017236003652215004, + 0.017879944294691086, + 0.0028056122828274965, + 0.002250110963359475, + 0.0024278932251036167, + 0.0014850664883852005, + 0.0020291749387979507, + 0.0021110721863806248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.12705688178539276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a modern, slim smartphone with a matte graphite-colored frame visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the smartphone resting on a smooth, lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, warm, rectangular beam of sunlight visible casting light onto the table and the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone rotate clockwise while drifting in gradual, unpredictable arcs across the wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static, high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369879350066185 + ], + "opposite_logprob_tokens": [ + -3.408698797225952 + ], + "logprob_reward": -0.03369879350066185, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369879350066185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37192216515541077, + 0.2097243219614029, + 0.13488291203975677, + 0.05406757816672325, + 0.053263161331415176, + 0.03191828355193138, + 0.05299583449959755, + 0.02924956940114498, + 0.0473957434296608, + 0.014580409973859787 + ], + "credit_map": [ + 0.03448549658060074, + 0.07071386277675629, + 0.037435322999954224, + 0.030138874426484108, + 0.0726848915219307, + 0.0466931089758873, + 0.03532177209854126, + 0.0444488525390625, + 0.017121970653533936, + 0.01517776120454073, + 0.031134745106101036, + 0.017641620710492134, + 0.015238326042890549, + 0.05062714219093323, + 0.020932478830218315, + 0.04185029864311218, + 0.009941739961504936, + 0.006230256520211697, + 0.019254373386502266, + 0.014248660765588284, + 0.010933966375887394, + 0.03910638391971588, + 0.007891455665230751, + 0.027276070788502693, + 0.013229791074991226, + 0.007399207446724176, + 0.008305867202579975, + 0.004910583607852459, + 0.004086153581738472, + 0.007464504800736904, + 0.0036183949559926987, + 0.005053075961768627, + 0.007518548518419266, + 0.005333204288035631, + 0.008212217129766941, + 0.00509312329813838, + 0.007751820143312216, + 0.004455392714589834, + 0.006301495712250471, + 0.008597359992563725, + 0.003589974483475089, + 0.007605795282870531, + 0.004506934434175491, + 0.003185815643519163, + 0.002379470271989703, + 0.004010199569165707, + 0.003982342313975096, + 0.002657749457284808, + 0.0070265308022499084, + 0.0035974085330963135, + 0.007098479196429253, + 0.010208938270807266, + 0.005844054743647575, + 0.008123211562633514, + 0.004580492153763771, + 0.0065167187713086605, + 0.0038748777005821466, + 0.00291444337926805, + 0.006949854549020529, + 0.0016241068951785564, + 0.0028365927282720804, + 0.0038937435019761324, + 0.0027701689396053553, + 0.004385780077427626, + 0.0070950849913060665, + 0.007303951773792505, + 0.004490336868911982, + 0.0027218919713050127, + 0.004545950330793858, + 0.011959667317569256, + 0.0032889440190047026, + 0.0059899138286709785, + 0.003943254239857197, + 0.0039823586121201515, + 0.0011156946420669556, + 0.0018305617850273848, + 0.0008542645955458283, + 0.0007589381420984864, + 0.00106980896089226, + 0.0010255282977595925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03369879350066185 + } + ], + "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/0000028/group_7/index_tensor([5718, 5718])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_tensor([5718, 5718])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8040ac8709ddd6c452a320c7e7dff946bbf072ba --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_tensor([5718, 5718])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Stunning high-speed Phantom camera recordings initiate a sweeping, low-angle tracking shot that tilts upward, relentlessly following a sleek, matte-black reconnaissance aircraft as it cuts through the upper atmosphere, climbing steeply toward the stratosphere. The aircraft's design is sharp and angular, its seamless fuselage and delta wings constructed from a non-reflective, radar-absorbent composite material that absorbs the harsh, unfiltered sunlight, creating a stark silhouette. From its powerful engines, afterburners emit an intense, fierce purple glow that pulses with controlled fury, casting a vibrant, otherworldly glow across the rear control surfaces. In the hyper-real clarity of the slow-motion capture, subtle, shimmering shockwaves become visible, rippled across the fuselage and wings like distortions of heat haze as they compress the thin air. The camera’s sweeping view reveals the vast, breathtaking curvature of the Earth below, with a majestic white cloud bank hovering over a deep blue ocean, all framed by thin, glowing blue threads of atmosphere that contrast starkly with the black, starless space, creating a breathtaking and powerful atmosphere of cutting-edge technology at the edge of the void.", + "seed": 11900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_5718_seed_11900_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": -0.04454972976818681, + "vlm_logprob": -0.04454972976818681, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04454972976818681, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09214398264884949 + ], + "opposite_logprob_tokens": [ + -2.467144012451172 + ], + "logprob_reward": -0.09214398264884949, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09214398264884949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037656642496585846, + 0.025056757032871246, + 0.020766200497746468, + 0.037595268338918686, + 0.2076052874326706, + 0.14405828714370728, + 0.36613285541534424, + 0.04232587665319443, + 0.06489800661802292, + 0.05390486121177673 + ], + "credit_map": [ + 0.004890955053269863, + 0.008224581368267536, + 0.003395383246243, + 0.0038536665961146355, + 0.002618276746943593, + 0.0024873956572264433, + 0.00707970978692174, + 0.005106669384986162, + 0.002989196917042136, + 0.006902496796101332, + 0.0022814886178821325, + 0.004523479379713535, + 0.0020559290423989296, + 0.0011308366665616632, + 0.0026737647131085396, + 0.0024995619896799326, + 0.0014480252284556627, + 0.0034941688645631075, + 0.0040972973220050335, + 0.003866820130497217, + 0.0013992113526910543, + 0.0014994723023846745, + 0.003010648535564542, + 0.001950555481016636, + 0.003195680445060134, + 0.00454956479370594, + 0.007856255397200584, + 0.006366827990859747, + 0.004169650375843048, + 0.003087321761995554, + 0.0036868038587272167, + 0.004683159291744232, + 0.019219087436795235, + 0.03881673514842987, + 0.027510447427630424, + 0.04353289306163788, + 0.01716856099665165, + 0.023074395954608917, + 0.02300318144261837, + 0.015279978513717651, + 0.012247583828866482, + 0.015760691836476326, + 0.02767797000706196, + 0.01625380478799343, + 0.013420488685369492, + 0.018225450068712234, + 0.017999576404690742, + 0.022472722455859184, + 0.03025784157216549, + 0.05722415819764137, + 0.06573942303657532, + 0.05189981311559677, + 0.03367897495627403, + 0.04696372523903847, + 0.04135168343782425, + 0.0390172116458416, + 0.006186075042933226, + 0.004456311464309692, + 0.0038977523799985647, + 0.005063827149569988, + 0.004974227864295244, + 0.005223143845796585, + 0.007061492186039686, + 0.005463043227791786, + 0.006714751478284597, + 0.007366021163761616, + 0.008037179708480835, + 0.009299036115407944, + 0.006979650352150202, + 0.009737360291182995, + 0.008165874518454075, + 0.00859813392162323, + 0.006742617581039667, + 0.005157794337719679, + 0.010796424001455307, + 0.015987064689397812, + 0.0030063446611166, + 0.002691194647923112, + 0.0062136962078511715, + 0.0033097201958298683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.09214398264884949 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301584646105766 + ], + "opposite_logprob_tokens": [ + -3.773301601409912 + ], + "logprob_reward": -0.023301584646105766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301584646105766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07681816071271896, + 0.04710673168301582, + 0.02533830888569355, + 0.06883223354816437, + 0.17825251817703247, + 0.05521165207028389, + 0.033323269337415695, + 0.3078104555606842, + 0.027475249022245407, + 0.17983144521713257 + ], + "credit_map": [ + 0.009670697152614594, + 0.015594908967614174, + 0.008560625836253166, + 0.009062832221388817, + 0.003920626826584339, + 0.006566491909325123, + 0.014765787869691849, + 0.008676184341311455, + 0.005811277776956558, + 0.01008113007992506, + 0.00799629371613264, + 0.004384984727948904, + 0.003621070645749569, + 0.002937617478892207, + 0.004535344894975424, + 0.007739011198282242, + 0.0025816543493419886, + 0.0034242779947817326, + 0.005706231575459242, + 0.003271738300099969, + 0.0017948723398149014, + 0.0020019456278532743, + 0.003875745926052332, + 0.002681838348507881, + 0.006991088390350342, + 0.007058601826429367, + 0.00613682996481657, + 0.013829069212079048, + 0.008543627336621284, + 0.007175950799137354, + 0.00851424690335989, + 0.010582812130451202, + 0.01419021375477314, + 0.03354668244719505, + 0.021229753270745277, + 0.027154002338647842, + 0.01698356121778488, + 0.022212231531739235, + 0.026411814615130424, + 0.016524245962500572, + 0.0053315432742238045, + 0.00618446059525013, + 0.006793193519115448, + 0.005610718857496977, + 0.00588582968339324, + 0.007438641041517258, + 0.006314330268651247, + 0.011652929708361626, + 0.003449666313827038, + 0.003447720780968666, + 0.005756756756454706, + 0.006106329150497913, + 0.002991722198203206, + 0.003801723476499319, + 0.004179010633379221, + 0.003590337233617902, + 0.05603417754173279, + 0.030963506549596786, + 0.06411571055650711, + 0.050423864275217056, + 0.016509028151631355, + 0.03681762143969536, + 0.03406457230448723, + 0.018881965428590775, + 0.0031766672618687153, + 0.00293273595161736, + 0.0038042492233216763, + 0.005439281929284334, + 0.0018578232266008854, + 0.003816179698333144, + 0.004269699100404978, + 0.0021786103025078773, + 0.020368285477161407, + 0.017612386494874954, + 0.034678518772125244, + 0.058326225727796555, + 0.009302554652094841, + 0.007599524222314358, + 0.021117378026247025, + 0.010826567187905312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.023301584646105766 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902125060558319 + ], + "opposite_logprob_tokens": [ + -2.4640212059020996 + ], + "logprob_reward": -0.08902125060558319, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902125060558319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06423266977071762, + 0.05810224637389183, + 0.05790548399090767, + 0.07540322840213776, + 0.1659257560968399, + 0.0852019190788269, + 0.08004111796617508, + 0.037227120250463486, + 0.2762896716594696, + 0.09967079013586044 + ], + "credit_map": [ + 0.0068398090079426765, + 0.009943811222910881, + 0.006890507414937019, + 0.008652541786432266, + 0.003486934117972851, + 0.005646430887281895, + 0.012457954697310925, + 0.010314678773283958, + 0.006596976425498724, + 0.009800064377486706, + 0.00555538572371006, + 0.014308596961200237, + 0.0034776984248310328, + 0.003218579338863492, + 0.008905274793505669, + 0.006239664740860462, + 0.004964379593729973, + 0.008967207744717598, + 0.009829680435359478, + 0.008590580895543098, + 0.0030864563304930925, + 0.0033594744745641947, + 0.013153347186744213, + 0.00595435593277216, + 0.005590495653450489, + 0.007628812454640865, + 0.00729901110753417, + 0.013741685077548027, + 0.010066133923828602, + 0.007101934868842363, + 0.009125559590756893, + 0.014849592000246048, + 0.01145744975656271, + 0.03015466220676899, + 0.01614270731806755, + 0.030898001044988632, + 0.013689237646758556, + 0.022003550082445145, + 0.02409520372748375, + 0.01748492941260338, + 0.008216881193220615, + 0.008972037583589554, + 0.012428642250597477, + 0.0086364159360528, + 0.00893501564860344, + 0.01097087375819683, + 0.010490632615983486, + 0.016551416367292404, + 0.00756939547136426, + 0.009364490397274494, + 0.012792440131306648, + 0.013303050771355629, + 0.006932358257472515, + 0.009370403364300728, + 0.00989560503512621, + 0.010813364759087563, + 0.005199348088353872, + 0.004219238180667162, + 0.005281213670969009, + 0.004453213885426521, + 0.0039019822143018246, + 0.004724529106169939, + 0.00579273235052824, + 0.0036548576317727566, + 0.027269193902611732, + 0.02746225893497467, + 0.04089609161019325, + 0.056162700057029724, + 0.020455334335565567, + 0.03982342779636383, + 0.03931678831577301, + 0.024903874844312668, + 0.011871135793626308, + 0.008747641928493977, + 0.01913726143538952, + 0.034405991435050964, + 0.004752317443490028, + 0.003417203202843666, + 0.010818904265761375, + 0.006520318798720837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.08902125060558319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419012900441885 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.01419012900441885, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419012900441885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0821714848279953, + 0.049627549946308136, + 0.05235057696700096, + 0.13130594789981842, + 0.07224984467029572, + 0.09070327877998352, + 0.25660768151283264, + 0.046747270971536636, + 0.11802522838115692, + 0.10021110624074936 + ], + "credit_map": [ + 0.01167611125856638, + 0.017852794378995895, + 0.006395298056304455, + 0.009579609148204327, + 0.005596049129962921, + 0.00581689178943634, + 0.014962437562644482, + 0.010292291641235352, + 0.005181432235985994, + 0.0066483463160693645, + 0.006180203054100275, + 0.007917172275483608, + 0.0028551765717566013, + 0.0037001511082053185, + 0.007927849888801575, + 0.009217221289873123, + 0.003163478570058942, + 0.0063849627040326595, + 0.010593791492283344, + 0.01086416095495224, + 0.004391686525195837, + 0.004564754199236631, + 0.007806703448295593, + 0.0045810420997440815, + 0.011215459555387497, + 0.012462801299989223, + 0.013712463900446892, + 0.027381794527173042, + 0.015920348465442657, + 0.01546754501760006, + 0.0176998320966959, + 0.017445717006921768, + 0.006548789329826832, + 0.008437980897724628, + 0.011436710134148598, + 0.0140983359888196, + 0.006959362421184778, + 0.006898119114339352, + 0.00969130638986826, + 0.008179237134754658, + 0.007736187893897295, + 0.010950075462460518, + 0.016792166978120804, + 0.010741110891103745, + 0.009106654673814774, + 0.011750757694244385, + 0.011093424633145332, + 0.01253290195018053, + 0.021295031532645226, + 0.02700822241604328, + 0.0468100868165493, + 0.048475295305252075, + 0.025681868195533752, + 0.031140673905611038, + 0.032984230667352676, + 0.02321227267384529, + 0.0065101985819637775, + 0.0073318397626280785, + 0.004676624201238155, + 0.006034000776708126, + 0.00538233295083046, + 0.005762050393968821, + 0.006863434799015522, + 0.004186791367828846, + 0.012631131336092949, + 0.020619535818696022, + 0.013505766168236732, + 0.011760206893086433, + 0.01421551126986742, + 0.01304597221314907, + 0.014851394109427929, + 0.017395704984664917, + 0.014725509099662304, + 0.01372987125068903, + 0.016607847064733505, + 0.016198808327317238, + 0.006594383157789707, + 0.009178788401186466, + 0.010776154696941376, + 0.01239974144846201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.01419012900441885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091701935976744 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091701935976744, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091701935976744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21682244539260864, + 0.06959498673677444, + 0.12723854184150696, + 0.03392178937792778, + 0.1363532841205597, + 0.06070365756750107, + 0.14759016036987305, + 0.06876573711633682, + 0.040604203939437866, + 0.09840518981218338 + ], + "credit_map": [ + 0.02285712957382202, + 0.021072877570986748, + 0.030569767579436302, + 0.03213385492563248, + 0.01184376422315836, + 0.02324533648788929, + 0.039508599787950516, + 0.03559111803770065, + 0.008807841688394547, + 0.014434191398322582, + 0.008222807198762894, + 0.01353362388908863, + 0.0040804436430335045, + 0.0040384018793702126, + 0.0097091319039464, + 0.006768546998500824, + 0.007493489421904087, + 0.015115992166101933, + 0.026492780074477196, + 0.02187742292881012, + 0.007773315999656916, + 0.009978250600397587, + 0.026884691789746284, + 0.011622602120041847, + 0.004339360166341066, + 0.0029511088505387306, + 0.0050242384895682335, + 0.005770185496658087, + 0.0032749029342085123, + 0.004159988835453987, + 0.0036553977988660336, + 0.004746603313833475, + 0.00957903079688549, + 0.025726420804858208, + 0.01393731776624918, + 0.024330271407961845, + 0.010634033940732479, + 0.01851990632712841, + 0.018866749480366707, + 0.014759538695216179, + 0.006372715812176466, + 0.0057634166441857815, + 0.012069405987858772, + 0.008928331546485424, + 0.005367903970181942, + 0.007946131750941277, + 0.006721124518662691, + 0.007534625940024853, + 0.013353678397834301, + 0.02270849049091339, + 0.021526813507080078, + 0.019262636080384254, + 0.013416205532848835, + 0.023696258664131165, + 0.016566962003707886, + 0.017059121280908585, + 0.011201671324670315, + 0.008577767759561539, + 0.00944302324205637, + 0.010040422901511192, + 0.004956544376909733, + 0.009433729574084282, + 0.008923937566578388, + 0.006188635714352131, + 0.005075251217931509, + 0.005489534232765436, + 0.005191907286643982, + 0.005553324706852436, + 0.0026794099248945713, + 0.005328205414116383, + 0.004826388321816921, + 0.006460186094045639, + 0.010028468444943428, + 0.009108140133321285, + 0.019130198284983635, + 0.03407640755176544, + 0.004810166545212269, + 0.0033514946699142456, + 0.010532399639487267, + 0.007367908954620361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.004091701935976744 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Stunning high-speed Phantom camera recordings initiate a sweeping, low-angle tracking shot that tilts upward, relentlessly following a sleek, matte-black reconnaissance aircraft as it cuts through the upper atmosphere, climbing steeply toward the stratosphere. The aircraft's design is sharp and angular, its seamless fuselage and delta wings constructed from a non-reflective, radar-absorbent composite material that absorbs the harsh, unfiltered sunlight, creating a stark silhouette. From its powerful engines, afterburners emit an intense, fierce purple glow that pulses with controlled fury, casting a vibrant, otherworldly glow across the rear control surfaces. In the hyper-real clarity of the slow-motion capture, subtle, shimmering shockwaves become visible, rippled across the fuselage and wings like distortions of heat haze as they compress the thin air. The camera’s sweeping view reveals the vast, breathtaking curvature of the Earth below, with a majestic white cloud bank hovering over a deep blue ocean, all framed by thin, glowing blue threads of atmosphere that contrast starkly with the black, starless space, creating a breathtaking and powerful atmosphere of cutting-edge technology at the edge of the void.", + "seed": 11901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_5718_seed_11901_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": -0.03528885748237372, + "vlm_logprob": -0.03528885748237372, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03528885748237372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033923059701919556 + ], + "opposite_logprob_tokens": [ + -3.5339231491088867 + ], + "logprob_reward": -0.033923059701919556, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.033923059701919556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2796934247016907, + 0.20392106473445892, + 0.031834185123443604, + 0.08267847448587418, + 0.06834876537322998, + 0.03408173844218254, + 0.05735188350081444, + 0.06448715180158615, + 0.09028415381908417, + 0.08731906116008759 + ], + "credit_map": [ + 0.026352431625127792, + 0.060037367045879364, + 0.023979729041457176, + 0.04516393318772316, + 0.02819743938744068, + 0.033765774220228195, + 0.017688054591417313, + 0.044508714228868484, + 0.03153373301029205, + 0.021120300516486168, + 0.0737416073679924, + 0.013225909322500229, + 0.019895130768418312, + 0.013094834052026272, + 0.007963551208376884, + 0.023346008732914925, + 0.00391168100759387, + 0.005872867535799742, + 0.0041927555575966835, + 0.00370731670409441, + 0.0052148811519145966, + 0.001854251022450626, + 0.0036682633217424154, + 0.003412168938666582, + 0.006561025511473417, + 0.018045762553811073, + 0.011336185038089752, + 0.013885012827813625, + 0.008840469643473625, + 0.0054603214375674725, + 0.010716836899518967, + 0.007832862436771393, + 0.007666006218641996, + 0.007586373016238213, + 0.007582349237054586, + 0.009733479470014572, + 0.006253680679947138, + 0.009903889149427414, + 0.008545813150703907, + 0.011077174916863441, + 0.004001585301011801, + 0.005152765661478043, + 0.0036020453553646803, + 0.0039550866931676865, + 0.007354036904871464, + 0.0047153630293905735, + 0.0024093433748930693, + 0.002891512820497155, + 0.006709782872349024, + 0.009704695083200932, + 0.0071548218838870525, + 0.012170996516942978, + 0.009182332083582878, + 0.004065821412950754, + 0.005088305100798607, + 0.0032751313410699368, + 0.007204247638583183, + 0.016905896365642548, + 0.0112961670383811, + 0.005912734195590019, + 0.006670480128377676, + 0.004698528908193111, + 0.006504708435386419, + 0.005294389091432095, + 0.007757141254842281, + 0.006343710236251354, + 0.012862615287303925, + 0.009748153388500214, + 0.006561171729117632, + 0.011769657954573631, + 0.020041868090629578, + 0.015199843794107437, + 0.01990886963903904, + 0.01409983355551958, + 0.022091902792453766, + 0.006571956444531679, + 0.005541305989027023, + 0.006894859950989485, + 0.006432449910789728, + 0.00577788520604372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.033923059701919556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206173077225685 + ], + "opposite_logprob_tokens": [ + -2.8120617866516113 + ], + "logprob_reward": -0.06206173077225685, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206173077225685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07771119475364685, + 0.4903159737586975, + 0.04866912215948105, + 0.06640350818634033, + 0.031655896455049515, + 0.05176984891295433, + 0.03874129429459572, + 0.01530552376061678, + 0.038480374962091446, + 0.14094726741313934 + ], + "credit_map": [ + 0.0076108588837087154, + 0.01113633532077074, + 0.009839381091296673, + 0.00982898473739624, + 0.009594229981303215, + 0.010718037374317646, + 0.007684130221605301, + 0.011299229227006435, + 0.06942160427570343, + 0.051887396723032, + 0.16608668863773346, + 0.04862102493643761, + 0.05289090797305107, + 0.025595946237444878, + 0.027521969750523567, + 0.04829040542244911, + 0.004951374139636755, + 0.006505692843347788, + 0.007545800413936377, + 0.007110008038580418, + 0.0060721831396222115, + 0.0024909169878810644, + 0.009486444294452667, + 0.004506698809564114, + 0.005391756072640419, + 0.009876029565930367, + 0.007528048008680344, + 0.011206605471670628, + 0.007945174351334572, + 0.00473520765081048, + 0.011849350295960903, + 0.00787133164703846, + 0.004270489327609539, + 0.003573833964765072, + 0.003372980048879981, + 0.005011740606278181, + 0.002831963822245598, + 0.004133676644414663, + 0.004214748274534941, + 0.004246459808200598, + 0.013948393985629082, + 0.012709840200841427, + 0.0034323642030358315, + 0.003056320594623685, + 0.006150849163532257, + 0.003714821534231305, + 0.004463801626116037, + 0.0042934538796544075, + 0.0027570289094001055, + 0.00982074998319149, + 0.0036333950702100992, + 0.00849062204360962, + 0.005453523248434067, + 0.002449513878673315, + 0.004303756169974804, + 0.0018327005673199892, + 0.0015267995186150074, + 0.0017139299307018518, + 0.0021626162342727184, + 0.0017143626464530826, + 0.002246855990961194, + 0.0013443164061754942, + 0.002715639304369688, + 0.0018810018664225936, + 0.0031510633416473866, + 0.0033197717275470495, + 0.005469661671668291, + 0.004418420605361462, + 0.0027811338659375906, + 0.005231072194874287, + 0.004320476669818163, + 0.009788770228624344, + 0.03327900916337967, + 0.020627735182642937, + 0.03929119557142258, + 0.010639145970344543, + 0.009651570580899715, + 0.007345558609813452, + 0.008297634311020374, + 0.011815405450761318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.06206173077225685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204122677445412 + ], + "opposite_logprob_tokens": [ + -2.8120412826538086 + ], + "logprob_reward": -0.06204122677445412, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204122677445412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10363101959228516, + 0.07669740915298462, + 0.0671549066901207, + 0.062364183366298676, + 0.08531571924686432, + 0.11330810934305191, + 0.1292969137430191, + 0.0941973552107811, + 0.06154518574476242, + 0.2064892053604126 + ], + "credit_map": [ + 0.0081165861338377, + 0.013837849721312523, + 0.010070402175188065, + 0.01681496761739254, + 0.010095654986798763, + 0.01907176338136196, + 0.009073317982256413, + 0.016550488770008087, + 0.008253054693341255, + 0.008807445876300335, + 0.01881372183561325, + 0.01078745722770691, + 0.009902150370180607, + 0.005633853375911713, + 0.006480623967945576, + 0.008019108325242996, + 0.009229704737663269, + 0.014541695825755596, + 0.010461349971592426, + 0.0075737545266747475, + 0.010156256146728992, + 0.003307042410597205, + 0.0057655018754303455, + 0.006119605153799057, + 0.00663718581199646, + 0.01628691330552101, + 0.008567607961595058, + 0.007181680295616388, + 0.008692571893334389, + 0.004731276072561741, + 0.00551076652482152, + 0.004756186623126268, + 0.01039097923785448, + 0.016127223148941994, + 0.0077453916892409325, + 0.010518427938222885, + 0.00953760463744402, + 0.018373019993305206, + 0.005989226512610912, + 0.0066338591277599335, + 0.03077266924083233, + 0.03903762251138687, + 0.006517417263239622, + 0.005822482984513044, + 0.013408155180513859, + 0.007418559864163399, + 0.004725817125290632, + 0.0056054070591926575, + 0.015776364132761955, + 0.035057175904512405, + 0.015442119911313057, + 0.021276993677020073, + 0.0165697168558836, + 0.009903840720653534, + 0.00827785488218069, + 0.006992855109274387, + 0.011559351347386837, + 0.025914406403899193, + 0.019423477351665497, + 0.009765582159161568, + 0.009780977852642536, + 0.005961619783192873, + 0.005486103240400553, + 0.0063058459199965, + 0.00658960547298193, + 0.008372576907277107, + 0.01036969292908907, + 0.008006568066775799, + 0.005853712093085051, + 0.006013910751789808, + 0.008772631175816059, + 0.0075664883479475975, + 0.04995185509324074, + 0.03269005939364433, + 0.059938300400972366, + 0.01769240014255047, + 0.013788362964987755, + 0.011366061866283417, + 0.008865347132086754, + 0.012196828611195087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.06204122677445412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10323754698038101, + 0.052553873509168625, + 0.08258993178606033, + 0.1547657996416092, + 0.08103688806295395, + 0.13971231877803802, + 0.09579075872898102, + 0.12343459576368332, + 0.040522947907447815, + 0.12635532021522522 + ], + "credit_map": [ + 0.0073360358364880085, + 0.013182483613491058, + 0.009903786703944206, + 0.014334111474454403, + 0.011713740415871143, + 0.018903713673353195, + 0.01276329718530178, + 0.01510037574917078, + 0.00414844648912549, + 0.007149158976972103, + 0.0077075292356312275, + 0.006449171341955662, + 0.010384557768702507, + 0.0036239824257791042, + 0.007076962385326624, + 0.006014064885675907, + 0.011495732702314854, + 0.014656712301075459, + 0.011724913492798805, + 0.009336180984973907, + 0.010051263496279716, + 0.006856225896626711, + 0.010836871340870857, + 0.007632033433765173, + 0.014562815427780151, + 0.037281956523656845, + 0.019897760823369026, + 0.017340345308184624, + 0.017703484743833542, + 0.011351793073117733, + 0.020334536209702492, + 0.016293112188577652, + 0.009736443869769573, + 0.011129921302199364, + 0.007577868178486824, + 0.007884659804403782, + 0.008591193705797195, + 0.02126038447022438, + 0.0066205658949911594, + 0.008235851302742958, + 0.034649286419153214, + 0.04891689494252205, + 0.00753020541742444, + 0.007618235424160957, + 0.016910120844841003, + 0.008768646977841854, + 0.006851836107671261, + 0.008467099629342556, + 0.007890667766332626, + 0.022043898701667786, + 0.01140602957457304, + 0.0256489310413599, + 0.010003600269556046, + 0.005754362791776657, + 0.009116997011005878, + 0.0039262715727090836, + 0.014543901197612286, + 0.018942901864647865, + 0.023890681564807892, + 0.019480889663100243, + 0.016375185921788216, + 0.007747791241854429, + 0.0118098808452487, + 0.01064335834234953, + 0.0049450271762907505, + 0.008682521060109138, + 0.004910249263048172, + 0.005243405699729919, + 0.004683199804276228, + 0.004866552073508501, + 0.003753479104489088, + 0.0034385104663670063, + 0.021018091589212418, + 0.016721518710255623, + 0.025376999750733376, + 0.01300638448446989, + 0.011042927391827106, + 0.01273070927709341, + 0.015722474083304405, + 0.010736219584941864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.008639931678771973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778338484466076 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778338484466076, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778338484466076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11666467040777206, + 0.08643796294927597, + 0.01868947222828865, + 0.036669645458459854, + 0.10674995183944702, + 0.1488730013370514, + 0.08620341867208481, + 0.22485199570655823, + 0.07529821991920471, + 0.09956157207489014 + ], + "credit_map": [ + 0.013198758475482464, + 0.020692534744739532, + 0.011437950655817986, + 0.014698049984872341, + 0.01525294128805399, + 0.01767350733280182, + 0.006746454630047083, + 0.016964484006166458, + 0.009726013988256454, + 0.009389525279402733, + 0.027137110009789467, + 0.006463357247412205, + 0.009710067883133888, + 0.005826723761856556, + 0.004737749230116606, + 0.01344742439687252, + 0.002867310307919979, + 0.0033087602350860834, + 0.002783851930871606, + 0.0023444502148777246, + 0.0027850763872265816, + 0.001157687627710402, + 0.001379281049594283, + 0.0020630578510463238, + 0.003988165408372879, + 0.008088518865406513, + 0.004507144913077354, + 0.004663859028369188, + 0.004220137372612953, + 0.002657967619597912, + 0.0046373168006539345, + 0.003906541038304567, + 0.02093036286532879, + 0.010790460743010044, + 0.013434515334665775, + 0.011250675655901432, + 0.012784766964614391, + 0.018699729815125465, + 0.008814401924610138, + 0.010045050643384457, + 0.036996904760599136, + 0.030589811503887177, + 0.015241703949868679, + 0.009477191604673862, + 0.022443657740950584, + 0.015030731447041035, + 0.009197939187288284, + 0.009895073249936104, + 0.0065410928800702095, + 0.01496698148548603, + 0.010537785477936268, + 0.01126112975180149, + 0.02294738031923771, + 0.0069899060763418674, + 0.007829438894987106, + 0.00512971356511116, + 0.02834242582321167, + 0.040379464626312256, + 0.032207783311605453, + 0.03002787008881569, + 0.03988311067223549, + 0.016454948112368584, + 0.01706562750041485, + 0.020490780472755432, + 0.007865355350077152, + 0.012914935126900673, + 0.016350889578461647, + 0.013257316313683987, + 0.008309850469231606, + 0.006811425089836121, + 0.004949077032506466, + 0.004839375149458647, + 0.018471889197826385, + 0.02155686914920807, + 0.022825870662927628, + 0.009782487526535988, + 0.006979998666793108, + 0.006705010309815407, + 0.0069053941406309605, + 0.006334059871733189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.009778338484466076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Stunning high-speed Phantom camera recordings initiate a sweeping, low-angle tracking shot that tilts upward, relentlessly following a sleek, matte-black reconnaissance aircraft as it cuts through the upper atmosphere, climbing steeply toward the stratosphere. The aircraft's design is sharp and angular, its seamless fuselage and delta wings constructed from a non-reflective, radar-absorbent composite material that absorbs the harsh, unfiltered sunlight, creating a stark silhouette. From its powerful engines, afterburners emit an intense, fierce purple glow that pulses with controlled fury, casting a vibrant, otherworldly glow across the rear control surfaces. In the hyper-real clarity of the slow-motion capture, subtle, shimmering shockwaves become visible, rippled across the fuselage and wings like distortions of heat haze as they compress the thin air. The camera’s sweeping view reveals the vast, breathtaking curvature of the Earth below, with a majestic white cloud bank hovering over a deep blue ocean, all framed by thin, glowing blue threads of atmosphere that contrast starkly with the black, starless space, creating a breathtaking and powerful atmosphere of cutting-edge technology at the edge of the void.", + "seed": 11902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_5718_seed_11902_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": -0.03610343309119344, + "vlm_logprob": -0.03610343309119344, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03610343309119344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07284021377563477 + ], + "opposite_logprob_tokens": [ + -2.6978402137756348 + ], + "logprob_reward": -0.07284021377563477, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07284021377563477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13595956563949585, + 0.08435878902673721, + 0.0745668038725853, + 0.033043112605810165, + 0.09738752990961075, + 0.31592264771461487, + 0.046547211706638336, + 0.11390996724367142, + 0.043603334575891495, + 0.05470108240842819 + ], + "credit_map": [ + 0.011593738570809364, + 0.00861295871436596, + 0.018572919070720673, + 0.019170716404914856, + 0.011932839639484882, + 0.02232799492776394, + 0.027890769764780998, + 0.01585761271417141, + 0.00561674265190959, + 0.004038278013467789, + 0.008613943122327328, + 0.012838225811719894, + 0.009796258993446827, + 0.012982473708689213, + 0.016939213499426842, + 0.013533644378185272, + 0.0037014633417129517, + 0.004414950963109732, + 0.012049148790538311, + 0.009194567799568176, + 0.008831229992210865, + 0.008759215474128723, + 0.017411895096302032, + 0.01020432822406292, + 0.002090359339490533, + 0.0025463325437158346, + 0.006206247955560684, + 0.004850334487855434, + 0.0036382016260176897, + 0.00292455917224288, + 0.0047433581203222275, + 0.0060437205247581005, + 0.005389906000345945, + 0.005468728486448526, + 0.015304584056138992, + 0.013651685789227486, + 0.007839776575565338, + 0.013914690352976322, + 0.015248010866343975, + 0.02057013474404812, + 0.035098522901535034, + 0.023256130516529083, + 0.07531274110078812, + 0.028912488371133804, + 0.016863858327269554, + 0.017450863495469093, + 0.02535427361726761, + 0.09367376565933228, + 0.005072626285254955, + 0.004003426991403103, + 0.006023844704031944, + 0.008149944245815277, + 0.007343676872551441, + 0.00615282217040658, + 0.004443799611181021, + 0.005357065703719854, + 0.01728649064898491, + 0.018789304420351982, + 0.01151293609291315, + 0.02042938955128193, + 0.013625560328364372, + 0.011881094425916672, + 0.009255259297788143, + 0.011129927821457386, + 0.005291997455060482, + 0.005895684473216534, + 0.005077971611171961, + 0.008500606752932072, + 0.006369180046021938, + 0.003677359316498041, + 0.003704289672896266, + 0.005086240824311972, + 0.008095850236713886, + 0.010145565494894981, + 0.006534272339195013, + 0.0032117303926497698, + 0.009696162305772305, + 0.008250880986452103, + 0.004346990492194891, + 0.004419631790369749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.07284021377563477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646339945495129 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646339945495129, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646339945495129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08704280853271484, + 0.06809373944997787, + 0.07924345135688782, + 0.080318383872509, + 0.056301314383745193, + 0.06017434597015381, + 0.10502971708774567, + 0.34145116806030273, + 0.06701768934726715, + 0.055327389389276505 + ], + "credit_map": [ + 0.007497943937778473, + 0.004347353242337704, + 0.011786531656980515, + 0.010189528577029705, + 0.01152814831584692, + 0.01326722651720047, + 0.017332958057522774, + 0.011093118228018284, + 0.00415703235194087, + 0.005288328509777784, + 0.00886316504329443, + 0.010452471673488617, + 0.007851239293813705, + 0.009354371577501297, + 0.01257332693785429, + 0.009553804993629456, + 0.0048640999011695385, + 0.0035016692709177732, + 0.012670119293034077, + 0.008924931287765503, + 0.00904321949928999, + 0.008646382950246334, + 0.018671918660402298, + 0.012921109795570374, + 0.004681450314819813, + 0.0064093912951648235, + 0.014826417900621891, + 0.011443776078522205, + 0.008609693497419357, + 0.007181473076343536, + 0.011924104765057564, + 0.0152420774102211, + 0.003968782257288694, + 0.005014942027628422, + 0.008691693656146526, + 0.009664018638432026, + 0.004822419490665197, + 0.0049917614087462425, + 0.010429909452795982, + 0.008717789314687252, + 0.008326116017997265, + 0.005362401716411114, + 0.009251495823264122, + 0.009241463616490364, + 0.003043246688321233, + 0.0042910282500088215, + 0.007025488186627626, + 0.013633104041218758, + 0.013761979527771473, + 0.012006527744233608, + 0.008710740134119987, + 0.012960565276443958, + 0.01955128274857998, + 0.01891680434346199, + 0.008796525187790394, + 0.010325300507247448, + 0.0629839152097702, + 0.06974691897630692, + 0.03141447901725769, + 0.05666756629943848, + 0.03309761732816696, + 0.03132332116365433, + 0.026880983263254166, + 0.02933638170361519, + 0.008471029810607433, + 0.007496223319321871, + 0.0068601141683757305, + 0.011201468296349049, + 0.011817279271781445, + 0.008477858267724514, + 0.006415968760848045, + 0.006277751177549362, + 0.0056797233410179615, + 0.008067558519542217, + 0.010185511782765388, + 0.006788043305277824, + 0.007421438116580248, + 0.00670234952121973, + 0.005745421629399061, + 0.004737345036119223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.007646339945495129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001148164272308 + ], + "opposite_logprob_tokens": [ + -2.6950113773345947 + ], + "logprob_reward": -0.07001148164272308, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001148164272308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10332117974758148, + 0.0792158991098404, + 0.05968216434121132, + 0.04133499413728714, + 0.04627180099487305, + 0.14320692420005798, + 0.12980589270591736, + 0.18662354350090027, + 0.0956069603562355, + 0.11493068933486938 + ], + "credit_map": [ + 0.004358417354524136, + 0.0062638940289616585, + 0.013698841445147991, + 0.012892255559563637, + 0.008948223665356636, + 0.02654198370873928, + 0.019955964758992195, + 0.010661597363650799, + 0.006075331941246986, + 0.004224346484988928, + 0.008677063509821892, + 0.009642764925956726, + 0.008933822624385357, + 0.014800043776631355, + 0.01690615899860859, + 0.0099563617259264, + 0.003411938203498721, + 0.002968159969896078, + 0.011661460623145103, + 0.008502824231982231, + 0.0061461166478693485, + 0.00747588649392128, + 0.011725779622793198, + 0.00778999924659729, + 0.0021232834551483393, + 0.002482781885191798, + 0.007287217304110527, + 0.006548316217958927, + 0.004951107781380415, + 0.0034449114464223385, + 0.007286570034921169, + 0.0072108060121536255, + 0.0036869258619844913, + 0.0037776653189212084, + 0.007817111909389496, + 0.007854413241147995, + 0.004527289420366287, + 0.004862668924033642, + 0.005827350541949272, + 0.00791836902499199, + 0.008641411550343037, + 0.008031132631003857, + 0.03775960952043533, + 0.02141484059393406, + 0.007358560338616371, + 0.008486331440508366, + 0.011935724876821041, + 0.03957930579781532, + 0.015841329470276833, + 0.015092807821929455, + 0.01069654244929552, + 0.01713656447827816, + 0.022368302568793297, + 0.023264655843377113, + 0.012969656847417355, + 0.012436025775969028, + 0.02297007106244564, + 0.028314605355262756, + 0.014021248556673527, + 0.022487791255116463, + 0.048009660094976425, + 0.01810923032462597, + 0.015753906220197678, + 0.01695702038705349, + 0.015558188781142235, + 0.012196864932775497, + 0.011540224775671959, + 0.015134330838918686, + 0.013862358406186104, + 0.010634372010827065, + 0.00836341455578804, + 0.008317200466990471, + 0.015519215725362301, + 0.020367154851555824, + 0.01156051829457283, + 0.010653145611286163, + 0.02217693254351616, + 0.017732983455061913, + 0.0072276233695447445, + 0.009693101979792118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.07001148164272308 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328341454267502 + ], + "opposite_logprob_tokens": [ + -3.7732834815979004 + ], + "logprob_reward": -0.02328341454267502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328341454267502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18460246920585632, + 0.24457712471485138, + 0.029672643169760704, + 0.03284678980708122, + 0.07854711264371872, + 0.10367495566606522, + 0.14928916096687317, + 0.08834578841924667, + 0.03382005915045738, + 0.05462395399808884 + ], + "credit_map": [ + 0.0137433847412467, + 0.011784118600189686, + 0.02667444944381714, + 0.030077239498496056, + 0.021170955151319504, + 0.02107366733253002, + 0.038388706743717194, + 0.021689923480153084, + 0.017840338870882988, + 0.021887430921196938, + 0.03213675320148468, + 0.04401380568742752, + 0.02666189707815647, + 0.031759459525346756, + 0.03712696582078934, + 0.03315046429634094, + 0.001516271848231554, + 0.0017464060802012682, + 0.004445062484592199, + 0.003966085612773895, + 0.0024819548707455397, + 0.003232267452403903, + 0.007875721901655197, + 0.00440886989235878, + 0.0019568249117583036, + 0.002169405808672309, + 0.005830591544508934, + 0.005519385449588299, + 0.003813317744061351, + 0.0027180416509509087, + 0.004969116300344467, + 0.00587010383605957, + 0.004456635098904371, + 0.004795493092387915, + 0.01182705257087946, + 0.014065251685678959, + 0.006956293247640133, + 0.007904600352048874, + 0.01626378297805786, + 0.012277992442250252, + 0.01107458770275116, + 0.00750291021540761, + 0.014256066642701626, + 0.019760092720389366, + 0.011505486443638802, + 0.010674973949790001, + 0.015145867131650448, + 0.013754959218204021, + 0.014776124618947506, + 0.01589040644466877, + 0.0170682892203331, + 0.02804999239742756, + 0.030648652464151382, + 0.013188048265874386, + 0.014492875896394253, + 0.015174747444689274, + 0.01693063974380493, + 0.0169567558914423, + 0.008289236575365067, + 0.01419632975012064, + 0.008137581869959831, + 0.008082810789346695, + 0.007840338163077831, + 0.007912084460258484, + 0.00396390026435256, + 0.004519281443208456, + 0.0039536613039672375, + 0.006318243220448494, + 0.0050987317226827145, + 0.003103793365880847, + 0.003097980283200741, + 0.003764465218409896, + 0.007784300483763218, + 0.006729183252900839, + 0.008074472658336163, + 0.0058212559670209885, + 0.00862597580999136, + 0.00681281229481101, + 0.005085707642138004, + 0.005690242163836956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.02328341454267502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10172678530216217, + 0.3025963306427002, + 0.045308224856853485, + 0.06918936967849731, + 0.05380988121032715, + 0.16572260856628418, + 0.033329084515571594, + 0.13259153068065643, + 0.04325312748551369, + 0.0524730421602726 + ], + "credit_map": [ + 0.004291458986699581, + 0.00761211384087801, + 0.01744087040424347, + 0.014668574556708336, + 0.011971687898039818, + 0.016083983704447746, + 0.02014017663896084, + 0.009517919272184372, + 0.02041778154671192, + 0.026547899469733238, + 0.045534856617450714, + 0.054845180362463, + 0.02878751792013645, + 0.04643849655985832, + 0.04293547943234444, + 0.03708909451961517, + 0.0031867455691099167, + 0.0030934757087379694, + 0.009583881124854088, + 0.0049904282204806805, + 0.004810753744095564, + 0.006023271009325981, + 0.00913624931126833, + 0.0044834204018116, + 0.005337091162800789, + 0.008095714263617992, + 0.012940535321831703, + 0.010019644163548946, + 0.009105819277465343, + 0.005860425066202879, + 0.008718758821487427, + 0.009111379273235798, + 0.004160087555646896, + 0.004775165114551783, + 0.009248214773833752, + 0.006284809205681086, + 0.0036395497154444456, + 0.0050124069675803185, + 0.008214045315980911, + 0.012475601397454739, + 0.02278127521276474, + 0.015860553830862045, + 0.03806246444582939, + 0.019405262544751167, + 0.007507406175136566, + 0.007721510715782642, + 0.008690956979990005, + 0.04569317400455475, + 0.0033727576956152916, + 0.0038403666112571955, + 0.0031121307983994484, + 0.005416914355009794, + 0.0051265014335513115, + 0.004930870141834021, + 0.0034354389645159245, + 0.004094103816896677, + 0.019438089802861214, + 0.020757567137479782, + 0.010121425613760948, + 0.019748982042074203, + 0.022956879809498787, + 0.012941699475049973, + 0.01117295864969492, + 0.015453933738172054, + 0.0058021037839353085, + 0.005599319934844971, + 0.005576937459409237, + 0.008345766924321651, + 0.006011052522808313, + 0.004087465815246105, + 0.0030222071800380945, + 0.004808275029063225, + 0.005455791484564543, + 0.0065657347440719604, + 0.00903086457401514, + 0.004629182629287243, + 0.007769349496811628, + 0.008626928552985191, + 0.005033405032008886, + 0.005361790303140879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.006735715549439192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Stunning high-speed Phantom camera recordings initiate a sweeping, low-angle tracking shot that tilts upward, relentlessly following a sleek, matte-black reconnaissance aircraft as it cuts through the upper atmosphere, climbing steeply toward the stratosphere. The aircraft's design is sharp and angular, its seamless fuselage and delta wings constructed from a non-reflective, radar-absorbent composite material that absorbs the harsh, unfiltered sunlight, creating a stark silhouette. From its powerful engines, afterburners emit an intense, fierce purple glow that pulses with controlled fury, casting a vibrant, otherworldly glow across the rear control surfaces. In the hyper-real clarity of the slow-motion capture, subtle, shimmering shockwaves become visible, rippled across the fuselage and wings like distortions of heat haze as they compress the thin air. The camera’s sweeping view reveals the vast, breathtaking curvature of the Earth below, with a majestic white cloud bank hovering over a deep blue ocean, all framed by thin, glowing blue threads of atmosphere that contrast starkly with the black, starless space, creating a breathtaking and powerful atmosphere of cutting-edge technology at the edge of the void.", + "seed": 11903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_5718_seed_11903_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": -0.057767427247017625, + "vlm_logprob": -0.057767427247017625, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.057767427247017625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18256138265132904 + ], + "opposite_logprob_tokens": [ + -1.8075613975524902 + ], + "logprob_reward": -0.18256138265132904, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18256138265132904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1230800524353981, + 0.05567459017038345, + 0.05939637869596481, + 0.052552033215761185, + 0.0773339793086052, + 0.15501923859119415, + 0.10041557997465134, + 0.0714636743068695, + 0.12639805674552917, + 0.17866642773151398 + ], + "credit_map": [ + 0.013736434280872345, + 0.017038637772202492, + 0.0161899421364069, + 0.01460269745439291, + 0.023990411311388016, + 0.01635921001434326, + 0.012356539256870747, + 0.008806176483631134, + 0.00405900040641427, + 0.005866069812327623, + 0.0051014176569879055, + 0.005612798035144806, + 0.009982659481465816, + 0.010871002450585365, + 0.007462197449058294, + 0.0067194439470767975, + 0.005539715755730867, + 0.005755248479545116, + 0.007757971994578838, + 0.007621983997523785, + 0.0072936685755848885, + 0.006889178417623043, + 0.008743217214941978, + 0.009795395657420158, + 0.009331757202744484, + 0.005375795066356659, + 0.00398253882303834, + 0.0034409803338348866, + 0.007631019223481417, + 0.013045738451182842, + 0.004417106043547392, + 0.0053271036595106125, + 0.008314567618072033, + 0.0063791051506996155, + 0.006828525569289923, + 0.018254779279232025, + 0.006307738833129406, + 0.011055545881390572, + 0.01093022245913744, + 0.009263493120670319, + 0.041933391243219376, + 0.045032862573862076, + 0.006280338857322931, + 0.008632197976112366, + 0.010788297280669212, + 0.011422963812947273, + 0.01628466695547104, + 0.014644516631960869, + 0.011598738841712475, + 0.013644421473145485, + 0.008053767494857311, + 0.011924326419830322, + 0.009864282794296741, + 0.01984689012169838, + 0.011962726712226868, + 0.013520422391593456, + 0.006469612941145897, + 0.009778821840882301, + 0.006207144353538752, + 0.005956295412033796, + 0.014368844218552113, + 0.012325653806328773, + 0.0075674159452319145, + 0.008789888583123684, + 0.009432620368897915, + 0.024824973195791245, + 0.025595100596547127, + 0.008573167026042938, + 0.02518404647707939, + 0.007943356409668922, + 0.01675632782280445, + 0.008088473230600357, + 0.012643400579690933, + 0.022760407999157906, + 0.023515714332461357, + 0.04952555149793625, + 0.019019844010472298, + 0.022818123921751976, + 0.0203578881919384, + 0.008025490678846836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.18256138265132904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016088545322418213 + ], + "opposite_logprob_tokens": [ + -4.141088485717773 + ], + "logprob_reward": -0.016088545322418213, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016088545322418213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15456973016262054, + 0.09581145644187927, + 0.23508375883102417, + 0.06886398792266846, + 0.021599644795060158, + 0.018164608627557755, + 0.026830976828932762, + 0.030311880633234978, + 0.12477268278598785, + 0.22399123013019562 + ], + "credit_map": [ + 0.010748756118118763, + 0.0196733046323061, + 0.017400307580828667, + 0.02132590487599373, + 0.031220661476254463, + 0.02469557523727417, + 0.017527854070067406, + 0.011977355927228928, + 0.005956895183771849, + 0.007762185297906399, + 0.009677829220890999, + 0.012079460546374321, + 0.012186698615550995, + 0.019517092034220695, + 0.016846677288413048, + 0.011784610338509083, + 0.015902429819107056, + 0.013388185761868954, + 0.02370094507932663, + 0.01937960647046566, + 0.02364959940314293, + 0.03404819965362549, + 0.07640203833580017, + 0.028612760826945305, + 0.010648416355252266, + 0.0070488634519279, + 0.0038943428080528975, + 0.0031186919659376144, + 0.010969494469463825, + 0.015088248997926712, + 0.011235533282160759, + 0.006860391702502966, + 0.0027074627578258514, + 0.002343045547604561, + 0.001974948449060321, + 0.00304634147323668, + 0.0017858182545751333, + 0.0022163130342960358, + 0.004678915720432997, + 0.002846800023689866, + 0.00223755044862628, + 0.0019085360690951347, + 0.0014164467575028539, + 0.0016890402184799314, + 0.0015889831120148301, + 0.0030520856380462646, + 0.003439077641814947, + 0.0028328874614089727, + 0.0025901938788592815, + 0.003433993784710765, + 0.002367249922826886, + 0.0036089399363845587, + 0.0022929662372916937, + 0.005345289129763842, + 0.002812743652611971, + 0.00437959935516119, + 0.003918239381164312, + 0.006466906983405352, + 0.00283215194940567, + 0.002452690387144685, + 0.006589986849576235, + 0.0032539060339331627, + 0.0025726533494889736, + 0.0022253443021327257, + 0.014385664835572243, + 0.027703622356057167, + 0.017143825069069862, + 0.0074423146434128284, + 0.024075869470834732, + 0.009945373982191086, + 0.016029490157961845, + 0.008046526461839676, + 0.01623713970184326, + 0.031226759776473045, + 0.0300351120531559, + 0.05484699085354805, + 0.025763511657714844, + 0.027081673964858055, + 0.02911117859184742, + 0.00968885701149702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.016088545322418213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002303749322891 + ], + "opposite_logprob_tokens": [ + -2.6950230598449707 + ], + "logprob_reward": -0.07002303749322891, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002303749322891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12102971971035004, + 0.09924749284982681, + 0.11091014742851257, + 0.09511607140302658, + 0.07565323263406754, + 0.035903338342905045, + 0.058045677840709686, + 0.09215077012777328, + 0.1506279557943344, + 0.16131560504436493 + ], + "credit_map": [ + 0.009227867238223553, + 0.015292109921574593, + 0.016623003408312798, + 0.012859314680099487, + 0.023798637092113495, + 0.01943289116024971, + 0.012528092600405216, + 0.011267797090113163, + 0.0076578701846301556, + 0.010627856478095055, + 0.008249878883361816, + 0.01295130979269743, + 0.014948028139770031, + 0.02572804130613804, + 0.011717463843524456, + 0.007367035374045372, + 0.009564959444105625, + 0.009501289576292038, + 0.01121813990175724, + 0.008495393209159374, + 0.01616315171122551, + 0.014495966024696827, + 0.02226182445883751, + 0.019209416583180428, + 0.014945334754884243, + 0.00935427751392126, + 0.006018063053488731, + 0.005906036123633385, + 0.017649222165346146, + 0.023007286712527275, + 0.010779805481433868, + 0.007456043269485235, + 0.010404794476926327, + 0.00926203466951847, + 0.009327730163931847, + 0.009689558297395706, + 0.00647976016625762, + 0.00895653385668993, + 0.0118750911206007, + 0.00965772196650505, + 0.006022122222930193, + 0.005793497897684574, + 0.0026815973687916994, + 0.0040858229622244835, + 0.0029219603165984154, + 0.002892067190259695, + 0.006485629826784134, + 0.005020636133849621, + 0.00664198724552989, + 0.00915681105107069, + 0.005442099180072546, + 0.005931082181632519, + 0.005567834246903658, + 0.011987224221229553, + 0.005303357727825642, + 0.008015278726816177, + 0.00915372371673584, + 0.014267444610595703, + 0.0108824223279953, + 0.00850498117506504, + 0.021106263622641563, + 0.008952206932008266, + 0.009014900773763657, + 0.010268822312355042, + 0.013792975805699825, + 0.03983719274401665, + 0.04133453965187073, + 0.008076316677033901, + 0.020047049969434738, + 0.00842915941029787, + 0.01162857934832573, + 0.007482137531042099, + 0.011954881250858307, + 0.025590071454644203, + 0.021825842559337616, + 0.03058624640107155, + 0.019376980140805244, + 0.020742828026413918, + 0.02328953891992569, + 0.00794919952750206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.07002303749322891 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538178823888302 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538178823888302, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538178823888302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12427513301372528, + 0.08283324539661407, + 0.05083639919757843, + 0.09477652609348297, + 0.08843744546175003, + 0.06597093492746353, + 0.09368407726287842, + 0.054977480322122574, + 0.25880223512649536, + 0.08540653437376022 + ], + "credit_map": [ + 0.009001466445624828, + 0.022302739322185516, + 0.018521999940276146, + 0.017694663256406784, + 0.01870262436568737, + 0.01544178742915392, + 0.011854865588247776, + 0.01075498666614294, + 0.005441696383059025, + 0.004907731432467699, + 0.00813574530184269, + 0.009945944882929325, + 0.01246615033596754, + 0.014438756741583347, + 0.016212312504649162, + 0.011284912936389446, + 0.004644597414880991, + 0.004635813646018505, + 0.005473271943628788, + 0.006291740573942661, + 0.007652370724827051, + 0.006724882405251265, + 0.008447416126728058, + 0.006966307759284973, + 0.015225371345877647, + 0.009075854904949665, + 0.007187594193965197, + 0.005999843589961529, + 0.012863317504525185, + 0.02199045941233635, + 0.010315668769180775, + 0.012118421494960785, + 0.012261921539902687, + 0.010568322613835335, + 0.006427316460758448, + 0.011250772513449192, + 0.007685177028179169, + 0.01136965211480856, + 0.014607290737330914, + 0.014266988262534142, + 0.007834197953343391, + 0.005882325116544962, + 0.005819719284772873, + 0.007793103810399771, + 0.0064774262718856335, + 0.013955879956483841, + 0.010249597020447254, + 0.007958685047924519, + 0.012115886434912682, + 0.015041711740195751, + 0.008621251210570335, + 0.009957433678209782, + 0.01150945108383894, + 0.010825119912624359, + 0.012673594988882542, + 0.012939627282321453, + 0.005012785084545612, + 0.007977217435836792, + 0.006118291057646275, + 0.004385316278785467, + 0.010006384924054146, + 0.0067193773575127125, + 0.007548387162387371, + 0.007209717761725187, + 0.026151517406105995, + 0.047627564519643784, + 0.061828237026929855, + 0.013082887046039104, + 0.05135476589202881, + 0.01604449935257435, + 0.027170056477189064, + 0.015542698092758656, + 0.007232919335365295, + 0.012140195816755295, + 0.010918902233242989, + 0.016795942559838295, + 0.012555917724967003, + 0.00904124416410923, + 0.012401419691741467, + 0.004319989588111639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.012538178823888302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, matte-black reconnaissance aircraft with sharp, angular design and delta wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the aircraft's engines emit an intense, fierce purple glow from the afterburners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are subtle, shimmering shockwaves visible rippling across the fuselage and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot that tilts upward while following the aircraft climbing steeply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background view a vast curvature of Earth showing a white cloud bank over a deep blue ocean framed by thin, glowing blue atmospheric threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625991944223642 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625991944223642, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625991944223642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07953644543886185, + 0.06009320914745331, + 0.053730353713035583, + 0.08797325193881989, + 0.11281296610832214, + 0.08936935663223267, + 0.08110389113426208, + 0.175080344080925, + 0.09115011990070343, + 0.16915002465248108 + ], + "credit_map": [ + 0.006979782599955797, + 0.011992063373327255, + 0.009185981936752796, + 0.01113707385957241, + 0.014321746304631233, + 0.010190854780375957, + 0.008136242628097534, + 0.007592698559165001, + 0.004225885029882193, + 0.0051362598314881325, + 0.004914663266390562, + 0.007583032362163067, + 0.008433719165623188, + 0.01449343841522932, + 0.009744449518620968, + 0.00556176109239459, + 0.0052336836233735085, + 0.005152255762368441, + 0.006933123804628849, + 0.006427451968193054, + 0.010049102827906609, + 0.007633245550096035, + 0.0068413564004004, + 0.005460134241729975, + 0.014307347126305103, + 0.016207227483391762, + 0.010356010869145393, + 0.009963467717170715, + 0.009084023535251617, + 0.009253851138055325, + 0.008800837211310863, + 0.010000480338931084, + 0.015171448700129986, + 0.01788829080760479, + 0.016052819788455963, + 0.01353432610630989, + 0.01172228716313839, + 0.009310412220656872, + 0.015437318943440914, + 0.01369607076048851, + 0.013943169265985489, + 0.013425345532596111, + 0.007848244160413742, + 0.011325668543577194, + 0.0074693295173347, + 0.010784829035401344, + 0.015266100876033306, + 0.00930667296051979, + 0.009969054721295834, + 0.011810584925115108, + 0.009715682826936245, + 0.014852084219455719, + 0.006652428302913904, + 0.008306284435093403, + 0.009962616488337517, + 0.00983515940606594, + 0.014970514923334122, + 0.027106499299407005, + 0.01762930117547512, + 0.017123369500041008, + 0.033560846000909805, + 0.02512778341770172, + 0.02200889401137829, + 0.01755315065383911, + 0.010600106790661812, + 0.021751295775175095, + 0.01872045174241066, + 0.0060709090903401375, + 0.017884869128465652, + 0.00530611677095294, + 0.00684373639523983, + 0.0039726318791508675, + 0.013154462911188602, + 0.032031718641519547, + 0.022144339978694916, + 0.04088350757956505, + 0.017651569098234177, + 0.022093063220381737, + 0.014793364331126213, + 0.006397996563464403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.007625991944223642 + } + ], + "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/0000028/group_8/index_tensor([6045, 6045])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_8/index_tensor([6045, 6045])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9776e2d7ea66b9902be3b961147a226a0b4e3289 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_8/index_tensor([6045, 6045])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "macro close-up shot from a slightly overhead angle, the camera remains still, the frame focused squarely on a hand clad in a pristine white cotton glove, tightly and possessively clutching an antique brass pocket watch. Illuminated from the side by a single, intense, warm light source (perhaps a table lamp or fireplace flame), the watch's cover, intricately carved with Baroque scrollwork, is vividly rendered. Each raised edge reflects bright highlights, while the recessed grooves are immersed in deep shadows, giving the brass a warm, rich patina imbued with age. The glove's fabric is clearly visible in the light, and the grip creates subtle wrinkles and tension at the knuckles, suggesting a sense of tension or cherishment. The entire subject is suspended above an almost completely black, soft-focus background. This extreme contrast of light and shadow not only highlights the exquisite craftsmanship of the pocket watch and the pure texture of the gloves, but also creates a tranquil atmosphere of extreme focus, suspense and nostalgia, as if time has frozen at this moment.", + "seed": 8791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_8/index_6045_seed_8791_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": -0.2080956857651472, + "vlm_logprob": -0.2080956857651472, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.2080956857651472, + "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 macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9003161787986755 + ], + "opposite_logprob_tokens": [ + -0.5253161787986755 + ], + "logprob_reward": -0.9003161787986755, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9003161787986755, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0802183449268341, + 0.05038554593920708, + 0.02729480341076851, + 0.018572771921753883, + 0.020350275561213493, + 0.12766431272029877, + 0.11542462557554245, + 0.23572751879692078, + 0.26953965425491333, + 0.05482206121087074 + ], + "credit_map": [ + 0.006795794237405062, + 0.00824009533971548, + 0.00960773415863514, + 0.014426779001951218, + 0.014554006047546864, + 0.009410717524588108, + 0.00800385419279337, + 0.009179367683827877, + 0.005775343626737595, + 0.0064036063849925995, + 0.006887918803840876, + 0.008415572345256805, + 0.007727486547082663, + 0.005603808909654617, + 0.003566507715731859, + 0.006005300674587488, + 0.0034974028822034597, + 0.003333266358822584, + 0.004080331884324551, + 0.0036454633809626102, + 0.004504770506173372, + 0.003012338187545538, + 0.0019210402388125658, + 0.00330019136890769, + 0.002643541432917118, + 0.002075399737805128, + 0.002713459078222513, + 0.002605520188808441, + 0.0026300798635929823, + 0.0020649980287998915, + 0.0015043370658531785, + 0.0023354378063231707, + 0.002227574586868286, + 0.001871467218734324, + 0.0025032174307852983, + 0.003514822805300355, + 0.0035796761512756348, + 0.0022574863396584988, + 0.0015654233284294605, + 0.002830610843375325, + 0.008273066952824593, + 0.010102269239723682, + 0.01966422237455845, + 0.011593633331358433, + 0.04321041330695152, + 0.0099033759906888, + 0.006258600391447544, + 0.018658742308616638, + 0.00939125008881092, + 0.010924610309302807, + 0.02713441662490368, + 0.017200184985995293, + 0.014136974699795246, + 0.013592490926384926, + 0.008831947110593319, + 0.014212758280336857, + 0.02363460510969162, + 0.029229125007987022, + 0.05183868482708931, + 0.029105544090270996, + 0.028009574860334396, + 0.03087560273706913, + 0.018819885328412056, + 0.02421453222632408, + 0.020177826285362244, + 0.02549358271062374, + 0.03908989951014519, + 0.016168447211384773, + 0.0968734472990036, + 0.01875113882124424, + 0.0131526505574584, + 0.03983268514275551, + 0.005828415974974632, + 0.006344277411699295, + 0.00852634571492672, + 0.008509250357747078, + 0.009658160619437695, + 0.004604793153703213, + 0.005103189032524824, + 0.006247634533792734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.9003161787986755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062668278813362 + ], + "opposite_logprob_tokens": [ + -3.8956267833709717 + ], + "logprob_reward": -0.02062668278813362, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062668278813362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19077084958553314, + 0.0874628871679306, + 0.04103279113769531, + 0.018716992810368538, + 0.06236451119184494, + 0.057310767471790314, + 0.16030225157737732, + 0.16615904867649078, + 0.11066455394029617, + 0.1052154004573822 + ], + "credit_map": [ + 0.017716841772198677, + 0.014614710584282875, + 0.02392699383199215, + 0.04441516473889351, + 0.021511279046535492, + 0.019975176081061363, + 0.023097893223166466, + 0.025512786582112312, + 0.006776377093046904, + 0.007922359742224216, + 0.011662520468235016, + 0.023190585896372795, + 0.008690687827765942, + 0.006943071261048317, + 0.009588520042598248, + 0.012688759714365005, + 0.0034946082159876823, + 0.004234315827488899, + 0.005188118200749159, + 0.009090235456824303, + 0.005076215602457523, + 0.004391920752823353, + 0.003819969017058611, + 0.005737406201660633, + 0.0027451880741864443, + 0.0017873892793431878, + 0.0031442122999578714, + 0.003631063736975193, + 0.0021024299785494804, + 0.001245232648216188, + 0.0015538333682343364, + 0.002507643075659871, + 0.005516144447028637, + 0.006627708673477173, + 0.007338114082813263, + 0.01657954789698124, + 0.0063432203605771065, + 0.0047646318562328815, + 0.004841613117605448, + 0.010353527031838894, + 0.004907122813165188, + 0.006167565006762743, + 0.008679845370352268, + 0.010026206262409687, + 0.012885503470897675, + 0.004002280067652464, + 0.00346036977134645, + 0.0071818651631474495, + 0.012675352394580841, + 0.01976829767227173, + 0.0296074990183115, + 0.029200924560427666, + 0.01869351416826248, + 0.014823567122220993, + 0.014934287406504154, + 0.020598802715539932, + 0.015887467190623283, + 0.019417287781834602, + 0.026959385722875595, + 0.029406556859612465, + 0.016207903623580933, + 0.020500503480434418, + 0.01713275909423828, + 0.020647160708904266, + 0.007749909069389105, + 0.010414664633572102, + 0.015366115607321262, + 0.010022830218076706, + 0.036209236830472946, + 0.008485116995871067, + 0.00674482760950923, + 0.015671847388148308, + 0.007318213116377592, + 0.012483756057918072, + 0.018694326281547546, + 0.023404326289892197, + 0.011533410288393497, + 0.008028434589505196, + 0.010758849792182446, + 0.012994070537388325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.02062668278813362 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100250080227852 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100250080227852, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100250080227852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1128731444478035, + 0.04961149021983147, + 0.03993688151240349, + 0.05027393251657486, + 0.0178531426936388, + 0.12479891628026962, + 0.041999414563179016, + 0.060472503304481506, + 0.4544564485549927, + 0.04772414639592171 + ], + "credit_map": [ + 0.010534326545894146, + 0.011173411272466183, + 0.01772160455584526, + 0.019154828041791916, + 0.01498978678137064, + 0.01656608283519745, + 0.009125443175435066, + 0.01360765565186739, + 0.007013658061623573, + 0.006195449735969305, + 0.008342419750988483, + 0.006479349918663502, + 0.006423826329410076, + 0.005573113914579153, + 0.0035606902092695236, + 0.006022982764989138, + 0.004917931277304888, + 0.004107655957341194, + 0.007258681580424309, + 0.005349535495042801, + 0.005954555701464415, + 0.004439757205545902, + 0.0033950572833418846, + 0.004513701424002647, + 0.00666620722040534, + 0.004762629047036171, + 0.009380760602653027, + 0.005852303933352232, + 0.007612881250679493, + 0.005887256469577551, + 0.0040311142802238464, + 0.006080777384340763, + 0.00198094779625535, + 0.0015296798665076494, + 0.0037440555170178413, + 0.002690497087314725, + 0.0027600107714533806, + 0.001881101750768721, + 0.0011811753502115607, + 0.002085672924295068, + 0.010204415768384933, + 0.012669552117586136, + 0.02067425660789013, + 0.013259293511509895, + 0.035991765558719635, + 0.008626008406281471, + 0.006702647544443607, + 0.016670970246195793, + 0.0038002151995897293, + 0.003458726219832897, + 0.00905737653374672, + 0.006601977162063122, + 0.005215385928750038, + 0.005724768154323101, + 0.0025075546000152826, + 0.005633408669382334, + 0.004545759875327349, + 0.007600437384098768, + 0.01096748560667038, + 0.00725750345736742, + 0.013714900240302086, + 0.005877277813851833, + 0.00334183475933969, + 0.0071672964841127396, + 0.03427867591381073, + 0.04302971437573433, + 0.0665583610534668, + 0.025286564603447914, + 0.16362136602401733, + 0.03104759193956852, + 0.02221241593360901, + 0.06842169910669327, + 0.0035153913777321577, + 0.004308334551751614, + 0.008964692242443562, + 0.006495890207588673, + 0.008701153099536896, + 0.004349443130195141, + 0.005052013788372278, + 0.0063372207805514336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.004100250080227852 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906694501638412 + ], + "opposite_logprob_tokens": [ + -2.464066982269287 + ], + "logprob_reward": -0.08906694501638412, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906694501638412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06919727474451065, + 0.05181640759110451, + 0.057134512811899185, + 0.07590700685977936, + 0.021237792447209358, + 0.3572388291358948, + 0.05361483246088028, + 0.0366048589348793, + 0.18987858295440674, + 0.08736983686685562 + ], + "credit_map": [ + 0.006087139714509249, + 0.006667003966867924, + 0.013247362338006496, + 0.012190352194011211, + 0.007679027505218983, + 0.007840989157557487, + 0.007071088533848524, + 0.008414316922426224, + 0.00524553656578064, + 0.005582860205322504, + 0.008941534906625748, + 0.00983051024377346, + 0.005635042209178209, + 0.00476415641605854, + 0.005342485848814249, + 0.006474283989518881, + 0.0061685494147241116, + 0.005355659406632185, + 0.01071489043533802, + 0.00944028701633215, + 0.008049427531659603, + 0.005335415713489056, + 0.005389357451349497, + 0.006680931430310011, + 0.00842579547315836, + 0.007051501423120499, + 0.013937907293438911, + 0.010034509003162384, + 0.010934427380561829, + 0.009113254956901073, + 0.007488688454031944, + 0.008920930325984955, + 0.0020905889105051756, + 0.001984934089705348, + 0.0037480355240404606, + 0.003108375705778599, + 0.0031959745101630688, + 0.002921047620475292, + 0.0016073432052507997, + 0.002581495326012373, + 0.03035292960703373, + 0.03202202916145325, + 0.06591153889894485, + 0.050396628677845, + 0.07960007339715958, + 0.025450192391872406, + 0.02261308766901493, + 0.05089236795902252, + 0.005657273810356855, + 0.004883424378931522, + 0.013156135566532612, + 0.008337592706084251, + 0.007176590152084827, + 0.00403814110904932, + 0.003954960498958826, + 0.006410722155123949, + 0.002356435637921095, + 0.0045247687958180904, + 0.008217540569603443, + 0.0049708569422364235, + 0.005663462448865175, + 0.0035309407394379377, + 0.002776210894808173, + 0.004564644303172827, + 0.014354613609611988, + 0.01798349805176258, + 0.029049905017018318, + 0.011143212206661701, + 0.0668201670050621, + 0.013036946766078472, + 0.009468765929341316, + 0.028021475300192833, + 0.011980780400335789, + 0.009944084100425243, + 0.018934691324830055, + 0.00872349925339222, + 0.011153743602335453, + 0.009231967851519585, + 0.007242357823997736, + 0.010158722288906574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.08906694501638412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13198992609977722, + 0.041870106011629105, + 0.06565625965595245, + 0.09134276956319809, + 0.06707033514976501, + 0.17281809449195862, + 0.08524612337350845, + 0.05395790934562683, + 0.24267643690109253, + 0.04737207293510437 + ], + "credit_map": [ + 0.014846852980554104, + 0.012918801046907902, + 0.018376657739281654, + 0.022435465827584267, + 0.017257124185562134, + 0.018281403928995132, + 0.011705056764185429, + 0.016168566420674324, + 0.00789718795567751, + 0.004534071311354637, + 0.008512919768691063, + 0.0045120869763195515, + 0.005072946194559336, + 0.0042888326570391655, + 0.0022419667802751064, + 0.004810092505067587, + 0.008560994639992714, + 0.006709787994623184, + 0.012017030268907547, + 0.00891588069498539, + 0.009613965637981892, + 0.006774697918444872, + 0.005367120262235403, + 0.007696780376136303, + 0.01179024763405323, + 0.008429513312876225, + 0.015797430649399757, + 0.01218966394662857, + 0.013550486415624619, + 0.010372869670391083, + 0.008263313211500645, + 0.010949241928756237, + 0.00884889904409647, + 0.005381186958402395, + 0.013538563624024391, + 0.007930070161819458, + 0.010277418419718742, + 0.007921851240098476, + 0.004844134673476219, + 0.008328214287757874, + 0.013900639489293098, + 0.013826555572450161, + 0.025397486984729767, + 0.028298407793045044, + 0.03963712975382805, + 0.013780159875750542, + 0.011183132417500019, + 0.026794590055942535, + 0.009619880467653275, + 0.011012701317667961, + 0.015091793611645699, + 0.012183190323412418, + 0.008917538449168205, + 0.009820803068578243, + 0.008071046322584152, + 0.010529173538088799, + 0.005189531948417425, + 0.007099711801856756, + 0.009745893068611622, + 0.006644480861723423, + 0.010418922640383244, + 0.005676640663295984, + 0.0029495747294276953, + 0.006233152002096176, + 0.01813580095767975, + 0.02283639833331108, + 0.03500129655003548, + 0.01358610950410366, + 0.08732966333627701, + 0.01744428463280201, + 0.012253722175955772, + 0.036089155822992325, + 0.004661780782043934, + 0.0055802869610488415, + 0.007948657497763634, + 0.006770438980311155, + 0.006839906796813011, + 0.004645827226340771, + 0.005206105299293995, + 0.005719066131860018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.02636837214231491 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot from a slightly overhead angle, the camera remains still, the frame focused squarely on a hand clad in a pristine white cotton glove, tightly and possessively clutching an antique brass pocket watch. Illuminated from the side by a single, intense, warm light source (perhaps a table lamp or fireplace flame), the watch's cover, intricately carved with Baroque scrollwork, is vividly rendered. Each raised edge reflects bright highlights, while the recessed grooves are immersed in deep shadows, giving the brass a warm, rich patina imbued with age. The glove's fabric is clearly visible in the light, and the grip creates subtle wrinkles and tension at the knuckles, suggesting a sense of tension or cherishment. The entire subject is suspended above an almost completely black, soft-focus background. This extreme contrast of light and shadow not only highlights the exquisite craftsmanship of the pocket watch and the pure texture of the gloves, but also creates a tranquil atmosphere of extreme focus, suspense and nostalgia, as if time has frozen at this moment.", + "seed": 8792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_8/index_6045_seed_8792_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": -0.009435600973665715, + "vlm_logprob": -0.009435600973665715, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.009435600973665715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010602337308228016 + ], + "opposite_logprob_tokens": [ + -4.6356024742126465 + ], + "logprob_reward": -0.010602337308228016, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010602337308228016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36508649587631226, + 0.1634386032819748, + 0.08129213750362396, + 0.05012434720993042, + 0.06555458903312683, + 0.05033041909337044, + 0.04827377200126648, + 0.054258767515420914, + 0.06679728627204895, + 0.05484361574053764 + ], + "credit_map": [ + 0.03063414804637432, + 0.03253861144185066, + 0.04004557803273201, + 0.06354115903377533, + 0.037303626537323, + 0.05302932858467102, + 0.0401080921292305, + 0.06788593530654907, + 0.013488339260220528, + 0.01812339387834072, + 0.014679097570478916, + 0.01808285340666771, + 0.021885497495532036, + 0.027508879080414772, + 0.018752356991171837, + 0.030918171629309654, + 0.0056771147064864635, + 0.011394779197871685, + 0.007359590381383896, + 0.009650787338614464, + 0.01468871533870697, + 0.010733727365732193, + 0.011168863624334335, + 0.010618554428219795, + 0.004041727166622877, + 0.004453393165022135, + 0.005357209127396345, + 0.006081659812480211, + 0.005244161933660507, + 0.008191109634935856, + 0.006687662098556757, + 0.010067418217658997, + 0.0048027075827121735, + 0.005875921342521906, + 0.006628502160310745, + 0.010709097608923912, + 0.007777748163789511, + 0.00654660863801837, + 0.012763796374201775, + 0.010450206696987152, + 0.0038108881562948227, + 0.004524120595306158, + 0.0038822160568088293, + 0.006051098927855492, + 0.007541571743786335, + 0.007112647872418165, + 0.009407413192093372, + 0.008000455796718597, + 0.003436947474256158, + 0.004132562782615423, + 0.003684913506731391, + 0.007851755246520042, + 0.005489187315106392, + 0.006720826495438814, + 0.00819872785359621, + 0.00875884760171175, + 0.005008023697882891, + 0.0034281897824257612, + 0.009540777653455734, + 0.006135441828519106, + 0.004211729392409325, + 0.00917092151939869, + 0.005126741714775562, + 0.011636937968432903, + 0.006497577764093876, + 0.007227003574371338, + 0.011189954355359077, + 0.010274497792124748, + 0.008477519266307354, + 0.007302915211766958, + 0.007107746787369251, + 0.008720062673091888, + 0.005341298412531614, + 0.005918124690651894, + 0.004122024402022362, + 0.0071031153202056885, + 0.007152517791837454, + 0.009144781157374382, + 0.007971045561134815, + 0.00809070561081171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.010602337308228016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032042134553194046 + ], + "opposite_logprob_tokens": [ + -5.753204345703125 + ], + "logprob_reward": -0.0032042134553194046, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032042134553194046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18198533356189728, + 0.1268434226512909, + 0.17419055104255676, + 0.058296434581279755, + 0.09694567322731018, + 0.07506085187196732, + 0.07785475254058838, + 0.04059946909546852, + 0.07687084376811981, + 0.0913526639342308 + ], + "credit_map": [ + 0.012694722972810268, + 0.018693946301937103, + 0.020304925739765167, + 0.02901928685605526, + 0.023431653156876564, + 0.022954847663640976, + 0.026453394442796707, + 0.028432561084628105, + 0.009304066188633442, + 0.010187818668782711, + 0.014898058958351612, + 0.014158189296722412, + 0.016711046919226646, + 0.020747745409607887, + 0.013972572982311249, + 0.026863928884267807, + 0.010926397517323494, + 0.025724604725837708, + 0.013276103883981705, + 0.01742430217564106, + 0.03476959839463234, + 0.022259077057242393, + 0.022963840514421463, + 0.026846639811992645, + 0.0038838034961372614, + 0.006301511079072952, + 0.005058296490460634, + 0.0063033695332705975, + 0.01183810643851757, + 0.007321985438466072, + 0.00913523230701685, + 0.008454136550426483, + 0.006041004788130522, + 0.0067042200826108456, + 0.0063764555379748344, + 0.014809426851570606, + 0.012087903916835785, + 0.01194208487868309, + 0.021682217717170715, + 0.017302365973591805, + 0.005973547697067261, + 0.006607558112591505, + 0.007131831254810095, + 0.012540207244455814, + 0.007811755873262882, + 0.011126148514449596, + 0.009547724388539791, + 0.014322078786790371, + 0.006552243605256081, + 0.006113035604357719, + 0.014884663745760918, + 0.008147592656314373, + 0.008520488627254963, + 0.01058378629386425, + 0.007735204417258501, + 0.015317743644118309, + 0.0023742904886603355, + 0.004452049266546965, + 0.0025665624998509884, + 0.006462910678237677, + 0.007774778176099062, + 0.005320150405168533, + 0.00534568727016449, + 0.006303046364337206, + 0.006084414664655924, + 0.00692171324044466, + 0.01048815157264471, + 0.00801727082580328, + 0.011272447183728218, + 0.012357715517282486, + 0.008399960584938526, + 0.013329172506928444, + 0.006110806483775377, + 0.010083827190101147, + 0.006863943766802549, + 0.012340173125267029, + 0.012119892984628677, + 0.01588059589266777, + 0.011866468004882336, + 0.016086967661976814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0032042134553194046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818987913429737 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.01818987913429737, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818987913429737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40804651379585266, + 0.08375748991966248, + 0.19789522886276245, + 0.03835764527320862, + 0.06381168216466904, + 0.08429981023073196, + 0.049062736332416534, + 0.04231453314423561, + 0.011239114217460155, + 0.021215209737420082 + ], + "credit_map": [ + 0.03288654237985611, + 0.034364596009254456, + 0.06287763267755508, + 0.053909074515104294, + 0.03388466313481331, + 0.06288102269172668, + 0.04520860314369202, + 0.0820344015955925, + 0.006896442733705044, + 0.007777696009725332, + 0.008672104217112064, + 0.012133976444602013, + 0.01121942512691021, + 0.012730512768030167, + 0.01079328078776598, + 0.013534050434827805, + 0.016624826937913895, + 0.01965237967669964, + 0.02150239609181881, + 0.025825554504990578, + 0.016949862241744995, + 0.03598424419760704, + 0.02087680995464325, + 0.0404791533946991, + 0.0029003662057220936, + 0.005305042024701834, + 0.002890475559979677, + 0.004917672835290432, + 0.006367632653564215, + 0.004684867803007364, + 0.0056098951026797295, + 0.005681690294295549, + 0.002857242012396455, + 0.005110495258122683, + 0.002700899261981249, + 0.008633371442556381, + 0.010191603563725948, + 0.007377986796200275, + 0.016420481726527214, + 0.010519606992602348, + 0.006071483716368675, + 0.01007351279258728, + 0.009673404507339, + 0.014214822091162205, + 0.008156831376254559, + 0.012607002630829811, + 0.00965164415538311, + 0.013851115480065346, + 0.0036176785361021757, + 0.0039572264067828655, + 0.008002922870218754, + 0.00604536896571517, + 0.005414553452283144, + 0.006569357588887215, + 0.006306280381977558, + 0.009149347431957722, + 0.0025852948892861605, + 0.00405165646225214, + 0.003852785797789693, + 0.0061713410541415215, + 0.0050201574340462685, + 0.007161241956055164, + 0.005380758084356785, + 0.008091294206678867, + 0.0008620248991064727, + 0.0013328278437256813, + 0.0009269907022826374, + 0.0017209158977493644, + 0.0016257318202406168, + 0.0012842344585806131, + 0.001948708202689886, + 0.001537679578177631, + 0.0011463603004813194, + 0.002095650415867567, + 0.0015733817126601934, + 0.002377904485911131, + 0.002912316471338272, + 0.004247779026627541, + 0.0027893814258277416, + 0.004072437062859535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01818987913429737 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097519442439079 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097519442439079, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097519442439079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29805850982666016, + 0.08536506444215775, + 0.1686916947364807, + 0.02848360873758793, + 0.05571424961090088, + 0.06748732924461365, + 0.0637013390660286, + 0.1418779492378235, + 0.021114131435751915, + 0.06950607150793076 + ], + "credit_map": [ + 0.022117992863059044, + 0.021210666745901108, + 0.03593111410737038, + 0.0536084808409214, + 0.03065105527639389, + 0.03767998144030571, + 0.03989572450518608, + 0.05696352571249008, + 0.007555137388408184, + 0.008769124746322632, + 0.009558125399053097, + 0.013035374693572521, + 0.007903934456408024, + 0.014091785065829754, + 0.008508097380399704, + 0.015943489968776703, + 0.016540609300136566, + 0.01590275764465332, + 0.020502055063843727, + 0.03250216692686081, + 0.009206542745232582, + 0.024700824171304703, + 0.019315406680107117, + 0.030021339654922485, + 0.002459755865857005, + 0.002844100119546056, + 0.0035751473624259233, + 0.004429592750966549, + 0.00338233495131135, + 0.0036540047731250525, + 0.003985572140663862, + 0.004153101705014706, + 0.004815266467630863, + 0.005341803189367056, + 0.006948265712708235, + 0.011000960133969784, + 0.004175711888819933, + 0.007660772651433945, + 0.0071366592310369015, + 0.008634812198579311, + 0.005478079430758953, + 0.007928092032670975, + 0.007896268740296364, + 0.010408448055386543, + 0.00750214047729969, + 0.009311498142778873, + 0.008954365737736225, + 0.010008440352976322, + 0.004534274339675903, + 0.006023635156452656, + 0.008307420648634434, + 0.008899317122995853, + 0.007439493667334318, + 0.00802238006144762, + 0.010301387868821621, + 0.010173434391617775, + 0.013808542862534523, + 0.009950894862413406, + 0.028770681470632553, + 0.018842950463294983, + 0.014157191850244999, + 0.018856381997466087, + 0.015073291026055813, + 0.022418029606342316, + 0.0017495083156973124, + 0.001863489393144846, + 0.0032144892029464245, + 0.0022849703673273325, + 0.002239895984530449, + 0.003237873548641801, + 0.0026282852049916983, + 0.0038956196513026953, + 0.0049503883346915245, + 0.007559711579233408, + 0.00730171799659729, + 0.01122345868498087, + 0.004646791610866785, + 0.013096094131469727, + 0.006853609345853329, + 0.013874304480850697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.004097519442439079 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110840555280447 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.0110840555280447, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110840555280447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15127044916152954, + 0.053983233869075775, + 0.07031457126140594, + 0.09751129895448685, + 0.4309653639793396, + 0.10648521780967712, + 0.035397905856370926, + 0.02171134389936924, + 0.017499754205346107, + 0.014860833063721657 + ], + "credit_map": [ + 0.011320562101900578, + 0.013624426908791065, + 0.017978433519601822, + 0.03159840404987335, + 0.014023217372596264, + 0.017869822680950165, + 0.021398024633526802, + 0.023457564413547516, + 0.004503706470131874, + 0.005740126129239798, + 0.00777303334325552, + 0.009832020848989487, + 0.007128506433218718, + 0.0053007411770522594, + 0.00534208444878459, + 0.008363023400306702, + 0.006859481800347567, + 0.0074547166004776955, + 0.006553948391228914, + 0.009838427416980267, + 0.00816943310201168, + 0.01087051909416914, + 0.009520523250102997, + 0.011047521606087685, + 0.007242895197123289, + 0.011719238944351673, + 0.006585338618606329, + 0.014886564575135708, + 0.015111560933291912, + 0.013730993494391441, + 0.014738739468157291, + 0.013495968654751778, + 0.023816436529159546, + 0.030326535925269127, + 0.023547431454062462, + 0.060514044016599655, + 0.054037999361753464, + 0.05610653758049011, + 0.10463989526033401, + 0.0779765397310257, + 0.006830460857599974, + 0.011022270657122135, + 0.007464846130460501, + 0.017101489007472992, + 0.013985596597194672, + 0.017246825620532036, + 0.017523985356092453, + 0.015309754759073257, + 0.0023961886763572693, + 0.003090179292485118, + 0.004444193094968796, + 0.005194440484046936, + 0.003807696746662259, + 0.0038907562848180532, + 0.007710434030741453, + 0.004864019807428122, + 0.0017212979728356004, + 0.0018617348978295922, + 0.0038386546075344086, + 0.002355630975216627, + 0.002613010350614786, + 0.0029382153879851103, + 0.0028880080208182335, + 0.003494794713333249, + 0.0010962478118017316, + 0.002343833912163973, + 0.0012042210437357426, + 0.002296434948220849, + 0.003227260196581483, + 0.002156028524041176, + 0.002340270671993494, + 0.0028354579117149115, + 0.001094938488677144, + 0.0018552187830209732, + 0.0013661050470545888, + 0.0021528059151023626, + 0.001899189897812903, + 0.0017773130675777793, + 0.002555383602157235, + 0.0021598790772259235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0110840555280447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot from a slightly overhead angle, the camera remains still, the frame focused squarely on a hand clad in a pristine white cotton glove, tightly and possessively clutching an antique brass pocket watch. Illuminated from the side by a single, intense, warm light source (perhaps a table lamp or fireplace flame), the watch's cover, intricately carved with Baroque scrollwork, is vividly rendered. Each raised edge reflects bright highlights, while the recessed grooves are immersed in deep shadows, giving the brass a warm, rich patina imbued with age. The glove's fabric is clearly visible in the light, and the grip creates subtle wrinkles and tension at the knuckles, suggesting a sense of tension or cherishment. The entire subject is suspended above an almost completely black, soft-focus background. This extreme contrast of light and shadow not only highlights the exquisite craftsmanship of the pocket watch and the pure texture of the gloves, but also creates a tranquil atmosphere of extreme focus, suspense and nostalgia, as if time has frozen at this moment.", + "seed": 8793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_8/index_6045_seed_8793_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": -0.1839376379735768, + "vlm_logprob": -0.1839376379735768, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.1839376379735768, + "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 macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6937994956970215 + ], + "opposite_logprob_tokens": [ + -0.6937994956970215 + ], + "logprob_reward": -0.6937994956970215, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6937994956970215, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.42734426259994507, + 0.12270783632993698, + 0.06573837250471115, + 0.052672866731882095, + 0.056313782930374146, + 0.06188786029815674, + 0.042755041271448135, + 0.029622595757246017, + 0.07131753116846085, + 0.06963988393545151 + ], + "credit_map": [ + 0.04166432097554207, + 0.03852645680308342, + 0.03428898751735687, + 0.06515638530254364, + 0.04689125716686249, + 0.06033794581890106, + 0.08509434759616852, + 0.05538458004593849, + 0.01241992972791195, + 0.01542881689965725, + 0.01019190438091755, + 0.020464690402150154, + 0.013672581873834133, + 0.015239626169204712, + 0.02270067296922207, + 0.01258961670100689, + 0.006150107830762863, + 0.00899508222937584, + 0.005385376513004303, + 0.01040654070675373, + 0.008159908466041088, + 0.008282464928925037, + 0.010205277241766453, + 0.008153610862791538, + 0.004004079382866621, + 0.007254524156451225, + 0.00426239799708128, + 0.011175479739904404, + 0.007519968785345554, + 0.006689481437206268, + 0.007012096233665943, + 0.004754840396344662, + 0.006011124700307846, + 0.0059685539454221725, + 0.004341456573456526, + 0.006697417236864567, + 0.008603638969361782, + 0.009273924864828587, + 0.009879271499812603, + 0.005538397002965212, + 0.008051317185163498, + 0.008212761022150517, + 0.004953522235155106, + 0.008939771912992, + 0.007599113509058952, + 0.008848337456583977, + 0.008828461170196533, + 0.0064545730128884315, + 0.005575546994805336, + 0.004157075192779303, + 0.004129636101424694, + 0.006585813593119383, + 0.0051809074357151985, + 0.004581172950565815, + 0.006369598675519228, + 0.006175290793180466, + 0.004045806359499693, + 0.0030172511469572783, + 0.0024399079848080873, + 0.0041780536994338036, + 0.003336068009957671, + 0.00461154617369175, + 0.004539312794804573, + 0.0034546509850770235, + 0.005844346713274717, + 0.010077293030917645, + 0.0072058443911373615, + 0.011888585053384304, + 0.007797542028129101, + 0.008967554196715355, + 0.012131714262068272, + 0.00740465335547924, + 0.008740958757698536, + 0.006981472484767437, + 0.004820417147129774, + 0.011552838608622551, + 0.007778811268508434, + 0.010203970596194267, + 0.010796246118843555, + 0.008765175938606262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.6937994956970215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20166485011577606 + ], + "opposite_logprob_tokens": [ + -1.7016648054122925 + ], + "logprob_reward": -0.20166485011577606, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20166485011577606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25410085916519165, + 0.1335706263780594, + 0.12259294092655182, + 0.061790015548467636, + 0.11018994450569153, + 0.07578159123659134, + 0.10407323390245438, + 0.06482453644275665, + 0.03093269094824791, + 0.04214349761605263 + ], + "credit_map": [ + 0.0263496246188879, + 0.02207137458026409, + 0.024629676714539528, + 0.03994010388851166, + 0.027143150568008423, + 0.028927627950906754, + 0.05040467530488968, + 0.03463463857769966, + 0.014916406944394112, + 0.014224355109035969, + 0.01138380728662014, + 0.015593509189784527, + 0.01832738146185875, + 0.023183541372418404, + 0.01989094913005829, + 0.01605067029595375, + 0.01450190506875515, + 0.014756995253264904, + 0.009370463900268078, + 0.020701687783002853, + 0.01363759208470583, + 0.021708548069000244, + 0.014512814581394196, + 0.013402929529547691, + 0.0072356220334768295, + 0.006854556966573, + 0.004157514311373234, + 0.010043834336102009, + 0.007460305467247963, + 0.012756623327732086, + 0.0067438362166285515, + 0.006537724751979113, + 0.013247876428067684, + 0.009248090907931328, + 0.006915450096130371, + 0.009929624386131763, + 0.011477864347398281, + 0.03152903541922569, + 0.01055004820227623, + 0.017291957512497902, + 0.006914585363119841, + 0.007754017598927021, + 0.0056258030235767365, + 0.008646195754408836, + 0.008576788008213043, + 0.017470017075538635, + 0.010408730246126652, + 0.010385450907051563, + 0.012731279246509075, + 0.013135252520442009, + 0.007917487062513828, + 0.017949754372239113, + 0.01107969880104065, + 0.012027380988001823, + 0.011419747956097126, + 0.017812645062804222, + 0.006097211968153715, + 0.007673552725464106, + 0.004784945398569107, + 0.01105231698602438, + 0.008250493556261063, + 0.010620291344821453, + 0.009998916648328304, + 0.006346806883811951, + 0.003803693223744631, + 0.00372878834605217, + 0.003110485849902034, + 0.0048447297886013985, + 0.003466809168457985, + 0.0033254241570830345, + 0.004560776986181736, + 0.00409198272973299, + 0.005071926396340132, + 0.004582997877150774, + 0.003222440369427204, + 0.00574173079803586, + 0.004329740535467863, + 0.007532734423875809, + 0.005546321161091328, + 0.0061156065203249454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20166485011577606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002823891816660762 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.002823891816660762, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002823891816660762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24107088148593903, + 0.10697316378355026, + 0.06398583203554153, + 0.08134272694587708, + 0.18007086217403412, + 0.07792627811431885, + 0.043422166258096695, + 0.06080293282866478, + 0.07484903931617737, + 0.0695560872554779 + ], + "credit_map": [ + 0.02406373992562294, + 0.025125950574874878, + 0.01718490943312645, + 0.04756851866841316, + 0.022352103143930435, + 0.030066508799791336, + 0.029819857329130173, + 0.044889286160469055, + 0.0138394171372056, + 0.012167575769126415, + 0.010462996549904346, + 0.014920977875590324, + 0.012660413049161434, + 0.01917220465838909, + 0.012405602261424065, + 0.011343972757458687, + 0.0080359922721982, + 0.005021905992180109, + 0.0050550661981105804, + 0.0069813355803489685, + 0.007346360012888908, + 0.01385575346648693, + 0.008214683271944523, + 0.009474735707044601, + 0.007836158387362957, + 0.008695843629539013, + 0.005827626213431358, + 0.010956792160868645, + 0.009656255133450031, + 0.016730498522520065, + 0.012869767844676971, + 0.00876978412270546, + 0.021242618560791016, + 0.01411864161491394, + 0.009639803320169449, + 0.020423073321580887, + 0.01930985599756241, + 0.06088487058877945, + 0.012014106847345829, + 0.02243790216743946, + 0.008095955476164818, + 0.010777942836284637, + 0.00613878620788455, + 0.01639036275446415, + 0.00778109161183238, + 0.010765648446977139, + 0.010527490638196468, + 0.007448997348546982, + 0.0050377617590129375, + 0.005213658791035414, + 0.004055490251630545, + 0.0058470191434025764, + 0.005155399441719055, + 0.006686436012387276, + 0.005881709046661854, + 0.005544692277908325, + 0.006778456270694733, + 0.005985544994473457, + 0.0057543120346963406, + 0.011198962107300758, + 0.00716603035107255, + 0.008953569456934929, + 0.008719679899513721, + 0.006246380507946014, + 0.008085943758487701, + 0.00893387570977211, + 0.005852447357028723, + 0.011969747953116894, + 0.009960385039448738, + 0.008825116790831089, + 0.010971594601869583, + 0.010249927639961243, + 0.006253383122384548, + 0.007716331165283918, + 0.005080456379801035, + 0.014954496175050735, + 0.009700738824903965, + 0.009104537777602673, + 0.00968991033732891, + 0.0070562344044446945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.002823891816660762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031967272516340017 + ], + "opposite_logprob_tokens": [ + -5.753196716308594 + ], + "logprob_reward": -0.0031967272516340017, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031967272516340017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1262679398059845, + 0.06088254600763321, + 0.07513704150915146, + 0.11896276473999023, + 0.2105645388364792, + 0.21971562504768372, + 0.027878575026988983, + 0.0766613706946373, + 0.03251101076602936, + 0.05141858011484146 + ], + "credit_map": [ + 0.011108887381851673, + 0.019379274919629097, + 0.01238267868757248, + 0.024352649226784706, + 0.0167342908680439, + 0.015498539432883263, + 0.013239418156445026, + 0.013572202995419502, + 0.006349621340632439, + 0.008893799036741257, + 0.004827731288969517, + 0.008859117515385151, + 0.005387707147747278, + 0.010264123789966106, + 0.008879484608769417, + 0.007420959882438183, + 0.006636932957917452, + 0.010096844285726547, + 0.006003654096275568, + 0.017794281244277954, + 0.005954562686383724, + 0.01063879020512104, + 0.010631006211042404, + 0.007380971685051918, + 0.013524564914405346, + 0.01618324965238571, + 0.010167543776333332, + 0.029107779264450073, + 0.01021519023925066, + 0.014580515213310719, + 0.01533367671072483, + 0.009850249625742435, + 0.02291320078074932, + 0.019785616546869278, + 0.011277884244918823, + 0.026121094822883606, + 0.026693051680922508, + 0.06278753280639648, + 0.015723099932074547, + 0.025263041257858276, + 0.025971047580242157, + 0.034105949103832245, + 0.016802888363599777, + 0.0455801822245121, + 0.020991677418351173, + 0.025597432628273964, + 0.028056684881448746, + 0.022609762847423553, + 0.003600990865379572, + 0.004194504581391811, + 0.0022893829736858606, + 0.0033563077449798584, + 0.0031113214790821075, + 0.0034289767500013113, + 0.0036245896480977535, + 0.004272498656064272, + 0.008479775860905647, + 0.006910393945872784, + 0.006683167070150375, + 0.009039080701768398, + 0.009426864795386791, + 0.016064133495092392, + 0.012589804828166962, + 0.007468159776180983, + 0.0035742674954235554, + 0.003788632107898593, + 0.00210339087061584, + 0.005522342398762703, + 0.0036124896723777056, + 0.005547722801566124, + 0.004796725232154131, + 0.003565437626093626, + 0.003939828369766474, + 0.0065114982426166534, + 0.0044311038218438625, + 0.013235155493021011, + 0.00506379222497344, + 0.006669485010206699, + 0.0065638720989227295, + 0.0050038425251841545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0031967272516340017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820322498679161 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.01820322498679161, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820322498679161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1777244359254837, + 0.0899498239159584, + 0.06684114784002304, + 0.06290064007043839, + 0.10308646410703659, + 0.06273283809423447, + 0.0905507504940033, + 0.21401700377464294, + 0.0744962990283966, + 0.05770053341984749 + ], + "credit_map": [ + 0.017014237120747566, + 0.0207589790225029, + 0.012121034786105156, + 0.035306014120578766, + 0.014791594818234444, + 0.025459706783294678, + 0.020231356844305992, + 0.032041508704423904, + 0.011818807572126389, + 0.008149324916303158, + 0.00951866153627634, + 0.015282021835446358, + 0.01100290659815073, + 0.012423527427017689, + 0.010540805757045746, + 0.011213765479624271, + 0.007395414635539055, + 0.007198892068117857, + 0.0058063636533916, + 0.009413663297891617, + 0.0076571619138121605, + 0.00786060094833374, + 0.012520170770585537, + 0.008988876827061176, + 0.006636329460889101, + 0.006022065877914429, + 0.004852169193327427, + 0.010665325447916985, + 0.005678737536072731, + 0.011210872791707516, + 0.0088110426440835, + 0.009024100378155708, + 0.008702541701495647, + 0.01448107324540615, + 0.007351384032517672, + 0.012867358513176441, + 0.018123086541891098, + 0.014260579831898212, + 0.015110790729522705, + 0.012189643457531929, + 0.005390224512666464, + 0.008799265138804913, + 0.004242366179823875, + 0.00950002670288086, + 0.007630399893969297, + 0.010135304182767868, + 0.01000964641571045, + 0.007025605998933315, + 0.01008443720638752, + 0.00952076818794012, + 0.008249317295849323, + 0.011656856164336205, + 0.010982302017509937, + 0.013885009102523327, + 0.013186432421207428, + 0.01298562902957201, + 0.022251885384321213, + 0.019643522799015045, + 0.019994212314486504, + 0.028527677059173584, + 0.026227299124002457, + 0.039208099246025085, + 0.03619357943534851, + 0.021970734000205994, + 0.007331317290663719, + 0.0075083221308887005, + 0.006815718952566385, + 0.011468237265944481, + 0.008038063533604145, + 0.011686920188367367, + 0.012946768663823605, + 0.008700953796505928, + 0.0049348934553563595, + 0.0056016696617007256, + 0.005399893503636122, + 0.009578961879014969, + 0.008497560396790504, + 0.006980588659644127, + 0.01049109734594822, + 0.006215869914740324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.01820322498679161 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot from a slightly overhead angle, the camera remains still, the frame focused squarely on a hand clad in a pristine white cotton glove, tightly and possessively clutching an antique brass pocket watch. Illuminated from the side by a single, intense, warm light source (perhaps a table lamp or fireplace flame), the watch's cover, intricately carved with Baroque scrollwork, is vividly rendered. Each raised edge reflects bright highlights, while the recessed grooves are immersed in deep shadows, giving the brass a warm, rich patina imbued with age. The glove's fabric is clearly visible in the light, and the grip creates subtle wrinkles and tension at the knuckles, suggesting a sense of tension or cherishment. The entire subject is suspended above an almost completely black, soft-focus background. This extreme contrast of light and shadow not only highlights the exquisite craftsmanship of the pocket watch and the pure texture of the gloves, but also creates a tranquil atmosphere of extreme focus, suspense and nostalgia, as if time has frozen at this moment.", + "seed": 8794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_8/index_6045_seed_8794_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": -0.06933544455096126, + "vlm_logprob": -0.06933544455096126, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.06933544455096126, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.253074049949646 + ], + "opposite_logprob_tokens": [ + -1.503074049949646 + ], + "logprob_reward": -0.253074049949646, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.253074049949646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42655086517333984, + 0.11405108869075775, + 0.06252583861351013, + 0.03462761268019676, + 0.20398662984371185, + 0.033878982067108154, + 0.050672441720962524, + 0.016700156033039093, + 0.02725697122514248, + 0.02974943444132805 + ], + "credit_map": [ + 0.03970681130886078, + 0.03375060111284256, + 0.04527239128947258, + 0.04359891638159752, + 0.056671854108572006, + 0.07199385017156601, + 0.07974330335855484, + 0.05581314489245415, + 0.01166515052318573, + 0.028566274791955948, + 0.004482089076191187, + 0.00870231632143259, + 0.011601007543504238, + 0.030278997495770454, + 0.010931813158094883, + 0.007823442108929157, + 0.006610965821892023, + 0.02244386449456215, + 0.003255098592489958, + 0.004947887733578682, + 0.010946320369839668, + 0.007447659969329834, + 0.004157567396759987, + 0.0027164726052433252, + 0.0026949995663017035, + 0.00817671325057745, + 0.002418417716398835, + 0.004502140451222658, + 0.005667930468916893, + 0.004584309179335833, + 0.0040739839896559715, + 0.0025091171264648438, + 0.0199571680277586, + 0.050603047013282776, + 0.004782711621373892, + 0.011228114366531372, + 0.024511223658919334, + 0.06371228396892548, + 0.006489521823823452, + 0.022702567279338837, + 0.0031425224151462317, + 0.002486499957740307, + 0.0020547497551888227, + 0.0027015397790819407, + 0.005124276969581842, + 0.006872161291539669, + 0.005765712354332209, + 0.0057315225712955, + 0.004783990327268839, + 0.00458104582503438, + 0.005019139964133501, + 0.00466860830783844, + 0.009067629463970661, + 0.007799725513905287, + 0.006996302865445614, + 0.007755999453365803, + 0.0017535780789330602, + 0.0010214427020400763, + 0.0011943635763600469, + 0.0012742357794195414, + 0.0026426189579069614, + 0.0021821786649525166, + 0.0027499846182763577, + 0.0038817545864731073, + 0.0028379149734973907, + 0.002554624807089567, + 0.0020007409621030092, + 0.002917254576459527, + 0.003254752838984132, + 0.005393349565565586, + 0.0040539284236729145, + 0.004244403913617134, + 0.004167591687291861, + 0.0019261820707470179, + 0.003287807572633028, + 0.0023589213378727436, + 0.004696099553257227, + 0.005101094488054514, + 0.0029529472813010216, + 0.0052587916143238544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.253074049949646 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822909340262413 + ], + "opposite_logprob_tokens": [ + -4.018229007720947 + ], + "logprob_reward": -0.01822909340262413, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822909340262413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3422355651855469, + 0.11036405712366104, + 0.27741867303848267, + 0.015425067394971848, + 0.06691031903028488, + 0.057633861899375916, + 0.038299959152936935, + 0.018185822293162346, + 0.04475664719939232, + 0.028770027682185173 + ], + "credit_map": [ + 0.03234009072184563, + 0.02690144069492817, + 0.031585339456796646, + 0.035227444022893906, + 0.04741750657558441, + 0.05712692067027092, + 0.0696239098906517, + 0.04201289266347885, + 0.019050784409046173, + 0.013808019459247589, + 0.004941793158650398, + 0.006999849807471037, + 0.009343638084828854, + 0.03431606665253639, + 0.012855029664933681, + 0.009048881009221077, + 0.028822194784879684, + 0.10250840336084366, + 0.01313994824886322, + 0.02215079590678215, + 0.04932112246751785, + 0.03261882811784744, + 0.017210042104125023, + 0.01164732500910759, + 0.0013833681587129831, + 0.002559290500357747, + 0.0012866932665929198, + 0.00216474337503314, + 0.00197225552983582, + 0.0021735450718551874, + 0.0023666725028306246, + 0.0015184981748461723, + 0.007085470948368311, + 0.014307857491075993, + 0.0016323646996170282, + 0.0036817919462919235, + 0.00655062310397625, + 0.022262169048190117, + 0.003253889735788107, + 0.008136149495840073, + 0.004385337233543396, + 0.007833222858607769, + 0.004799669608473778, + 0.0054658264853060246, + 0.011077900417149067, + 0.007005338557064533, + 0.010006515309214592, + 0.007060049567371607, + 0.004268404096364975, + 0.003725716844201088, + 0.005184152629226446, + 0.004141735844314098, + 0.006324062589555979, + 0.00446160975843668, + 0.005572906229645014, + 0.004621370695531368, + 0.0020484942942857742, + 0.0012229487765580416, + 0.002145481528714299, + 0.0019355209078639746, + 0.0021470356732606888, + 0.002156923059374094, + 0.003259230637922883, + 0.0032701881136745214, + 0.005022234283387661, + 0.0024877553805708885, + 0.002119888551533222, + 0.003991090226918459, + 0.0047820559702813625, + 0.012993118725717068, + 0.004665940068662167, + 0.008694562129676342, + 0.003899508621543646, + 0.0020913765765726566, + 0.0030908870976418257, + 0.0025016821455210447, + 0.0028831956442445517, + 0.005727133248001337, + 0.0028925107326358557, + 0.005683733616024256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01822909340262413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646813031286001 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646813031286001, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646813031286001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13090814650058746, + 0.15468154847621918, + 0.16364939510822296, + 0.03179239481687546, + 0.1107817143201828, + 0.17377929389476776, + 0.07702699303627014, + 0.022633681073784828, + 0.08325787633657455, + 0.051488857716321945 + ], + "credit_map": [ + 0.011964889243245125, + 0.013142729178071022, + 0.017302220687270164, + 0.009141478687524796, + 0.024244938045740128, + 0.01912609301507473, + 0.018873121589422226, + 0.017112702131271362, + 0.012869827449321747, + 0.05080706626176834, + 0.0074196187779307365, + 0.013359314762055874, + 0.01785762421786785, + 0.03473085165023804, + 0.01013246551156044, + 0.007504807319492102, + 0.017370378598570824, + 0.05980110168457031, + 0.008728686720132828, + 0.012634605169296265, + 0.02918948605656624, + 0.018172478303313255, + 0.010735329240560532, + 0.007017356809228659, + 0.0030800425447523594, + 0.004812092520296574, + 0.004607425536960363, + 0.0037046715151518583, + 0.004085137974470854, + 0.0036315307952463627, + 0.0047822981141507626, + 0.003089203964918852, + 0.008226042613387108, + 0.03275652602314949, + 0.005181454122066498, + 0.006692170165479183, + 0.018752405419945717, + 0.02500043995678425, + 0.006126694846898317, + 0.008045997470617294, + 0.020226508378982544, + 0.012789237312972546, + 0.015871427953243256, + 0.017407169565558434, + 0.018483668565750122, + 0.04696584492921829, + 0.01969228684902191, + 0.022343169897794724, + 0.007650908548384905, + 0.0074024684727191925, + 0.00939378421753645, + 0.007998527027666569, + 0.013759525492787361, + 0.009974470362067223, + 0.011076684109866619, + 0.009770629927515984, + 0.001862101722508669, + 0.00172427820507437, + 0.002029313938692212, + 0.0012751987669616938, + 0.0037102869246155024, + 0.0032230273354798555, + 0.0035991808399558067, + 0.005210299510508776, + 0.009470885619521141, + 0.005880826152861118, + 0.0068457056768238544, + 0.008388077840209007, + 0.010710240341722965, + 0.018471334129571915, + 0.010228084400296211, + 0.013262739405035973, + 0.006652532611042261, + 0.004086350090801716, + 0.004489573184400797, + 0.003524579806253314, + 0.009553865529596806, + 0.009436647407710552, + 0.004993092268705368, + 0.008752224035561085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.004646813031286001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649740368127823 + ], + "opposite_logprob_tokens": [ + -4.758649826049805 + ], + "logprob_reward": -0.008649740368127823, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649740368127823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24760177731513977, + 0.08311565220355988, + 0.1802808791399002, + 0.03378104418516159, + 0.06009288132190704, + 0.13729090988636017, + 0.1259469836950302, + 0.02308754064142704, + 0.0618009977042675, + 0.04700133949518204 + ], + "credit_map": [ + 0.02799159847199917, + 0.021976707503199577, + 0.014639495871961117, + 0.020103221759200096, + 0.0414012148976326, + 0.042613957077264786, + 0.038208458572626114, + 0.04066712036728859, + 0.014031468890607357, + 0.013623160310089588, + 0.00863225944340229, + 0.005498900543898344, + 0.009758010506629944, + 0.014077900908887386, + 0.010374215431511402, + 0.007119736168533564, + 0.019561948254704475, + 0.0665389895439148, + 0.009968382306396961, + 0.014264981262385845, + 0.03149460256099701, + 0.020518403500318527, + 0.010444285348057747, + 0.0074893101118505, + 0.0030490243807435036, + 0.008700530044734478, + 0.0034822355955839157, + 0.003583531826734543, + 0.004906837362796068, + 0.004055099096149206, + 0.0037612351588904858, + 0.002242547459900379, + 0.005370614118874073, + 0.011039583012461662, + 0.0028154696337878704, + 0.0033295792527496815, + 0.006019512191414833, + 0.01802823692560196, + 0.005981504451483488, + 0.007508382201194763, + 0.011922301724553108, + 0.009348299354314804, + 0.018443506211042404, + 0.017024056985974312, + 0.01831343211233616, + 0.026291275396943092, + 0.01769179292023182, + 0.018256239593029022, + 0.013966882601380348, + 0.01373250037431717, + 0.016961615532636642, + 0.013176574371755123, + 0.019725220277905464, + 0.01395935658365488, + 0.018029313534498215, + 0.01639552414417267, + 0.002303738845512271, + 0.0017963696736842394, + 0.0021895787212997675, + 0.0024193956051021814, + 0.00288628158159554, + 0.002886770758777857, + 0.004960310645401478, + 0.003645094344392419, + 0.008183264173567295, + 0.005157931707799435, + 0.005318793933838606, + 0.006657800637185574, + 0.007303602993488312, + 0.012020756490528584, + 0.0080121960490942, + 0.00914665125310421, + 0.007369855418801308, + 0.003802662482485175, + 0.006745669059455395, + 0.00380616239272058, + 0.004823044408112764, + 0.008056439459323883, + 0.004846399184316397, + 0.007551105227321386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.008649740368127823 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a hand wearing a pristine white cotton glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand tightly clutching an antique brass pocket watch with Baroque scrollwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a single intense warm light source creating high contrast with deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background almost completely black and soft-focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead and does the camera remain still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207752600312233 + ], + "opposite_logprob_tokens": [ + -2.812077522277832 + ], + "logprob_reward": -0.06207752600312233, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207752600312233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19187170267105103, + 0.14736904203891754, + 0.03964919224381447, + 0.07904298603534698, + 0.08161783218383789, + 0.08178456872701645, + 0.11483954638242722, + 0.052640318870544434, + 0.1653587520122528, + 0.0458260215818882 + ], + "credit_map": [ + 0.021191468462347984, + 0.013706957921385765, + 0.012812193483114243, + 0.01860899291932583, + 0.0275966115295887, + 0.0346355140209198, + 0.03877125307917595, + 0.0245487242937088, + 0.012960479594767094, + 0.045814573764801025, + 0.005761171691119671, + 0.01279147993773222, + 0.016777778044342995, + 0.033564385026693344, + 0.012449873611330986, + 0.007249313406646252, + 0.004221076611429453, + 0.008376390673220158, + 0.0028664707206189632, + 0.003791047492995858, + 0.005040574353188276, + 0.006032160483300686, + 0.006058151833713055, + 0.0032633256632834673, + 0.006238017696887255, + 0.02358666993677616, + 0.006447445601224899, + 0.005720935761928558, + 0.011518730781972408, + 0.012706270441412926, + 0.007102935574948788, + 0.005721983499825001, + 0.009269943460822105, + 0.014774725772440434, + 0.0046755424700677395, + 0.003842244390398264, + 0.014944535680115223, + 0.019035639241337776, + 0.0071946983225643635, + 0.00788051262497902, + 0.007832897827029228, + 0.006624585948884487, + 0.006018877495080233, + 0.011914790607988834, + 0.009770316071808338, + 0.01439903024584055, + 0.011776459403336048, + 0.013447622768580914, + 0.011272860690951347, + 0.01202082447707653, + 0.012060251086950302, + 0.011611390858888626, + 0.02106402814388275, + 0.01432162057608366, + 0.016985109075903893, + 0.015503467991948128, + 0.004732092376798391, + 0.0037198346108198166, + 0.004868375137448311, + 0.004176883492618799, + 0.005286424420773983, + 0.00923917442560196, + 0.009076985530555248, + 0.011540555395185947, + 0.020933639258146286, + 0.012521608732640743, + 0.01702837646007538, + 0.017879728227853775, + 0.020154839381575584, + 0.024764519184827805, + 0.02841721847653389, + 0.023658830672502518, + 0.007492808625102043, + 0.0038771184626966715, + 0.008527005091309547, + 0.0033298309426754713, + 0.005105794407427311, + 0.008609099313616753, + 0.00399437639862299, + 0.004889992065727711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.06207752600312233 + } + ], + "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/0000028/group_9/index_tensor([1075, 1075])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_9/index_tensor([1075, 1075])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..00f8fd0b89cface571ea6c0d9827bb4c254a818e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_9/index_tensor([1075, 1075])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a low angle, slowly tilts upward with the climber, emphasizing the arduousness of the ascent and the height of the wall. The scene focuses on a single explorer, his face stern and his eyes focused and determined beneath his windswept black hair, as he strives to scale a nearly vertical, weathered section of the ancient Great Wall. He wears a dark green, windproof jacket, stained with dust and wear, and heavy, muddy hiking boots provide crucial support. The massive gray masonry is jagged, with tough moss and weeds growing in the cracks between them. The golden afterglow of the setting sun shines down at a sharp angle, casting long, dramatic shadows on the mottled wall. He carefully reached upward, his frayed leather-gloved hand, his fingers searching the weathered cracks of the stone for a firm grip. As he stabilized his center of gravity and lifted one foot onto a raised rock, a small patch of gravel and sand beneath his feet loosened under the force and slid down, echoing softly yet distinctly in the silent valley. He clung to the cold wall, each breath punctuated by condensed white air. Though his movements were labored, his rhythm was steady and filled with determination. The entire scene was stained with a warm golden hue by the setting sun, a stark contrast to the cold gray of the wall and the blue-green of the distant mountains, creating an atmosphere of loneliness, perseverance, and epic proportions.", + "seed": 20947, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_9/index_1075_seed_20947_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": -0.22269645258784293, + "vlm_logprob": -0.22269645258784293, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.22269645258784293, + "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 shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.1030474379658699 + ], + "opposite_logprob_tokens": [ + -2.7280473709106445 + ], + "logprob_reward": -0.1030474379658699, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.1030474379658699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1095346212387085, + 0.06283488124608994, + 0.06960023194551468, + 0.08314068615436554, + 0.1791187971830368, + 0.12158948928117752, + 0.1435983031988144, + 0.07460309565067291, + 0.06868915259838104, + 0.08729072660207748 + ], + "credit_map": [ + 0.019317839294672012, + 0.01986219920217991, + 0.012869484722614288, + 0.014214256778359413, + 0.010054808109998703, + 0.007722074165940285, + 0.016157468780875206, + 0.009336492046713829, + 0.007962302304804325, + 0.009313346818089485, + 0.00798342376947403, + 0.0072234198451042175, + 0.007670569233596325, + 0.005783683154731989, + 0.011046512052416801, + 0.0058516282588243484, + 0.006869818549603224, + 0.00846103485673666, + 0.00945315882563591, + 0.010298375971615314, + 0.007854286581277847, + 0.008107883855700493, + 0.012033929117023945, + 0.006521743256598711, + 0.012625095434486866, + 0.015034371986985207, + 0.012324769981205463, + 0.011104726232588291, + 0.009205417707562447, + 0.008018744178116322, + 0.008549914695322514, + 0.006277649197727442, + 0.0212527047842741, + 0.020205873996019363, + 0.039558615535497665, + 0.023524263873696327, + 0.01968638226389885, + 0.01989285834133625, + 0.020308848470449448, + 0.014689264819025993, + 0.013020453043282032, + 0.009844604879617691, + 0.02169159986078739, + 0.018231436610221863, + 0.01602977328002453, + 0.015646398067474365, + 0.014856011606752872, + 0.012269217520952225, + 0.014793765731155872, + 0.014525296166539192, + 0.02412181906402111, + 0.018980203196406364, + 0.01827387325465679, + 0.021985376253724098, + 0.016418421640992165, + 0.014499529264867306, + 0.007612057030200958, + 0.009567826986312866, + 0.013547848910093307, + 0.014005052857100964, + 0.0075716120190918446, + 0.005762089509516954, + 0.008595302700996399, + 0.007941308431327343, + 0.0050600203685462475, + 0.009749586693942547, + 0.009304721839725971, + 0.011161142028868198, + 0.004664058797061443, + 0.011066734790802002, + 0.007545447908341885, + 0.010137439705431461, + 0.009287957102060318, + 0.010047469288110733, + 0.011447400785982609, + 0.009348437190055847, + 0.009342597797513008, + 0.012399200350046158, + 0.013164245523512363, + 0.01225342508405447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.1030474379658699 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.07900188863277435 + ], + "opposite_logprob_tokens": [ + -2.5790019035339355 + ], + "logprob_reward": -0.07900188863277435, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900188863277435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1467519998550415, + 0.08741267770528793, + 0.09100861847400665, + 0.10140267759561539, + 0.11351954191923141, + 0.18453097343444824, + 0.07886794209480286, + 0.048886023461818695, + 0.06139581277966499, + 0.0862237960100174 + ], + "credit_map": [ + 0.018857572227716446, + 0.030750634148716927, + 0.023757454007864, + 0.019828621298074722, + 0.014503419399261475, + 0.01082583423703909, + 0.01807088777422905, + 0.010157559998333454, + 0.010305558331310749, + 0.01747649349272251, + 0.016005197539925575, + 0.009918474592268467, + 0.010039412416517735, + 0.006436259485781193, + 0.010064330883324146, + 0.007166942581534386, + 0.016937419772148132, + 0.02754215896129608, + 0.007377210073173046, + 0.016573436558246613, + 0.0076161897741258144, + 0.004314152989536524, + 0.005502365063875914, + 0.005145684815943241, + 0.018253054469823837, + 0.01413806527853012, + 0.015761876478791237, + 0.015274985693395138, + 0.007817389443516731, + 0.007279630750417709, + 0.014756197109818459, + 0.008121476508677006, + 0.01459646038711071, + 0.012307871133089066, + 0.015491710044443607, + 0.018620969727635384, + 0.013281230814754963, + 0.011242473497986794, + 0.015802858397364616, + 0.012175964191555977, + 0.015795323997735977, + 0.01710161752998829, + 0.029522761702537537, + 0.023198198527097702, + 0.024939822033047676, + 0.0318915881216526, + 0.0203026682138443, + 0.02177898772060871, + 0.010110193863511086, + 0.00976913794875145, + 0.012698712758719921, + 0.009535751305520535, + 0.009780583903193474, + 0.008360563777387142, + 0.011826936155557632, + 0.00678605493158102, + 0.0052461219020187855, + 0.007414272055029869, + 0.005586571060121059, + 0.008593268692493439, + 0.005445057526230812, + 0.0048411013558506966, + 0.006135983392596245, + 0.005623644217848778, + 0.006924560759216547, + 0.009849485009908676, + 0.007296718657016754, + 0.010481037199497223, + 0.005096822511404753, + 0.009313101880252361, + 0.005916280206292868, + 0.006517806090414524, + 0.009521140716969967, + 0.00963551551103592, + 0.012355797924101353, + 0.008259306661784649, + 0.00867633055895567, + 0.01329736690968275, + 0.014508835971355438, + 0.009969495236873627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.07900188863277435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.07897104322910309 + ], + "opposite_logprob_tokens": [ + -2.5789711475372314 + ], + "logprob_reward": -0.07897104322910309, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897104322910309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40054455399513245, + 0.060346394777297974, + 0.09274233877658844, + 0.0500783696770668, + 0.06409970670938492, + 0.10412237793207169, + 0.09762800484895706, + 0.0513482429087162, + 0.03821716085076332, + 0.040872856974601746 + ], + "credit_map": [ + 0.04437867924571037, + 0.07938692718744278, + 0.0673297718167305, + 0.052597615867853165, + 0.03481287881731987, + 0.03207658603787422, + 0.06011407449841499, + 0.02984800934791565, + 0.008950812742114067, + 0.008925225585699081, + 0.008678524754941463, + 0.0068577188067138195, + 0.0078034126199781895, + 0.004835057538002729, + 0.007671778090298176, + 0.0066238646395504475, + 0.010303924791514874, + 0.02090635895729065, + 0.009186687879264355, + 0.015356654301285744, + 0.007971866987645626, + 0.007790334988385439, + 0.012122665531933308, + 0.009103848598897457, + 0.009974472224712372, + 0.008968692272901535, + 0.008692807517945766, + 0.008154292590916157, + 0.0029892127495259047, + 0.003277040319517255, + 0.004565118812024593, + 0.0034567334223538637, + 0.009692194871604443, + 0.006868549156934023, + 0.01213319506496191, + 0.011869854293763638, + 0.005456665530800819, + 0.005273445975035429, + 0.007126441225409508, + 0.00567935686558485, + 0.011661223135888577, + 0.008801552467048168, + 0.019960300996899605, + 0.016603197902441025, + 0.010825475677847862, + 0.01444175373762846, + 0.00842967163771391, + 0.013399210758507252, + 0.011336660012602806, + 0.010008377023041248, + 0.010934768244624138, + 0.00884063821285963, + 0.013399121351540089, + 0.015578259713947773, + 0.015588009729981422, + 0.011942173354327679, + 0.005041212774813175, + 0.006671006791293621, + 0.006162074860185385, + 0.010610523633658886, + 0.007525265216827393, + 0.005444198846817017, + 0.00546062458306551, + 0.004433335270732641, + 0.0031344022136181593, + 0.005557318218052387, + 0.005318045616149902, + 0.006049197632819414, + 0.0032772186677902937, + 0.005242851562798023, + 0.004561931826174259, + 0.005076194182038307, + 0.004224644973874092, + 0.005553964525461197, + 0.005719585809856653, + 0.003862673882395029, + 0.003461119718849659, + 0.007854554802179337, + 0.006313389632850885, + 0.003882922464981675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.07897104322910309 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261101245880127 + ], + "opposite_logprob_tokens": [ + -0.5761101245880127 + ], + "logprob_reward": -0.8261101245880127, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261101245880127, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07318572700023651, + 0.15647977590560913, + 0.05563512444496155, + 0.04299252852797508, + 0.05635509267449379, + 0.180863156914711, + 0.2331540286540985, + 0.05719693377614021, + 0.04491129890084267, + 0.09922628104686737 + ], + "credit_map": [ + 0.008647401817142963, + 0.014918629080057144, + 0.010778842493891716, + 0.009398255497217178, + 0.006865016184747219, + 0.00644640251994133, + 0.010509973391890526, + 0.005621213000267744, + 0.024985160678625107, + 0.029203632846474648, + 0.025758912786841393, + 0.018837330862879753, + 0.01815587468445301, + 0.012600095942616463, + 0.015227209776639938, + 0.01171154621988535, + 0.004548602271825075, + 0.00805016327649355, + 0.005831350106745958, + 0.00906346831470728, + 0.004184845834970474, + 0.005927216727286577, + 0.009246347472071648, + 0.008783130906522274, + 0.006180217489600182, + 0.005627300590276718, + 0.007196853868663311, + 0.008511170744895935, + 0.004319588653743267, + 0.003769586095586419, + 0.004710022360086441, + 0.0026777859311550856, + 0.006180521100759506, + 0.008555272594094276, + 0.005988403223454952, + 0.00673580402508378, + 0.0073426044546067715, + 0.009281880222260952, + 0.00548756867647171, + 0.006783036980777979, + 0.013725983910262585, + 0.012357759289443493, + 0.027213485911488533, + 0.018523097038269043, + 0.023195089772343636, + 0.040225181728601456, + 0.020723706111311913, + 0.024898840114474297, + 0.025886237621307373, + 0.02525187097489834, + 0.027105581015348434, + 0.021325426176190376, + 0.035916052758693695, + 0.03978116437792778, + 0.031571321189403534, + 0.026316380128264427, + 0.00573069928213954, + 0.005703820381313562, + 0.006666671019047499, + 0.007887779735028744, + 0.010702013969421387, + 0.007766250520944595, + 0.007508988492190838, + 0.005230706185102463, + 0.004536319524049759, + 0.00535055436193943, + 0.006475038360804319, + 0.007995246909558773, + 0.0038913399912416935, + 0.0061377245001494884, + 0.005275654140859842, + 0.005249421577900648, + 0.010677376762032509, + 0.012072654440999031, + 0.011352894827723503, + 0.01045176386833191, + 0.01238970272243023, + 0.01681654527783394, + 0.013353856280446053, + 0.012111486867070198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.8261101245880127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.026351768523454666 + ], + "opposite_logprob_tokens": [ + -3.6513516902923584 + ], + "logprob_reward": -0.026351768523454666, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351768523454666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07204212248325348, + 0.08272818475961685, + 0.1312883198261261, + 0.04204694181680679, + 0.07333270460367203, + 0.10645434260368347, + 0.18346041440963745, + 0.05212679132819176, + 0.044759612530469894, + 0.21176058053970337 + ], + "credit_map": [ + 0.009158649481832981, + 0.01667121611535549, + 0.009402128867805004, + 0.009606041014194489, + 0.0053734928369522095, + 0.005944338161498308, + 0.009623382240533829, + 0.006262872833758593, + 0.014265967532992363, + 0.014793005771934986, + 0.01258579920977354, + 0.00957933533936739, + 0.009181130677461624, + 0.0056951516307890415, + 0.009044459089636803, + 0.007583338767290115, + 0.016756832599639893, + 0.02892833761870861, + 0.01687210611999035, + 0.021812548860907555, + 0.014195547439157963, + 0.012636176310479641, + 0.011317119002342224, + 0.008769653737545013, + 0.005866306368261576, + 0.004744929261505604, + 0.0058967601507902145, + 0.008314637467265129, + 0.0039397007785737514, + 0.0041267154738307, + 0.004998366814106703, + 0.004159525036811829, + 0.007784937974065542, + 0.012704201973974705, + 0.00951601192355156, + 0.01135919988155365, + 0.00892665982246399, + 0.008956585079431534, + 0.0062266867607831955, + 0.00785841979086399, + 0.009469319134950638, + 0.00672902213409543, + 0.01893697679042816, + 0.011012164875864983, + 0.013279997743666172, + 0.022266507148742676, + 0.010765034705400467, + 0.013995318673551083, + 0.026800446212291718, + 0.024213017895817757, + 0.022672250866889954, + 0.017533497884869576, + 0.025606082752346992, + 0.02361369878053665, + 0.02291839011013508, + 0.020103029906749725, + 0.005371346604079008, + 0.005564412102103233, + 0.006443768739700317, + 0.00831089448183775, + 0.009466884657740593, + 0.00606648251414299, + 0.006386346183717251, + 0.004516656510531902, + 0.004387119319289923, + 0.007477825973182917, + 0.006543202791363001, + 0.006413934286683798, + 0.003970163408666849, + 0.005860247649252415, + 0.004096128512173891, + 0.006010986398905516, + 0.026333196088671684, + 0.03977125138044357, + 0.02178490348160267, + 0.02040637657046318, + 0.02106877975165844, + 0.025756699964404106, + 0.03268281742930412, + 0.023956554010510445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.026351768523454666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a low angle, slowly tilts upward with the climber, emphasizing the arduousness of the ascent and the height of the wall. The scene focuses on a single explorer, his face stern and his eyes focused and determined beneath his windswept black hair, as he strives to scale a nearly vertical, weathered section of the ancient Great Wall. He wears a dark green, windproof jacket, stained with dust and wear, and heavy, muddy hiking boots provide crucial support. The massive gray masonry is jagged, with tough moss and weeds growing in the cracks between them. The golden afterglow of the setting sun shines down at a sharp angle, casting long, dramatic shadows on the mottled wall. He carefully reached upward, his frayed leather-gloved hand, his fingers searching the weathered cracks of the stone for a firm grip. As he stabilized his center of gravity and lifted one foot onto a raised rock, a small patch of gravel and sand beneath his feet loosened under the force and slid down, echoing softly yet distinctly in the silent valley. He clung to the cold wall, each breath punctuated by condensed white air. Though his movements were labored, his rhythm was steady and filled with determination. The entire scene was stained with a warm golden hue by the setting sun, a stark contrast to the cold gray of the wall and the blue-green of the distant mountains, creating an atmosphere of loneliness, perseverance, and epic proportions.", + "seed": 20948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_9/index_1075_seed_20948_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": -0.08114899266511202, + "vlm_logprob": -0.08114899266511202, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.08114899266511202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.07131951302289963 + ], + "opposite_logprob_tokens": [ + -3.446319580078125 + ], + "logprob_reward": -0.07131951302289963, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.07131951302289963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2688208818435669, + 0.1253076195716858, + 0.049844637513160706, + 0.10196078568696976, + 0.05065244063735008, + 0.1113639622926712, + 0.04524003714323044, + 0.11498308926820755, + 0.07972245663404465, + 0.05210411921143532 + ], + "credit_map": [ + 0.04778270050883293, + 0.030952846631407738, + 0.03094431757926941, + 0.02491282857954502, + 0.052266161888837814, + 0.026743602007627487, + 0.024143148213624954, + 0.031075263395905495, + 0.009444592520594597, + 0.014683037996292114, + 0.042777713388204575, + 0.016393380239605904, + 0.006695048417896032, + 0.02180165983736515, + 0.005208807066082954, + 0.008303368464112282, + 0.009168820455670357, + 0.006525526754558086, + 0.005616082809865475, + 0.005465836729854345, + 0.006049716845154762, + 0.004846590105444193, + 0.00738997058942914, + 0.0047820936888456345, + 0.010265017859637737, + 0.015155387111008167, + 0.012062320485711098, + 0.01817644014954567, + 0.018073726445436478, + 0.007088575512170792, + 0.009725203737616539, + 0.011414103209972382, + 0.005231940187513828, + 0.009311539120972157, + 0.005508757662028074, + 0.005009395536035299, + 0.008784116245806217, + 0.00395663408562541, + 0.005091315601021051, + 0.007758737076073885, + 0.03460962697863579, + 0.012101380154490471, + 0.009645994752645493, + 0.011592923663556576, + 0.010548759251832962, + 0.016331050544977188, + 0.007899212650954723, + 0.008635008707642555, + 0.005465300288051367, + 0.006110765039920807, + 0.00533183291554451, + 0.004181853495538235, + 0.007962983101606369, + 0.005608633626252413, + 0.006126741878688335, + 0.004451920744031668, + 0.010172967799007893, + 0.017981451004743576, + 0.013698693364858627, + 0.016921859234571457, + 0.02685905247926712, + 0.007780292071402073, + 0.011944450438022614, + 0.009624318219721317, + 0.009300670586526394, + 0.020126232877373695, + 0.007990853860974312, + 0.01346130482852459, + 0.009009760804474354, + 0.00582731980830431, + 0.00796954333782196, + 0.006036761216819286, + 0.009463644586503506, + 0.005042909644544125, + 0.004537427332252264, + 0.00735351862385869, + 0.004266984295099974, + 0.00851899292320013, + 0.005182921886444092, + 0.0077377138659358025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.07131951302289963 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.20156818628311157 + ], + "opposite_logprob_tokens": [ + -1.7015681266784668 + ], + "logprob_reward": -0.20156818628311157, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156818628311157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.142572820186615, + 0.3004310727119446, + 0.07975149899721146, + 0.0461510494351387, + 0.07969212532043457, + 0.04874706640839577, + 0.029345141723752022, + 0.0711771547794342, + 0.07103332877159119, + 0.13109880685806274 + ], + "credit_map": [ + 0.026381079107522964, + 0.019886592403054237, + 0.015062631107866764, + 0.020898349583148956, + 0.018501833081245422, + 0.013958144932985306, + 0.010014881379902363, + 0.017869291827082634, + 0.020971572026610374, + 0.037126775830984116, + 0.10735311359167099, + 0.04416939616203308, + 0.01079315971583128, + 0.0541214793920517, + 0.011042765341699123, + 0.014852778986096382, + 0.011999239213764668, + 0.00978355947881937, + 0.009567035362124443, + 0.00884197372943163, + 0.010275375097990036, + 0.00657041696831584, + 0.012959443032741547, + 0.009754450991749763, + 0.005631526932120323, + 0.006522520910948515, + 0.0058923703618347645, + 0.005853136070072651, + 0.0053800600580871105, + 0.0031954620499163866, + 0.0050931209698319435, + 0.008582847192883492, + 0.006777594331651926, + 0.0076074316166341305, + 0.008896901272237301, + 0.015605117194354534, + 0.007438227999955416, + 0.017634248360991478, + 0.006608216091990471, + 0.00912438239902258, + 0.01091134175658226, + 0.010102846659719944, + 0.004308884032070637, + 0.0049845753237605095, + 0.0038489948492497206, + 0.007349880412220955, + 0.0037236642092466354, + 0.0035168726462870836, + 0.0032624045852571726, + 0.004778102971613407, + 0.004848743323236704, + 0.0027939165011048317, + 0.0044996836222708225, + 0.002276686718687415, + 0.003715566825121641, + 0.0031700332183390856, + 0.009458040818572044, + 0.008056755177676678, + 0.009505377151072025, + 0.010714486241340637, + 0.011741183698177338, + 0.005690663121640682, + 0.010204815305769444, + 0.005805825814604759, + 0.007949641905725002, + 0.015115895308554173, + 0.007864847779273987, + 0.01068625133484602, + 0.007765157613903284, + 0.008786097168922424, + 0.006141279824078083, + 0.006724152248352766, + 0.03805213421583176, + 0.012336562387645245, + 0.012166302651166916, + 0.008872140198946, + 0.009033462032675743, + 0.030521061271429062, + 0.007590452674776316, + 0.012526670470833778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.20156818628311157 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.07002203911542892 + ], + "opposite_logprob_tokens": [ + -2.6950221061706543 + ], + "logprob_reward": -0.07002203911542892, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002203911542892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2853761613368988, + 0.40948864817619324, + 0.013912315480411053, + 0.022749070078134537, + 0.022455690428614616, + 0.08256667852401733, + 0.009574081748723984, + 0.061050351709127426, + 0.05551843345165253, + 0.03730856999754906 + ], + "credit_map": [ + 0.04748264327645302, + 0.02970898523926735, + 0.037584878504276276, + 0.029664363712072372, + 0.06110861152410507, + 0.02844463661313057, + 0.01797347329556942, + 0.03340855985879898, + 0.03435898199677467, + 0.0524446927011013, + 0.15547271072864532, + 0.046293146908283234, + 0.010124879889190197, + 0.0845230296254158, + 0.010944366455078125, + 0.01532683800905943, + 0.001806991989724338, + 0.0017851997399702668, + 0.0017888171132653952, + 0.0019414606504142284, + 0.0017027913127094507, + 0.001677828491665423, + 0.0014932408230379224, + 0.0017159844283014536, + 0.0026981402188539505, + 0.0029100063256919384, + 0.0028333349619060755, + 0.004887990187853575, + 0.0031626923009753227, + 0.0017070089234039187, + 0.0029277896974235773, + 0.001622107345610857, + 0.002781630028039217, + 0.002580214524641633, + 0.0029499416705220938, + 0.003983120433986187, + 0.002879169536754489, + 0.0034254153724759817, + 0.0015547630609944463, + 0.0023014352191239595, + 0.018262730911374092, + 0.014816905371844769, + 0.009054561145603657, + 0.0087510347366333, + 0.007628039922565222, + 0.009336649440228939, + 0.006574423052370548, + 0.008142328821122646, + 0.0010939199710264802, + 0.0012980663450434804, + 0.0014893197221681476, + 0.0011670718668028712, + 0.0011744055664166808, + 0.0009617864852771163, + 0.0012192631838843226, + 0.001170248375274241, + 0.004786630626767874, + 0.00883137620985508, + 0.006676960736513138, + 0.008135157637298107, + 0.014772960916161537, + 0.004490818362683058, + 0.007765010464936495, + 0.005591438151896, + 0.006906106602400541, + 0.005514926742762327, + 0.005303587298840284, + 0.016491146758198738, + 0.0032459713984280825, + 0.011370168067514896, + 0.0032079145312309265, + 0.00347860693000257, + 0.005829925183206797, + 0.006419579032808542, + 0.003939811140298843, + 0.0037684037815779448, + 0.005153960082679987, + 0.002607691567391157, + 0.004298664163798094, + 0.005290532950311899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.07002203911542892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.014192597009241581 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192597009241581, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192597009241581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10992541909217834, + 0.20095114409923553, + 0.030760688707232475, + 0.0317981131374836, + 0.15864916145801544, + 0.05030197650194168, + 0.01057808380573988, + 0.07016588747501373, + 0.1065821722149849, + 0.23028738796710968 + ], + "credit_map": [ + 0.02311006560921669, + 0.011584581807255745, + 0.015496143139898777, + 0.007760449778288603, + 0.019753703847527504, + 0.009879965335130692, + 0.00850778166204691, + 0.013832722790539265, + 0.016399817541241646, + 0.026089604943990707, + 0.07505349069833755, + 0.025033067911863327, + 0.005389050114899874, + 0.039765119552612305, + 0.005976674612611532, + 0.007244301959872246, + 0.004539662506431341, + 0.002955096308141947, + 0.004616980440914631, + 0.0028516994789242744, + 0.004203723277896643, + 0.003797288751229644, + 0.0034305877052247524, + 0.004365650936961174, + 0.004100827034562826, + 0.004808782134205103, + 0.004500182811170816, + 0.0036675967276096344, + 0.0051179854199290276, + 0.002674515824764967, + 0.003083439776673913, + 0.003844781778752804, + 0.021713146939873695, + 0.018070043995976448, + 0.022031065076589584, + 0.027437709271907806, + 0.016851074993610382, + 0.0286928191781044, + 0.011853641830384731, + 0.011999638751149178, + 0.01507823821157217, + 0.00783428456634283, + 0.003337438916787505, + 0.005124674644321203, + 0.004544591531157494, + 0.008130552247166634, + 0.0029604521114379168, + 0.003291736589744687, + 0.001679000211879611, + 0.0014312885468825698, + 0.001753499498590827, + 0.0008517883834429085, + 0.0011265079956501722, + 0.0013724048621952534, + 0.0010166724678128958, + 0.0013469199184328318, + 0.006212242878973484, + 0.011005561798810959, + 0.00849669799208641, + 0.007054678164422512, + 0.018970100209116936, + 0.004962825682014227, + 0.008090553805232048, + 0.005373221356421709, + 0.013721635565161705, + 0.01044432632625103, + 0.011168476194143295, + 0.027441436424851418, + 0.007331606931984425, + 0.02045990340411663, + 0.008397499099373817, + 0.007617282215505838, + 0.04392215609550476, + 0.0438072606921196, + 0.024717088788747787, + 0.019813982769846916, + 0.02653185836970806, + 0.022689851000905037, + 0.02219342067837715, + 0.02661176025867462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.014192597009241581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.04864262789487839 + ], + "opposite_logprob_tokens": [ + -3.048642635345459 + ], + "logprob_reward": -0.04864262789487839, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864262789487839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15663324296474457, + 0.3184430003166199, + 0.07065703719854355, + 0.05729877948760986, + 0.018834935501217842, + 0.16878393292427063, + 0.02864239178597927, + 0.07803786545991898, + 0.06063810735940933, + 0.0420306921005249 + ], + "credit_map": [ + 0.024698568508028984, + 0.017368588596582413, + 0.018215663731098175, + 0.012204084545373917, + 0.034397050738334656, + 0.01796696148812771, + 0.01283419318497181, + 0.018948141485452652, + 0.02544993720948696, + 0.04097935929894447, + 0.1216980516910553, + 0.03712308406829834, + 0.008581150323152542, + 0.06392409652471542, + 0.0091044120490551, + 0.0115829361602664, + 0.009631249122321606, + 0.007339798379689455, + 0.008048883639276028, + 0.006973204668611288, + 0.007019197102636099, + 0.011026235297322273, + 0.006236381363123655, + 0.014382085762917995, + 0.00710320333018899, + 0.00627388758584857, + 0.008334935642778873, + 0.008052167482674122, + 0.00814841315150261, + 0.004057362210005522, + 0.005726528353989124, + 0.00960228219628334, + 0.002494208747521043, + 0.002718566218391061, + 0.0024968248326331377, + 0.0024185620713979006, + 0.0025334779638797045, + 0.001750210183672607, + 0.0020148404873907566, + 0.0024082432501018047, + 0.04497537016868591, + 0.030862443149089813, + 0.014324843883514404, + 0.014851254411041737, + 0.015023086220026016, + 0.029306769371032715, + 0.010122650302946568, + 0.009317511692643166, + 0.002958166180178523, + 0.005399386864155531, + 0.004453189671039581, + 0.0022808483336120844, + 0.004895725753158331, + 0.0029613825026899576, + 0.0029859766364097595, + 0.002707716077566147, + 0.011444911360740662, + 0.008819169364869595, + 0.007536784280091524, + 0.01079209428280592, + 0.014765651896595955, + 0.008602267131209373, + 0.007759071420878172, + 0.008317911997437477, + 0.0075827608816325665, + 0.0068883588537573814, + 0.007173680700361729, + 0.014891518279910088, + 0.004058673977851868, + 0.01044524647295475, + 0.003854617476463318, + 0.005743251647800207, + 0.00997496023774147, + 0.003657939378172159, + 0.003268277505412698, + 0.006292310077697039, + 0.0032200049608945847, + 0.007491216529160738, + 0.002854179125279188, + 0.005271806847304106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.04864262789487839 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a low angle, slowly tilts upward with the climber, emphasizing the arduousness of the ascent and the height of the wall. The scene focuses on a single explorer, his face stern and his eyes focused and determined beneath his windswept black hair, as he strives to scale a nearly vertical, weathered section of the ancient Great Wall. He wears a dark green, windproof jacket, stained with dust and wear, and heavy, muddy hiking boots provide crucial support. The massive gray masonry is jagged, with tough moss and weeds growing in the cracks between them. The golden afterglow of the setting sun shines down at a sharp angle, casting long, dramatic shadows on the mottled wall. He carefully reached upward, his frayed leather-gloved hand, his fingers searching the weathered cracks of the stone for a firm grip. As he stabilized his center of gravity and lifted one foot onto a raised rock, a small patch of gravel and sand beneath his feet loosened under the force and slid down, echoing softly yet distinctly in the silent valley. He clung to the cold wall, each breath punctuated by condensed white air. Though his movements were labored, his rhythm was steady and filled with determination. The entire scene was stained with a warm golden hue by the setting sun, a stark contrast to the cold gray of the wall and the blue-green of the distant mountains, creating an atmosphere of loneliness, perseverance, and epic proportions.", + "seed": 20949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_9/index_1075_seed_20949_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": -0.05395795479416847, + "vlm_logprob": -0.05395795479416847, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.05395795479416847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.08070309460163116 + ], + "opposite_logprob_tokens": [ + -3.205703020095825 + ], + "logprob_reward": -0.08070309460163116, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.08070309460163116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09185020625591278, + 0.18813635408878326, + 0.2900504469871521, + 0.09540089964866638, + 0.05843015015125275, + 0.07398416846990585, + 0.061326075345277786, + 0.06602223962545395, + 0.0519646592438221, + 0.022834783419966698 + ], + "credit_map": [ + 0.01753152348101139, + 0.00571883050724864, + 0.022494299337267876, + 0.0108881751075387, + 0.00982304010540247, + 0.009812845848500729, + 0.007922441698610783, + 0.0076590487733483315, + 0.03128839656710625, + 0.02613051049411297, + 0.030463455244898796, + 0.018065255135297775, + 0.02892518788576126, + 0.016249848529696465, + 0.019463414326310158, + 0.01755027286708355, + 0.04685633257031441, + 0.02239564061164856, + 0.04148903116583824, + 0.026464972645044327, + 0.042097046971321106, + 0.058020807802677155, + 0.02287730760872364, + 0.029849303886294365, + 0.016434984281659126, + 0.016180967912077904, + 0.018452800810337067, + 0.010424401611089706, + 0.010030961595475674, + 0.008406751789152622, + 0.006924771703779697, + 0.008545259945094585, + 0.006698875222355127, + 0.007873909547924995, + 0.007307781372219324, + 0.006403542589396238, + 0.006749843247234821, + 0.010106117464601994, + 0.007602959871292114, + 0.00568711943924427, + 0.005257318262010813, + 0.00887264683842659, + 0.006105497479438782, + 0.011906305328011513, + 0.014060513116419315, + 0.008321604691445827, + 0.008942208252847195, + 0.010518074966967106, + 0.006822456605732441, + 0.011219736188650131, + 0.00734257185831666, + 0.009122379124164581, + 0.006629038602113724, + 0.004812020342797041, + 0.00577317550778389, + 0.009604699909687042, + 0.0049882009625434875, + 0.008067362941801548, + 0.006886879913508892, + 0.007164617069065571, + 0.016395995393395424, + 0.00975535623729229, + 0.0071403696201741695, + 0.005623459350317717, + 0.004484078381210566, + 0.006951187737286091, + 0.008815078996121883, + 0.006655436009168625, + 0.007221282925456762, + 0.008734899573028088, + 0.004359674174338579, + 0.004743017256259918, + 0.0019616298377513885, + 0.0026585697196424007, + 0.003697932930663228, + 0.00420480128377676, + 0.0019910612609237432, + 0.0027322403620928526, + 0.0023896535858511925, + 0.0031988946720957756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.08070309460163116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.11295150965452194 + ], + "opposite_logprob_tokens": [ + -2.2379515171051025 + ], + "logprob_reward": -0.11295150965452194, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295150965452194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07711143046617508, + 0.09781119227409363, + 0.06366048753261566, + 0.17160311341285706, + 0.09508901089429855, + 0.264376163482666, + 0.08981378376483917, + 0.07057427614927292, + 0.03260942921042442, + 0.037351127713918686 + ], + "credit_map": [ + 0.013409730978310108, + 0.008883330971002579, + 0.013542553409934044, + 0.01007069181650877, + 0.008304806426167488, + 0.009198963642120361, + 0.007292189635336399, + 0.00640915846452117, + 0.014101307839155197, + 0.01119202096015215, + 0.019272761419415474, + 0.011225029826164246, + 0.016991669312119484, + 0.008092109113931656, + 0.007872010581195354, + 0.009064286947250366, + 0.009169358760118484, + 0.00781753845512867, + 0.010312969796359539, + 0.0068292561918497086, + 0.006067775189876556, + 0.008425862528383732, + 0.004626729059964418, + 0.010410996153950691, + 0.028089681640267372, + 0.03018137812614441, + 0.031026501208543777, + 0.01474145706743002, + 0.016744518652558327, + 0.019549822434782982, + 0.01547759398818016, + 0.01579216867685318, + 0.008672911673784256, + 0.0190429724752903, + 0.011947491206228733, + 0.012569407001137733, + 0.014888305217027664, + 0.008170727640390396, + 0.010597982443869114, + 0.009199218824505806, + 0.019106091931462288, + 0.027395060285925865, + 0.02169862762093544, + 0.029205074533820152, + 0.07351046055555344, + 0.027641180902719498, + 0.024611691012978554, + 0.04120795801281929, + 0.008344084955751896, + 0.018428027629852295, + 0.00969714019447565, + 0.013332678005099297, + 0.01199738122522831, + 0.0075945197604596615, + 0.009248531423509121, + 0.011171413585543633, + 0.0060771252028644085, + 0.008537122048437595, + 0.011675434187054634, + 0.01282892469316721, + 0.009310859255492687, + 0.010067632421851158, + 0.005319308023899794, + 0.0067578707821667194, + 0.003012021305039525, + 0.0026985546573996544, + 0.0049097067676484585, + 0.0037812581285834312, + 0.0024435240775346756, + 0.0061996676959097385, + 0.004667461849749088, + 0.00489723589271307, + 0.00367600005120039, + 0.004190512932837009, + 0.009507421404123306, + 0.005280937068164349, + 0.002741951262578368, + 0.004930031951516867, + 0.002937622833997011, + 0.004086650907993317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.11295150965452194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.0067411623895168304 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.0067411623895168304, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067411623895168304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09390959143638611, + 0.25185397267341614, + 0.04680028185248375, + 0.10448487848043442, + 0.07552094757556915, + 0.099921815097332, + 0.07862135767936707, + 0.08609265089035034, + 0.09045077860355377, + 0.07234374433755875 + ], + "credit_map": [ + 0.015367632731795311, + 0.009642470628023148, + 0.014592149294912815, + 0.014953884296119213, + 0.008818302303552628, + 0.009652205742895603, + 0.011089233681559563, + 0.00979371927678585, + 0.04181971028447151, + 0.03220370411872864, + 0.03740644082427025, + 0.025429951027035713, + 0.04179444909095764, + 0.02285890467464924, + 0.02313278056681156, + 0.027208009734749794, + 0.009618923999369144, + 0.005151818040758371, + 0.00695261312648654, + 0.004957348108291626, + 0.005834151990711689, + 0.006125309970229864, + 0.004178836941719055, + 0.0039812782779335976, + 0.01267634890973568, + 0.014824816957116127, + 0.0148688405752182, + 0.010765803046524525, + 0.010593199171125889, + 0.017255792394280434, + 0.013081450015306473, + 0.010418620891869068, + 0.007388468831777573, + 0.016308311372995377, + 0.008785360492765903, + 0.012226175516843796, + 0.008516447618603706, + 0.0053090257570147514, + 0.008947541937232018, + 0.008039616048336029, + 0.009073963388800621, + 0.011937370523810387, + 0.01131515670567751, + 0.013423674739897251, + 0.018375644460320473, + 0.010520256124436855, + 0.011988009326159954, + 0.013287732377648354, + 0.007811675313860178, + 0.01403745450079441, + 0.014265146106481552, + 0.01159420982003212, + 0.006049418356269598, + 0.008681324310600758, + 0.006857970729470253, + 0.009324159473180771, + 0.007537617348134518, + 0.015815511345863342, + 0.017251720651984215, + 0.01286559458822012, + 0.009768838994204998, + 0.008218885399401188, + 0.008229648694396019, + 0.006404826417565346, + 0.00993902888149023, + 0.009177603758871555, + 0.014971345663070679, + 0.007800049614161253, + 0.009262324310839176, + 0.01713758334517479, + 0.007954289205372334, + 0.01420854963362217, + 0.008641884662210941, + 0.009117236360907555, + 0.014201100915670395, + 0.01207585446536541, + 0.00512305274605751, + 0.011641334742307663, + 0.005829911213368177, + 0.005713366903364658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.0067411623895168304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.026342712342739105 + ], + "opposite_logprob_tokens": [ + -3.6513426303863525 + ], + "logprob_reward": -0.026342712342739105, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342712342739105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1449267864227295, + 0.21107575297355652, + 0.026775848120450974, + 0.12742209434509277, + 0.08944625407457352, + 0.1559852659702301, + 0.01373849343508482, + 0.04928018897771835, + 0.09564488381147385, + 0.08570442348718643 + ], + "credit_map": [ + 0.032676249742507935, + 0.007517764810472727, + 0.04367053881287575, + 0.01464496273547411, + 0.012689427472651005, + 0.01689646951854229, + 0.00856424868106842, + 0.008267131634056568, + 0.03398965299129486, + 0.025864891707897186, + 0.030300749465823174, + 0.021285030990839005, + 0.035063792020082474, + 0.019417734816670418, + 0.020407196134328842, + 0.0247467253357172, + 0.0035312275867909193, + 0.00258646160364151, + 0.003507988527417183, + 0.003712036181241274, + 0.0031465543434023857, + 0.0041184695437550545, + 0.003452522214502096, + 0.0027205871883779764, + 0.021874791011214256, + 0.02176649309694767, + 0.026771491393446922, + 0.010736098513007164, + 0.013664153404533863, + 0.011085905134677887, + 0.010666798800230026, + 0.01085637416690588, + 0.009255390614271164, + 0.012722634710371494, + 0.010781902819871902, + 0.007879410870373249, + 0.013567452318966389, + 0.013988178223371506, + 0.008833680301904678, + 0.012417607009410858, + 0.010114439763128757, + 0.015210750512778759, + 0.010290605016052723, + 0.012570160441100597, + 0.04853788763284683, + 0.01886701211333275, + 0.015540294349193573, + 0.024854116141796112, + 0.001441049505956471, + 0.002283612033352256, + 0.0015020687133073807, + 0.0017152009531855583, + 0.0016587675781920552, + 0.0017813090234994888, + 0.0015845239395275712, + 0.0017719622701406479, + 0.00480540981516242, + 0.007649289909750223, + 0.00659496383741498, + 0.005674134474247694, + 0.0060951076447963715, + 0.007240844424813986, + 0.00643780967220664, + 0.004782628733664751, + 0.011790269985795021, + 0.008251512423157692, + 0.010775325819849968, + 0.008258850313723087, + 0.008894553408026695, + 0.021017063409090042, + 0.00828382559120655, + 0.018373481929302216, + 0.009234989993274212, + 0.01046602800488472, + 0.016632482409477234, + 0.014016947709023952, + 0.006607391405850649, + 0.014084872789680958, + 0.0072582783177495, + 0.007403429597616196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.026342712342739105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.04305129498243332 + ], + "opposite_logprob_tokens": [ + -3.168051242828369 + ], + "logprob_reward": -0.04305129498243332, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305129498243332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16364262998104095, + 0.5547090172767639, + 0.04248279333114624, + 0.03359423577785492, + 0.02938787266612053, + 0.017313962802290916, + 0.06906642019748688, + 0.014800061471760273, + 0.028177140280604362, + 0.04682581499218941 + ], + "credit_map": [ + 0.03371131792664528, + 0.011163054965436459, + 0.04399862885475159, + 0.01600356586277485, + 0.014345891773700714, + 0.021740825846791267, + 0.013814671896398067, + 0.008864684961736202, + 0.07794656604528427, + 0.05473165586590767, + 0.08426813036203384, + 0.07326222956180573, + 0.11212549358606339, + 0.049523621797561646, + 0.06187712028622627, + 0.04097424075007439, + 0.0037516949232667685, + 0.003513636766001582, + 0.005882425233721733, + 0.003572052577510476, + 0.006197137292474508, + 0.008857566863298416, + 0.0036623827181756496, + 0.007045896723866463, + 0.004163590259850025, + 0.0034925902727991343, + 0.004796675406396389, + 0.004072547424584627, + 0.004017125815153122, + 0.00703374482691288, + 0.0028597568161785603, + 0.0031582035589963198, + 0.002812845865264535, + 0.003213786520063877, + 0.004869371652603149, + 0.004091725219041109, + 0.0037453616969287395, + 0.0034823082387447357, + 0.004615976009517908, + 0.002556496299803257, + 0.0019575534388422966, + 0.0017626225017011166, + 0.001269904663786292, + 0.0027066683396697044, + 0.0030057274270802736, + 0.002016232116147876, + 0.002378540812060237, + 0.0022167135030031204, + 0.007794019766151905, + 0.010823164135217667, + 0.006566931959241629, + 0.007893534377217293, + 0.009179466404020786, + 0.009898326359689236, + 0.00525211775675416, + 0.011658859439194202, + 0.0013115935726091266, + 0.002156593604013324, + 0.0017646020278334618, + 0.0016542816301807761, + 0.0018309683073312044, + 0.0028556871693581343, + 0.0016137106576934457, + 0.0016126236878335476, + 0.003744307206943631, + 0.002463937969878316, + 0.005451648961752653, + 0.002971279202029109, + 0.001738912076689303, + 0.006498451344668865, + 0.0017628204077482224, + 0.0035457832273095846, + 0.006095240358263254, + 0.0055182804353535175, + 0.005333007778972387, + 0.0040356856770813465, + 0.0031937926542013884, + 0.009317740797996521, + 0.0036180231254547834, + 0.009714045561850071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04305129498243332 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a low angle, slowly tilts upward with the climber, emphasizing the arduousness of the ascent and the height of the wall. The scene focuses on a single explorer, his face stern and his eyes focused and determined beneath his windswept black hair, as he strives to scale a nearly vertical, weathered section of the ancient Great Wall. He wears a dark green, windproof jacket, stained with dust and wear, and heavy, muddy hiking boots provide crucial support. The massive gray masonry is jagged, with tough moss and weeds growing in the cracks between them. The golden afterglow of the setting sun shines down at a sharp angle, casting long, dramatic shadows on the mottled wall. He carefully reached upward, his frayed leather-gloved hand, his fingers searching the weathered cracks of the stone for a firm grip. As he stabilized his center of gravity and lifted one foot onto a raised rock, a small patch of gravel and sand beneath his feet loosened under the force and slid down, echoing softly yet distinctly in the silent valley. He clung to the cold wall, each breath punctuated by condensed white air. Though his movements were labored, his rhythm was steady and filled with determination. The entire scene was stained with a warm golden hue by the setting sun, a stark contrast to the cold gray of the wall and the blue-green of the distant mountains, creating an atmosphere of loneliness, perseverance, and epic proportions.", + "seed": 20950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_9/index_1075_seed_20950_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": -0.04447676669806242, + "vlm_logprob": -0.04447676669806242, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04447676669806242, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.08088605850934982 + ], + "opposite_logprob_tokens": [ + -3.705886125564575 + ], + "logprob_reward": -0.08088605850934982, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.08088605850934982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23160360753536224, + 0.06928073614835739, + 0.1216353252530098, + 0.06421398371458054, + 0.04718708619475365, + 0.0717032179236412, + 0.07688982039690018, + 0.15045633912086487, + 0.082869753241539, + 0.08416011184453964 + ], + "credit_map": [ + 0.02531445398926735, + 0.02731040120124817, + 0.03914313763380051, + 0.019988182932138443, + 0.019567452371120453, + 0.04543958604335785, + 0.027811788022518158, + 0.027028638869524002, + 0.005976896733045578, + 0.007125990465283394, + 0.00975465402007103, + 0.0068206824362277985, + 0.005733294878154993, + 0.010839820839464664, + 0.006664942484349012, + 0.016364462673664093, + 0.018744828179478645, + 0.010996699333190918, + 0.03166959062218666, + 0.01641712337732315, + 0.009401250630617142, + 0.0077564953826367855, + 0.010361095890402794, + 0.01628824695944786, + 0.006006546318531036, + 0.009077976457774639, + 0.007564350962638855, + 0.009113702923059464, + 0.008498710580170155, + 0.006751141045242548, + 0.007520914543420076, + 0.00968064647167921, + 0.004923803731799126, + 0.009971770457923412, + 0.0051158820278942585, + 0.006530593149363995, + 0.005763056688010693, + 0.00489591620862484, + 0.005676041357219219, + 0.004310025833547115, + 0.005665482021868229, + 0.014470342546701431, + 0.00860647764056921, + 0.006513408850878477, + 0.016656765714287758, + 0.007783389650285244, + 0.007473733276128769, + 0.004533626139163971, + 0.007026343606412411, + 0.00748108234256506, + 0.008446456864476204, + 0.013038251549005508, + 0.007594027556478977, + 0.011616569012403488, + 0.011772103607654572, + 0.009914998896420002, + 0.015576284378767014, + 0.02802360989153385, + 0.020191622897982597, + 0.015803437680006027, + 0.017337653785943985, + 0.02203798107802868, + 0.014682180248200893, + 0.016803579404950142, + 0.00938038993626833, + 0.016612695530056953, + 0.008861563168466091, + 0.010784771293401718, + 0.00830209068953991, + 0.00830186065286398, + 0.01114931982010603, + 0.009477069601416588, + 0.011296676471829414, + 0.012445620261132717, + 0.01516637485474348, + 0.0075547718442976475, + 0.0073135304264724255, + 0.011570948176085949, + 0.01008775644004345, + 0.00872444175183773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.08088605850934982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.04866817593574524 + ], + "opposite_logprob_tokens": [ + -3.048668146133423 + ], + "logprob_reward": -0.04866817593574524, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866817593574524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09579280018806458, + 0.10759975016117096, + 0.07077311724424362, + 0.18447518348693848, + 0.03181319683790207, + 0.06832046806812286, + 0.04034513235092163, + 0.27110958099365234, + 0.06638336926698685, + 0.06338734179735184 + ], + "credit_map": [ + 0.008203353732824326, + 0.011704365722835064, + 0.011944293975830078, + 0.009519370272755623, + 0.0073596215806901455, + 0.02253188192844391, + 0.011151899583637714, + 0.01337801106274128, + 0.009165571071207523, + 0.009544813074171543, + 0.00906517542898655, + 0.011505421251058578, + 0.008225098252296448, + 0.018269946798682213, + 0.010935097932815552, + 0.030888622626662254, + 0.0137369679287076, + 0.007872921414673328, + 0.02227395586669445, + 0.00503894267603755, + 0.006175400223582983, + 0.0042422921396791935, + 0.0046015894040465355, + 0.006831055507063866, + 0.011690269224345684, + 0.032004617154598236, + 0.012098995968699455, + 0.03373856097459793, + 0.024842876940965652, + 0.02569732815027237, + 0.02563789114356041, + 0.018764642998576164, + 0.0032778531312942505, + 0.004280691500753164, + 0.0025140177458524704, + 0.005774197168648243, + 0.0036429755855351686, + 0.003456210484728217, + 0.005345124285668135, + 0.003522127866744995, + 0.006012861151248217, + 0.0104560237377882, + 0.00725132878869772, + 0.008807784877717495, + 0.014005275443196297, + 0.00842279102653265, + 0.008050783537328243, + 0.005313620902597904, + 0.0048563359305262566, + 0.003798285499215126, + 0.0033321843948215246, + 0.006848127115517855, + 0.0040422226302325726, + 0.006318769417703152, + 0.006709217093884945, + 0.004439990501850843, + 0.03936116397380829, + 0.04496648535132408, + 0.04327916353940964, + 0.03221585229039192, + 0.02837592363357544, + 0.02655748464167118, + 0.024945246055722237, + 0.03140827640891075, + 0.007859110832214355, + 0.015925435349345207, + 0.012663489207625389, + 0.006886181887239218, + 0.005890960339456797, + 0.005901383236050606, + 0.005827265791594982, + 0.005429541226476431, + 0.008134974166750908, + 0.009254134260118008, + 0.009753051213920116, + 0.007155546918511391, + 0.006287186872214079, + 0.006939386948943138, + 0.008214309811592102, + 0.007648753933608532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.04866817593574524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.014195770025253296 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195770025253296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195770025253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12519796192646027, + 0.059105124324560165, + 0.04384109377861023, + 0.08150820434093475, + 0.07215957343578339, + 0.07215163856744766, + 0.05167659744620323, + 0.2812976837158203, + 0.029890960082411766, + 0.1831711232662201 + ], + "credit_map": [ + 0.010171967558562756, + 0.009305206127464771, + 0.015703776851296425, + 0.016656482592225075, + 0.014736214652657509, + 0.02342778444290161, + 0.019532207399606705, + 0.015664322301745415, + 0.0061902934685349464, + 0.006053524557501078, + 0.0076119243167340755, + 0.009015442803502083, + 0.006029350683093071, + 0.009800609201192856, + 0.008338254876434803, + 0.006065722089260817, + 0.004915360826998949, + 0.005647946614772081, + 0.0041270325891673565, + 0.0057412199676036835, + 0.005638548173010349, + 0.007832211442291737, + 0.005736733321100473, + 0.0042020408436656, + 0.004727013409137726, + 0.013367016799747944, + 0.005004230886697769, + 0.015811985358595848, + 0.011944599449634552, + 0.011396383866667747, + 0.010967027395963669, + 0.0082899434491992, + 0.006825597491115332, + 0.014917551539838314, + 0.0073218937031924725, + 0.009867014363408089, + 0.00886860303580761, + 0.008008051663637161, + 0.01118775550276041, + 0.0051631079986691475, + 0.0063115255907177925, + 0.012572546489536762, + 0.008484777994453907, + 0.008110172115266323, + 0.016841402277350426, + 0.0059152389876544476, + 0.00811506062746048, + 0.005800911691039801, + 0.00863201916217804, + 0.004283802583813667, + 0.0065469443798065186, + 0.005850688554346561, + 0.004664827138185501, + 0.007995881140232086, + 0.00548890046775341, + 0.008213532157242298, + 0.029241936281323433, + 0.05359085276722908, + 0.027676010504364967, + 0.03494056686758995, + 0.035044148564338684, + 0.0421801321208477, + 0.026712777093052864, + 0.03191128000617027, + 0.0030079262796789408, + 0.005123194772750139, + 0.0044821021147072315, + 0.002865808317437768, + 0.0024789178278297186, + 0.0033651411067694426, + 0.0038943614345043898, + 0.004673509392887354, + 0.02794802188873291, + 0.01849222742021084, + 0.0553237609565258, + 0.011617699638009071, + 0.00960850901901722, + 0.01148291863501072, + 0.016269657760858536, + 0.032428331673145294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.014195770025253296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.008634022437036037 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634022437036037, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634022437036037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1958756446838379, + 0.14526085555553436, + 0.13949525356292725, + 0.05239380896091461, + 0.03363749012351036, + 0.06592495739459991, + 0.04448944330215454, + 0.04793032631278038, + 0.04158524051308632, + 0.2334069460630417 + ], + "credit_map": [ + 0.018909720703959465, + 0.012843970209360123, + 0.018364498391747475, + 0.01895781233906746, + 0.02230926789343357, + 0.054025501012802124, + 0.024606486782431602, + 0.025858383625745773, + 0.012208039872348309, + 0.009639664553105831, + 0.01515235472470522, + 0.010270536877214909, + 0.013819818384945393, + 0.033615559339523315, + 0.010777818970382214, + 0.03977707028388977, + 0.011436705477535725, + 0.012341288849711418, + 0.009879929013550282, + 0.014037738554179668, + 0.02395646460354328, + 0.0285740215331316, + 0.02188042923808098, + 0.017388662323355675, + 0.004571943543851376, + 0.00598797807469964, + 0.004224032163619995, + 0.010407086461782455, + 0.005749765317887068, + 0.008236754685640335, + 0.007421619258821011, + 0.005794628988951445, + 0.002651509828865528, + 0.005181735381484032, + 0.0024487271439284086, + 0.004466474521905184, + 0.004400673788040876, + 0.004920830950140953, + 0.005697337910532951, + 0.0038702015299350023, + 0.007630787789821625, + 0.010720648802816868, + 0.007078077178448439, + 0.007589471992105246, + 0.0095753725618124, + 0.0083481939509511, + 0.007522340398281813, + 0.0074600656516849995, + 0.007694634143263102, + 0.004122297745198011, + 0.0070379371754825115, + 0.003971119876950979, + 0.004298360086977482, + 0.007415733765810728, + 0.004294877871870995, + 0.005654483567923307, + 0.003617357462644577, + 0.007080005947500467, + 0.005741430446505547, + 0.005156693980097771, + 0.005846014712005854, + 0.008302326314151287, + 0.0058890460059046745, + 0.006297444924712181, + 0.003198238555341959, + 0.009392104111611843, + 0.005345489829778671, + 0.004868978168815374, + 0.004091727547347546, + 0.004662179388105869, + 0.005231213290244341, + 0.004795307293534279, + 0.031433410942554474, + 0.01975139230489731, + 0.06219139322638512, + 0.013726630248129368, + 0.016436798498034477, + 0.015216781757771969, + 0.030950289219617844, + 0.04370025545358658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008634022437036037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a medium shot framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while following the climber's ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the climber wearing a dark green windproof jacket and heavy muddy hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wall made of massive gray masonry with jagged edges, moss, and weeds in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a golden afterglow from a setting sun 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": [ + -0.0699998065829277 + ], + "opposite_logprob_tokens": [ + -2.6949996948242188 + ], + "logprob_reward": -0.0699998065829277, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0699998065829277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12759947776794434, + 0.028059756383299828, + 0.0740615576505661, + 0.08929244428873062, + 0.02612108737230301, + 0.030242610722780228, + 0.17542316019535065, + 0.22059662640094757, + 0.0523524172604084, + 0.17625084519386292 + ], + "credit_map": [ + 0.009837052784860134, + 0.011788303032517433, + 0.020564982667565346, + 0.013901891186833382, + 0.012303821742534637, + 0.02583429403603077, + 0.017838647589087486, + 0.015530486591160297, + 0.0024855502415448427, + 0.001705958042293787, + 0.0026541482657194138, + 0.0020813816227018833, + 0.0021130512468516827, + 0.005474934354424477, + 0.002553566126152873, + 0.008991165086627007, + 0.012093526311218739, + 0.00400868384167552, + 0.02347029373049736, + 0.004890527110546827, + 0.0047681150026619434, + 0.005376813001930714, + 0.005410989746451378, + 0.014042611233890057, + 0.01312724594026804, + 0.009267762303352356, + 0.009938830509781837, + 0.01453989278525114, + 0.009569180198013783, + 0.00965593010187149, + 0.011051618494093418, + 0.012141979299485683, + 0.002410542918369174, + 0.004911316558718681, + 0.002090248977765441, + 0.003856354858726263, + 0.0035294871777296066, + 0.002967611188068986, + 0.0037580179050564766, + 0.002597506856545806, + 0.0022248856257647276, + 0.00469097588211298, + 0.003043775213882327, + 0.0038115938659757376, + 0.0064293560571968555, + 0.003355639521032572, + 0.0033001613337546587, + 0.0033862253185361624, + 0.020273195579648018, + 0.010633137077093124, + 0.017423568293452263, + 0.015517902560532093, + 0.01501352246850729, + 0.043277185410261154, + 0.02596106566488743, + 0.02732357569038868, + 0.03206635266542435, + 0.03317106515169144, + 0.035557497292757034, + 0.02462884970009327, + 0.02352084405720234, + 0.023066259920597076, + 0.021232057362794876, + 0.027353696525096893, + 0.004815947730094194, + 0.01046788226813078, + 0.006568429991602898, + 0.005773280281573534, + 0.005196872167289257, + 0.007454092614352703, + 0.005627899896353483, + 0.006448009982705116, + 0.029820609837770462, + 0.01880771480500698, + 0.05898195877671242, + 0.009496159851551056, + 0.006957237143069506, + 0.01173553429543972, + 0.015309847891330719, + 0.02514178305864334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0699998065829277 + } + ], + "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/0000052/group_0/index_tensor([5842, 5842])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_0/index_tensor([5842, 5842])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c6136d109af1da37472a5f90a6d1f985d5ee9705 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_0/index_tensor([5842, 5842])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, intimate close-up captures a glowing paper lantern, perched precariously on the edge of a dark, weathered wooden ledge. The wood's distinct grain shimmers with a smooth, damp sheen from the humid air, capturing and reflecting the lantern's warm, golden glow. The lantern itself, made of fragile, translucent rice paper laid over a delicate bamboo frame, emanates a soft, flickering light from within. The lantern begins to slide slowly and inexorably, its base gently scraping against the damp surface. Then, the camera begins a steady, deliberate downward tilt, carefully tracking the object as it loses its grip and tumbles down the ledge. Its fragile body descends in a silent, graceful descent, plunging headfirst into a thick, swirling bank of night fog below. The warm point of light flickers in and out as it falls, casting an ethereal halo through the dense fog before being completely swallowed by the profound darkness, creating a poignant and eerie atmosphere.", + "seed": 17139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_0/index_5842_seed_17139_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": -0.01857706531882286, + "vlm_logprob": -0.01857706531882286, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.01857706531882286, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02922574244439602 + ], + "opposite_logprob_tokens": [ + -4.029225826263428 + ], + "logprob_reward": -0.02922574244439602, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02922574244439602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23820286989212036, + 0.08475913107395172, + 0.057543907314538956, + 0.04667206108570099, + 0.12301860004663467, + 0.09688384085893631, + 0.0582607164978981, + 0.1335364133119583, + 0.049253955483436584, + 0.1118684783577919 + ], + "credit_map": [ + 0.027004333212971687, + 0.015171296894550323, + 0.04095499962568283, + 0.043046046048402786, + 0.015207418240606785, + 0.04297051206231117, + 0.027577729895710945, + 0.026270531117916107, + 0.013275545090436935, + 0.005642513744533062, + 0.012885182164609432, + 0.008636591956019402, + 0.005486032925546169, + 0.01917291432619095, + 0.010872083716094494, + 0.008788270875811577, + 0.0059831403195858, + 0.003128218697384, + 0.015094323083758354, + 0.0038312841206789017, + 0.0027698299381881952, + 0.015846652910113335, + 0.004605320747941732, + 0.006285139359533787, + 0.005643395707011223, + 0.004807288758456707, + 0.0076628560200333595, + 0.007245101500302553, + 0.003292680950835347, + 0.008661179803311825, + 0.005508452653884888, + 0.0038511077873408794, + 0.0162254199385643, + 0.020711656659841537, + 0.021280905231833458, + 0.01531897857785225, + 0.009725607000291348, + 0.020638156682252884, + 0.00992038194090128, + 0.009197496809065342, + 0.007828367874026299, + 0.007338893599808216, + 0.008203797042369843, + 0.014148012734949589, + 0.007529099937528372, + 0.025024281814694405, + 0.020355314016342163, + 0.0064560784958302975, + 0.003979782573878765, + 0.0059842015616595745, + 0.004153660032898188, + 0.011820188723504543, + 0.009953242726624012, + 0.0037226693239063025, + 0.013776135630905628, + 0.004870836157351732, + 0.011889386922121048, + 0.027207186445593834, + 0.009082815609872341, + 0.02175806649029255, + 0.018317900598049164, + 0.017212359234690666, + 0.019999397918581963, + 0.008069301955401897, + 0.006281533278524876, + 0.004664345644414425, + 0.00515257753431797, + 0.006677794270217419, + 0.005703797098249197, + 0.0037800741847604513, + 0.011333558708429337, + 0.005660278722643852, + 0.008838978596031666, + 0.01493771094828844, + 0.014482693746685982, + 0.02483403868973255, + 0.01264149509370327, + 0.010727321729063988, + 0.014204386621713638, + 0.01120186410844326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.02922574244439602 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642790675163269 + ], + "opposite_logprob_tokens": [ + -4.88264274597168 + ], + "logprob_reward": -0.007642790675163269, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642790675163269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10805562138557434, + 0.17475390434265137, + 0.03553443029522896, + 0.04067809134721756, + 0.12169477343559265, + 0.08197071403265, + 0.08237152546644211, + 0.25411251187324524, + 0.037061646580696106, + 0.06376677751541138 + ], + "credit_map": [ + 0.012984094209969044, + 0.006172930356115103, + 0.016644923016428947, + 0.019301578402519226, + 0.005722492467612028, + 0.02229183539748192, + 0.011557996273040771, + 0.013379769399762154, + 0.02720416523516178, + 0.011796841397881508, + 0.03898762911558151, + 0.024378539994359016, + 0.006730922497808933, + 0.03226879611611366, + 0.010610167868435383, + 0.02277684025466442, + 0.003962719812989235, + 0.0018428004113957286, + 0.010158066637814045, + 0.0029718782752752304, + 0.0012962946202605963, + 0.010255786590278149, + 0.0019350269576534629, + 0.003111855359748006, + 0.004646230023354292, + 0.006257394328713417, + 0.007108235266059637, + 0.006559137254953384, + 0.0024856769014149904, + 0.0062218341045081615, + 0.0033249950502067804, + 0.00407458795234561, + 0.019782472401857376, + 0.011658878065645695, + 0.038063012063503265, + 0.013865143060684204, + 0.006276690401136875, + 0.01702435128390789, + 0.006670056842267513, + 0.008354170247912407, + 0.004898979794234037, + 0.012308242730796337, + 0.0047381226904690266, + 0.011025833897292614, + 0.006400371436029673, + 0.009648019447922707, + 0.020669525489211082, + 0.012281622737646103, + 0.0037822979502379894, + 0.005452033132314682, + 0.0036356449127197266, + 0.01332137081772089, + 0.012389381416141987, + 0.0045673539862036705, + 0.02681252360343933, + 0.012410921044647694, + 0.021939272060990334, + 0.05421411991119385, + 0.011890550144016743, + 0.03270045667886734, + 0.04906308278441429, + 0.044491253793239594, + 0.026280613616108894, + 0.013533172197639942, + 0.003052771557122469, + 0.004045861307531595, + 0.0032338674645870924, + 0.0063849338330328465, + 0.00442121597006917, + 0.003883002558723092, + 0.008097883313894272, + 0.003942112438380718, + 0.00393696827813983, + 0.010831035673618317, + 0.006826417520642281, + 0.013803355395793915, + 0.007983426563441753, + 0.005409968085587025, + 0.00999881885945797, + 0.0049767945893108845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.007642790675163269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060040831565857 + ], + "opposite_logprob_tokens": [ + -3.8956003189086914 + ], + "logprob_reward": -0.02060040831565857, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060040831565857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09597230702638626, + 0.04545906186103821, + 0.04550224915146828, + 0.0689452588558197, + 0.06723061949014664, + 0.17557840049266815, + 0.07233814895153046, + 0.26498308777809143, + 0.08776138722896576, + 0.07622946053743362 + ], + "credit_map": [ + 0.010934293270111084, + 0.003884737379848957, + 0.018477829173207283, + 0.012847636826336384, + 0.005162706132978201, + 0.020044978708028793, + 0.005229802802205086, + 0.019390325993299484, + 0.006504775956273079, + 0.0035271213855594397, + 0.007670678663998842, + 0.007906838320195675, + 0.0020413824822753668, + 0.007169664371758699, + 0.0043866331689059734, + 0.00625196797773242, + 0.005575284827500582, + 0.005561263300478458, + 0.010332098230719566, + 0.004282485693693161, + 0.001723719760775566, + 0.011449773795902729, + 0.0024497106205672026, + 0.004127910360693932, + 0.008343423716723919, + 0.006879297085106373, + 0.012223165482282639, + 0.012490827590227127, + 0.004722420126199722, + 0.010583716444671154, + 0.007843296974897385, + 0.005859110038727522, + 0.011232658289372921, + 0.007826494984328747, + 0.010387217625975609, + 0.011106743477284908, + 0.004888630472123623, + 0.009157189168035984, + 0.0057397023774683475, + 0.0068919844925403595, + 0.016091957688331604, + 0.018957341089844704, + 0.018646271899342537, + 0.02010861225426197, + 0.014960070140659809, + 0.051417019218206406, + 0.026110587641596794, + 0.009286539629101753, + 0.0040169525891542435, + 0.006879997905343771, + 0.00376883358694613, + 0.01366320252418518, + 0.011378795839846134, + 0.004056564066559076, + 0.020908595994114876, + 0.007665205281227827, + 0.019308555871248245, + 0.04546753317117691, + 0.011359515599906445, + 0.029049130156636238, + 0.061261136084795, + 0.045383673161268234, + 0.03014559857547283, + 0.023007957264780998, + 0.007194397039711475, + 0.012142146937549114, + 0.00777817377820611, + 0.01799103617668152, + 0.009353335946798325, + 0.008045670576393604, + 0.018146149814128876, + 0.007110475562512875, + 0.007188448682427406, + 0.02209698222577572, + 0.006778162904083729, + 0.011550738476216793, + 0.010397844016551971, + 0.005960413720458746, + 0.0063547780737280846, + 0.0059020924381911755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.02060040831565857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370029106736183 + ], + "opposite_logprob_tokens": [ + -3.4087002277374268 + ], + "logprob_reward": -0.03370029106736183, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370029106736183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08440951257944107, + 0.03633077070116997, + 0.0379100926220417, + 0.12046246230602264, + 0.24278941750526428, + 0.18605388700962067, + 0.08272344619035721, + 0.14768628776073456, + 0.02478824555873871, + 0.036845870316028595 + ], + "credit_map": [ + 0.00986175425350666, + 0.00466163968667388, + 0.014443639665842056, + 0.012063791044056416, + 0.004197826609015465, + 0.020432062447071075, + 0.00895630568265915, + 0.009792491793632507, + 0.004812188446521759, + 0.0023630319628864527, + 0.0072445012629032135, + 0.004591572564095259, + 0.0015317247016355395, + 0.005726255476474762, + 0.004601515829563141, + 0.005459979642182589, + 0.00602320022881031, + 0.0029831950087100267, + 0.008167491294443607, + 0.003478037426248193, + 0.001793495612218976, + 0.008832999505102634, + 0.002744535682722926, + 0.00388713669963181, + 0.012711365707218647, + 0.010609591379761696, + 0.022185636684298515, + 0.020869899541139603, + 0.008427885361015797, + 0.023171333596110344, + 0.013111945241689682, + 0.009374802932143211, + 0.03646291047334671, + 0.02476458065211773, + 0.06618737429380417, + 0.03341531753540039, + 0.01234174333512783, + 0.03409581631422043, + 0.019376158714294434, + 0.016145510599017143, + 0.01423889584839344, + 0.016992155462503433, + 0.010773692280054092, + 0.017657341435551643, + 0.015434347093105316, + 0.02940797060728073, + 0.04150810465216637, + 0.040041372179985046, + 0.004738466814160347, + 0.006399177946150303, + 0.004093443043529987, + 0.012367773801088333, + 0.01297792512923479, + 0.006389125250279903, + 0.0235440656542778, + 0.012213467620313168, + 0.011352281086146832, + 0.02954135648906231, + 0.006407547276467085, + 0.019173389300704002, + 0.03248142451047897, + 0.02618241123855114, + 0.01421916950494051, + 0.008328712545335293, + 0.0021956278942525387, + 0.003116885432973504, + 0.001850706641562283, + 0.0044288937933743, + 0.0023809114936739206, + 0.0024794330820441246, + 0.005286726634949446, + 0.0030490600038319826, + 0.003096438478678465, + 0.007932567037642002, + 0.003839501179754734, + 0.006719985511153936, + 0.004342211410403252, + 0.0027898920234292746, + 0.005160514265298843, + 0.0029647587798535824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.03370029106736183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017160940915346146 + ], + "opposite_logprob_tokens": [ + -6.376716136932373 + ], + "logprob_reward": -0.0017160940915346146, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017160940915346146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07131605595350266, + 0.034676358103752136, + 0.1097656637430191, + 0.034558121114969254, + 0.09431450814008713, + 0.09023025631904602, + 0.0656047984957695, + 0.07146754860877991, + 0.18012897670269012, + 0.24793773889541626 + ], + "credit_map": [ + 0.008257175795733929, + 0.0032958986703306437, + 0.017466459423303604, + 0.008082310669124126, + 0.004342738073319197, + 0.016481071710586548, + 0.004134179558604956, + 0.009256219491362572, + 0.005667082034051418, + 0.003307431936264038, + 0.00574817368760705, + 0.005765710491687059, + 0.0020676569547504187, + 0.00567408325150609, + 0.0027067968621850014, + 0.003739419160410762, + 0.015296570956707, + 0.006658952683210373, + 0.027566736564040184, + 0.005498225800693035, + 0.0033379425294697285, + 0.03257632255554199, + 0.0036721895448863506, + 0.015158725902438164, + 0.004731573164463043, + 0.004234993830323219, + 0.006604460068047047, + 0.005674629006534815, + 0.002296417485922575, + 0.004969412460923195, + 0.0029061934910714626, + 0.003140439745038748, + 0.013027418404817581, + 0.015422041527926922, + 0.015791313722729683, + 0.012719772756099701, + 0.00731499120593071, + 0.015293166041374207, + 0.007394431158900261, + 0.0073513756506145, + 0.008314785547554493, + 0.019869739189743996, + 0.008954399265348911, + 0.013105089776217937, + 0.012352949939668179, + 0.01126071810722351, + 0.009367743507027626, + 0.007004829589277506, + 0.008135623298585415, + 0.0069854967296123505, + 0.004691137932240963, + 0.013772114180028439, + 0.00897933915257454, + 0.009085536003112793, + 0.009069457650184631, + 0.004886091221123934, + 0.004992898553609848, + 0.008657232858240604, + 0.0060815042816102505, + 0.008499760180711746, + 0.0185867752879858, + 0.009525371715426445, + 0.009281591512262821, + 0.005842407234013081, + 0.012492012232542038, + 0.027345461770892143, + 0.0152791952714324, + 0.041011992841959, + 0.02103777416050434, + 0.019097765907645226, + 0.02964581735432148, + 0.014218964613974094, + 0.01717408373951912, + 0.06075139716267586, + 0.02577827125787735, + 0.046807099133729935, + 0.029852932319045067, + 0.020550042390823364, + 0.02323412336409092, + 0.023789800703525543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.0017160940915346146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up captures a glowing paper lantern, perched precariously on the edge of a dark, weathered wooden ledge. The wood's distinct grain shimmers with a smooth, damp sheen from the humid air, capturing and reflecting the lantern's warm, golden glow. The lantern itself, made of fragile, translucent rice paper laid over a delicate bamboo frame, emanates a soft, flickering light from within. The lantern begins to slide slowly and inexorably, its base gently scraping against the damp surface. Then, the camera begins a steady, deliberate downward tilt, carefully tracking the object as it loses its grip and tumbles down the ledge. Its fragile body descends in a silent, graceful descent, plunging headfirst into a thick, swirling bank of night fog below. The warm point of light flickers in and out as it falls, casting an ethereal halo through the dense fog before being completely swallowed by the profound darkness, creating a poignant and eerie atmosphere.", + "seed": 17140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_0/index_5842_seed_17140_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": -0.03326217862777412, + "vlm_logprob": -0.03326217862777412, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.03326217862777412, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11243549734354019 + ], + "opposite_logprob_tokens": [ + -2.3624355792999268 + ], + "logprob_reward": -0.11243549734354019, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11243549734354019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13733768463134766, + 0.02726077102124691, + 0.02762500010430813, + 0.07478859275579453, + 0.14332574605941772, + 0.13392871618270874, + 0.09077125042676926, + 0.07283449172973633, + 0.24973881244659424, + 0.042388882488012314 + ], + "credit_map": [ + 0.020662503316998482, + 0.008434361778199673, + 0.01674492657184601, + 0.016810160130262375, + 0.013013891875743866, + 0.017554419115185738, + 0.01925453543663025, + 0.024862894788384438, + 0.004219598602503538, + 0.0014514644863083959, + 0.0029257989954203367, + 0.002132271882146597, + 0.0018647800898179412, + 0.004202721640467644, + 0.003232717514038086, + 0.007231421768665314, + 0.004761999007314444, + 0.0019515067106112838, + 0.003415140323340893, + 0.0023013807367533445, + 0.002497101901099086, + 0.003810921683907509, + 0.003413259284570813, + 0.0054736933670938015, + 0.011996209621429443, + 0.008354326710104942, + 0.011291631497442722, + 0.0077881948091089725, + 0.005271426867693663, + 0.008938162587583065, + 0.007000715471804142, + 0.014147937297821045, + 0.021177949383854866, + 0.008273189887404442, + 0.011386191472411156, + 0.009071794338524342, + 0.005766397807747126, + 0.04257725551724434, + 0.009106182493269444, + 0.03596680238842964, + 0.007510271854698658, + 0.003190327435731888, + 0.03051012195646763, + 0.008604200556874275, + 0.004456890281289816, + 0.010839644819498062, + 0.014653002843260765, + 0.05416426435112953, + 0.010250315070152283, + 0.010449876077473164, + 0.01088673435151577, + 0.010076713748276234, + 0.01659749075770378, + 0.01317388191819191, + 0.008463401347398758, + 0.01087284553796053, + 0.006554543040692806, + 0.00899797584861517, + 0.007260006386786699, + 0.008778118528425694, + 0.013499610126018524, + 0.008822728879749775, + 0.012191890738904476, + 0.006729628425091505, + 0.021063726395368576, + 0.04105854406952858, + 0.038281530141830444, + 0.037112440913915634, + 0.02729758247733116, + 0.021396752446889877, + 0.01892649568617344, + 0.04460175335407257, + 0.003233071183785796, + 0.0049269734881818295, + 0.005897822789847851, + 0.0069502717815339565, + 0.004981995094567537, + 0.009467175230383873, + 0.0029213661327958107, + 0.004010213539004326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.11243549734354019 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656949736177921 + ], + "opposite_logprob_tokens": [ + -4.758656978607178 + ], + "logprob_reward": -0.008656949736177921, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656949736177921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16098056733608246, + 0.07931312918663025, + 0.0846615582704544, + 0.07713306695222855, + 0.1427098661661148, + 0.08554233610630035, + 0.0910327136516571, + 0.04805254563689232, + 0.15695473551750183, + 0.0736195296049118 + ], + "credit_map": [ + 0.02206909842789173, + 0.009358125738799572, + 0.020229414105415344, + 0.0135936439037323, + 0.013916191644966602, + 0.021699467673897743, + 0.023307926952838898, + 0.03680669888854027, + 0.009702317416667938, + 0.004871014971286058, + 0.009839539416134357, + 0.005781632848083973, + 0.004993903450667858, + 0.011262939311563969, + 0.011242262087762356, + 0.021619509905576706, + 0.0110962875187397, + 0.004463245160877705, + 0.012520610354840755, + 0.006258591543883085, + 0.009595504961907864, + 0.010883791372179985, + 0.016171807423233986, + 0.01367170736193657, + 0.011999715119600296, + 0.006737023591995239, + 0.009405515156686306, + 0.006963388063013554, + 0.004555440507829189, + 0.010433738119900227, + 0.007830268703401089, + 0.019207967445254326, + 0.014983918517827988, + 0.008151496760547161, + 0.01578015834093094, + 0.014663886278867722, + 0.012478324584662914, + 0.026502199470996857, + 0.023419475182890892, + 0.026730384677648544, + 0.005152228754013777, + 0.0033358908258378506, + 0.017197467386722565, + 0.007028466090559959, + 0.00394737021997571, + 0.007157719694077969, + 0.010633958503603935, + 0.03108922578394413, + 0.008557078428566456, + 0.006361857056617737, + 0.01350043248385191, + 0.007436763029545546, + 0.009266145527362823, + 0.010305560193955898, + 0.010744867846369743, + 0.02486000582575798, + 0.004041914362460375, + 0.005263699218630791, + 0.004591685254126787, + 0.005933376494795084, + 0.0102023771032691, + 0.007933875545859337, + 0.0048674000427126884, + 0.005218213424086571, + 0.013164986856281757, + 0.020568137988448143, + 0.017111485823988914, + 0.02158713899552822, + 0.014065924100577831, + 0.018668022006750107, + 0.012230202555656433, + 0.039558835327625275, + 0.0051143052987754345, + 0.010920063592493534, + 0.009094509296119213, + 0.009296342730522156, + 0.005806721746921539, + 0.020272059366106987, + 0.004859511740505695, + 0.008256010711193085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.008656949736177921 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301932960748672 + ], + "opposite_logprob_tokens": [ + -3.773301839828491 + ], + "logprob_reward": -0.023301932960748672, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301932960748672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07748186588287354, + 0.021749397739768028, + 0.01981992833316326, + 0.050356484949588776, + 0.08842507749795914, + 0.11373620480298996, + 0.09880746155977249, + 0.1084471046924591, + 0.3670930564403534, + 0.054083410650491714 + ], + "credit_map": [ + 0.009257810190320015, + 0.005099939182400703, + 0.011217579245567322, + 0.004121160134673119, + 0.0036048695910722017, + 0.01182239968329668, + 0.008106407709419727, + 0.02425169013440609, + 0.0034506937954574823, + 0.0022593196481466293, + 0.0023945425637066364, + 0.0017403867095708847, + 0.001372664701193571, + 0.002967200241982937, + 0.002742352895438671, + 0.0048222350887954235, + 0.002879047067835927, + 0.0015939990989863873, + 0.0016683782450854778, + 0.0020063461270183325, + 0.0013036488089710474, + 0.002656949684023857, + 0.0027856561355292797, + 0.004925901535898447, + 0.0066595603711903095, + 0.0034582801163196564, + 0.00557280657812953, + 0.00385859957896173, + 0.0031086343806236982, + 0.0071410550735890865, + 0.00591631093993783, + 0.014641236513853073, + 0.011669178493320942, + 0.005739024840295315, + 0.007691095117479563, + 0.006151329260319471, + 0.005339431576430798, + 0.021351411938667297, + 0.011847847141325474, + 0.018635757267475128, + 0.006863763090223074, + 0.004196728579699993, + 0.02369569055736065, + 0.00804098416119814, + 0.004058127291500568, + 0.007938622497022152, + 0.013567637652158737, + 0.045374635607004166, + 0.009991549886763096, + 0.007461007218807936, + 0.013011087663471699, + 0.010041379369795322, + 0.012220448814332485, + 0.009656606242060661, + 0.011737306602299213, + 0.02468806877732277, + 0.007114548701792955, + 0.008204805664718151, + 0.009205141104757786, + 0.014309177175164223, + 0.018741676583886147, + 0.019146209582686424, + 0.012834342196583748, + 0.01889120042324066, + 0.02879825234413147, + 0.04695605859160423, + 0.03587215021252632, + 0.05169686675071716, + 0.03870134428143501, + 0.04374566301703453, + 0.028113944455981255, + 0.09320873022079468, + 0.003911830019205809, + 0.006275859195739031, + 0.008054343983530998, + 0.006644034292548895, + 0.004310362972319126, + 0.01642514392733574, + 0.003356002736836672, + 0.005105834919959307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.023301932960748672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578688010573387 + ], + "opposite_logprob_tokens": [ + -3.895578622817993 + ], + "logprob_reward": -0.020578688010573387, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578688010573387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0453193336725235, + 0.0203529242426157, + 0.028862865641713142, + 0.026205893605947495, + 0.0788554772734642, + 0.2911311984062195, + 0.22863838076591492, + 0.08474724739789963, + 0.12866704165935516, + 0.06721966713666916 + ], + "credit_map": [ + 0.007064938545227051, + 0.003777687903493643, + 0.006277395412325859, + 0.0028740777634084225, + 0.002396230585873127, + 0.006556227337568998, + 0.004447089042514563, + 0.01192568615078926, + 0.0028809919022023678, + 0.002182228956371546, + 0.0024631647393107414, + 0.0018612582935020328, + 0.0016688545001670718, + 0.002515579340979457, + 0.0028049247339367867, + 0.003975920844823122, + 0.0036307142581790686, + 0.002157071838155389, + 0.0037882323376834393, + 0.0023319495376199484, + 0.002143305726349354, + 0.0038801266346126795, + 0.0040167467668652534, + 0.0069147213362157345, + 0.003717069048434496, + 0.0026948454324156046, + 0.0038665386382490396, + 0.0027866086456924677, + 0.0016686648596078157, + 0.0038489378057420254, + 0.0023937129881232977, + 0.005229515489190817, + 0.01334906741976738, + 0.0025885780341923237, + 0.004236748907715082, + 0.003324607154354453, + 0.006863726302981377, + 0.024941664189100266, + 0.006455618422478437, + 0.017095472663640976, + 0.016476945951581, + 0.005409370642155409, + 0.06761500239372253, + 0.017138469964265823, + 0.00766783906146884, + 0.022052854299545288, + 0.03495103120803833, + 0.11981968581676483, + 0.012861036695539951, + 0.009720309637486935, + 0.040595460683107376, + 0.016521653160452843, + 0.013185937888920307, + 0.019696727395057678, + 0.0321933776140213, + 0.08386386930942535, + 0.004730274900794029, + 0.00470577459782362, + 0.009629066102206707, + 0.005869915243238211, + 0.008881278336048126, + 0.01744319498538971, + 0.009305299259722233, + 0.024182438850402832, + 0.018711945042014122, + 0.009839973412454128, + 0.038192614912986755, + 0.01216295175254345, + 0.011138838715851307, + 0.011814936995506287, + 0.010905477218329906, + 0.015900306403636932, + 0.005037257447838783, + 0.010929892770946026, + 0.010587595403194427, + 0.012375731021165848, + 0.004518796224147081, + 0.014272211119532585, + 0.0034057453740388155, + 0.006092440336942673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.020578688010573387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013378250878304243 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013378250878304243, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013378250878304243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04599037393927574, + 0.01304456777870655, + 0.011369630694389343, + 0.03718814253807068, + 0.019485337659716606, + 0.27792349457740784, + 0.1565852165222168, + 0.17225106060504913, + 0.20620505511760712, + 0.05995716527104378 + ], + "credit_map": [ + 0.006280574481934309, + 0.003636791603639722, + 0.00589895760640502, + 0.0030508681666105986, + 0.004628341179341078, + 0.006985925137996674, + 0.00856931321322918, + 0.006939598824828863, + 0.0014847966376692057, + 0.0008457162184640765, + 0.0013215385843068361, + 0.0008017366053536534, + 0.0008125940221361816, + 0.00197460874915123, + 0.0019172425381839275, + 0.003886332968249917, + 0.0015018447302281857, + 0.0006559230387210846, + 0.0015721801901236176, + 0.0009070696542039514, + 0.0011089507024735212, + 0.0016011731931939721, + 0.0021778165828436613, + 0.0018446716712787747, + 0.006533499341458082, + 0.003555531846359372, + 0.004797372967004776, + 0.0033554232213646173, + 0.0022667543962597847, + 0.004754799418151379, + 0.003022425342351198, + 0.008902331814169884, + 0.0019027992384508252, + 0.0012839315459132195, + 0.002412061206996441, + 0.002260000677779317, + 0.0014871963066980243, + 0.003078020876273513, + 0.0033183014020323753, + 0.003743023844435811, + 0.015881666913628578, + 0.005078109446913004, + 0.06474687159061432, + 0.01648278906941414, + 0.006478945724666119, + 0.019839711487293243, + 0.03342561423778534, + 0.11598977446556091, + 0.01416939590126276, + 0.01039118878543377, + 0.022553063929080963, + 0.013268529437482357, + 0.020431235432624817, + 0.014859575778245926, + 0.01822112314403057, + 0.04269108176231384, + 0.013402863405644894, + 0.015073753893375397, + 0.014875728636980057, + 0.02224823087453842, + 0.0308711975812912, + 0.027498047798871994, + 0.021477267146110535, + 0.026803962886333466, + 0.0220539141446352, + 0.018710985779762268, + 0.031051959842443466, + 0.02438788115978241, + 0.015938643366098404, + 0.02658737264573574, + 0.01853361539542675, + 0.04894067719578743, + 0.004206502810120583, + 0.006466401740908623, + 0.00842208880931139, + 0.007603214122354984, + 0.0062997485511004925, + 0.017804518342018127, + 0.0039522419683635235, + 0.0052024382166564465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.0013378250878304243 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up captures a glowing paper lantern, perched precariously on the edge of a dark, weathered wooden ledge. The wood's distinct grain shimmers with a smooth, damp sheen from the humid air, capturing and reflecting the lantern's warm, golden glow. The lantern itself, made of fragile, translucent rice paper laid over a delicate bamboo frame, emanates a soft, flickering light from within. The lantern begins to slide slowly and inexorably, its base gently scraping against the damp surface. Then, the camera begins a steady, deliberate downward tilt, carefully tracking the object as it loses its grip and tumbles down the ledge. Its fragile body descends in a silent, graceful descent, plunging headfirst into a thick, swirling bank of night fog below. The warm point of light flickers in and out as it falls, casting an ethereal halo through the dense fog before being completely swallowed by the profound darkness, creating a poignant and eerie atmosphere.", + "seed": 17141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_0/index_5842_seed_17141_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": -0.016795304103288798, + "vlm_logprob": -0.016795304103288798, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.016795304103288798, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02020176313817501 + ], + "opposite_logprob_tokens": [ + -4.395201683044434 + ], + "logprob_reward": -0.02020176313817501, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02020176313817501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13843929767608643, + 0.07473831623792648, + 0.06703665107488632, + 0.07467810809612274, + 0.2057999223470688, + 0.15832315385341644, + 0.06952843070030212, + 0.049654390662908554, + 0.12046456336975098, + 0.04133718088269234 + ], + "credit_map": [ + 0.019321251660585403, + 0.019457494840025902, + 0.02382199652493, + 0.010529418475925922, + 0.014894148334860802, + 0.02330750599503517, + 0.009524284861981869, + 0.017583193257451057, + 0.007761471439152956, + 0.011039079166948795, + 0.009273279458284378, + 0.0076468754559755325, + 0.008104067295789719, + 0.010865617543458939, + 0.006063332315534353, + 0.013984584249556065, + 0.007236456032842398, + 0.0040544020012021065, + 0.018974512815475464, + 0.0075366622768342495, + 0.0026086498983204365, + 0.016822373494505882, + 0.004584993701428175, + 0.005218599457293749, + 0.006810581311583519, + 0.005442776717245579, + 0.011579655110836029, + 0.009821458719670773, + 0.005285554099828005, + 0.016363371163606644, + 0.009001677855849266, + 0.010373024269938469, + 0.014912508428096771, + 0.014368018135428429, + 0.043673254549503326, + 0.02997109666466713, + 0.012861311435699463, + 0.04796360060572624, + 0.024786151945590973, + 0.017263969406485558, + 0.008841553702950478, + 0.012543991208076477, + 0.025894127786159515, + 0.035401199012994766, + 0.007766327820718288, + 0.023844905197620392, + 0.029932206496596336, + 0.014098829589784145, + 0.0046763233840465546, + 0.005841387435793877, + 0.011837026104331017, + 0.009695355780422688, + 0.005314389243721962, + 0.010492086410522461, + 0.014620513655245304, + 0.007051348686218262, + 0.005067615304142237, + 0.0043979366309940815, + 0.00658447528257966, + 0.009009949862957, + 0.005000782664865255, + 0.007304536644369364, + 0.005303123965859413, + 0.006985967513173819, + 0.01029285229742527, + 0.008687314577400684, + 0.019938074052333832, + 0.015217171050608158, + 0.0068678888492286205, + 0.012799971736967564, + 0.01075593288987875, + 0.03590535745024681, + 0.0035690900404006243, + 0.003071525366976857, + 0.005252348259091377, + 0.00917972531169653, + 0.00332379387691617, + 0.0038356990553438663, + 0.00758050661534071, + 0.005524489097297192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.02020176313817501 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008657895028591156 + ], + "opposite_logprob_tokens": [ + -4.758657932281494 + ], + "logprob_reward": -0.008657895028591156, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008657895028591156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0518546961247921, + 0.2714824080467224, + 0.07133983820676804, + 0.08029088377952576, + 0.14388875663280487, + 0.07247904688119888, + 0.1577586531639099, + 0.06994973868131638, + 0.03632064908742905, + 0.0446353554725647 + ], + "credit_map": [ + 0.007485785987228155, + 0.012165561318397522, + 0.00714506022632122, + 0.004110245034098625, + 0.004728454630821943, + 0.00784123595803976, + 0.002945585874840617, + 0.00543276546522975, + 0.025376372039318085, + 0.0167752206325531, + 0.06609399616718292, + 0.0104612335562706, + 0.009063083678483963, + 0.10003253072500229, + 0.007164516486227512, + 0.036515459418296814, + 0.008050204254686832, + 0.006461767479777336, + 0.015379277057945728, + 0.013575472868978977, + 0.0031940387561917305, + 0.011455651372671127, + 0.006300848443061113, + 0.006922580301761627, + 0.011310238391160965, + 0.005540367215871811, + 0.011463981121778488, + 0.015381875447928905, + 0.006349179428070784, + 0.00952303409576416, + 0.011388297192752361, + 0.009333912283182144, + 0.010581526905298233, + 0.007207181304693222, + 0.035720765590667725, + 0.016468128189444542, + 0.006561240181326866, + 0.042399946600198746, + 0.014874855987727642, + 0.01007512304931879, + 0.005231008864939213, + 0.005969357676804066, + 0.012690586969256401, + 0.013761801645159721, + 0.0033307455014437437, + 0.01090262457728386, + 0.013344564475119114, + 0.007248361594974995, + 0.009346835315227509, + 0.004403396043926477, + 0.03295409306883812, + 0.008493155241012573, + 0.008669784292578697, + 0.05280502513051033, + 0.016685642302036285, + 0.024400724098086357, + 0.005710056982934475, + 0.0031397545244544744, + 0.009519496001303196, + 0.0055230059660971165, + 0.005062137730419636, + 0.01699993573129177, + 0.011362339369952679, + 0.012633010745048523, + 0.0035781566984951496, + 0.0038114096969366074, + 0.004841363988816738, + 0.007151924539357424, + 0.003550829365849495, + 0.003174792742356658, + 0.003630987135693431, + 0.0065811872482299805, + 0.003796529723331332, + 0.0038545995485037565, + 0.00499110110104084, + 0.007747815456241369, + 0.004272179678082466, + 0.003849973203614354, + 0.007977736182510853, + 0.00814542081207037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.008657895028591156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602861419320107 + ], + "opposite_logprob_tokens": [ + -3.8956029415130615 + ], + "logprob_reward": -0.020602861419320107, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602861419320107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04356281831860542, + 0.06597041338682175, + 0.0665423572063446, + 0.029979577288031578, + 0.17720669507980347, + 0.07200713455677032, + 0.32255372405052185, + 0.06529176980257034, + 0.09119497984647751, + 0.0656905472278595 + ], + "credit_map": [ + 0.004847712814807892, + 0.006985095329582691, + 0.009418079629540443, + 0.004780368413776159, + 0.0033029166515916586, + 0.006693645846098661, + 0.0024318082723766565, + 0.005103186704218388, + 0.006794590502977371, + 0.0066177137196063995, + 0.0089727146551013, + 0.0034981751814484596, + 0.0038494127802550793, + 0.02004290372133255, + 0.0033052044454962015, + 0.012889698147773743, + 0.007231257390230894, + 0.004094080068171024, + 0.019652118906378746, + 0.004369236994534731, + 0.0023871473968029022, + 0.020785054191946983, + 0.002953191753476858, + 0.0050702630542218685, + 0.004800952039659023, + 0.0017864672699943185, + 0.004443081095814705, + 0.005572284571826458, + 0.00151452561840415, + 0.00454657943919301, + 0.0037896744906902313, + 0.003526013344526291, + 0.01822945661842823, + 0.017010830342769623, + 0.023309148848056793, + 0.033965930342674255, + 0.01316860318183899, + 0.027775531634688377, + 0.028843576088547707, + 0.014903589151799679, + 0.005810353439301252, + 0.007332845125347376, + 0.011666245758533478, + 0.014213494956493378, + 0.004451734479516745, + 0.009175264276564121, + 0.013573145493865013, + 0.005784050095826387, + 0.019221043214201927, + 0.009265854023396969, + 0.06638442724943161, + 0.015494542196393013, + 0.01798451505601406, + 0.11010398715734482, + 0.026221580803394318, + 0.05787774175405502, + 0.004565256182104349, + 0.0029833728913217783, + 0.00722107058390975, + 0.008142402395606041, + 0.003836988238617778, + 0.012122217565774918, + 0.012812524102628231, + 0.013607938773930073, + 0.006230553612112999, + 0.007661656942218542, + 0.01830560900270939, + 0.02008282206952572, + 0.007319875061511993, + 0.008823854848742485, + 0.008883696049451828, + 0.01388690434396267, + 0.005021851044148207, + 0.005059400107711554, + 0.0061605279333889484, + 0.014354237355291843, + 0.006821074988692999, + 0.00504605146124959, + 0.015424714423716068, + 0.007802688982337713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.020602861419320107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369786962866783 + ], + "opposite_logprob_tokens": [ + -3.4086978435516357 + ], + "logprob_reward": -0.03369786962866783, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369786962866783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052017997950315475, + 0.05612609162926674, + 0.0674101784825325, + 0.03367775306105614, + 0.08319475501775742, + 0.15192818641662598, + 0.08348250389099121, + 0.13025812804698944, + 0.20171648263931274, + 0.14018788933753967 + ], + "credit_map": [ + 0.005740516819059849, + 0.007329608779400587, + 0.0077485209330916405, + 0.006471187807619572, + 0.005763971712440252, + 0.008302699774503708, + 0.005163630470633507, + 0.005497866775840521, + 0.0064328634180128574, + 0.00808265432715416, + 0.009240851737558842, + 0.0037343036383390427, + 0.0046296501532197, + 0.012154662050306797, + 0.004124097991734743, + 0.007727012969553471, + 0.0061706798151135445, + 0.006702290382236242, + 0.015191469341516495, + 0.008403685875236988, + 0.004279634449630976, + 0.014556176960468292, + 0.006158720701932907, + 0.00594753585755825, + 0.005390889476984739, + 0.0026497631333768368, + 0.006630214862525463, + 0.004549278412014246, + 0.0017952427733689547, + 0.005309768952429295, + 0.0030838751699775457, + 0.00426872493699193, + 0.012649944983422756, + 0.007240315433591604, + 0.01203231979161501, + 0.015269340947270393, + 0.004943739622831345, + 0.010458509437739849, + 0.01032787561416626, + 0.010272720828652382, + 0.017317989841103554, + 0.012859323062002659, + 0.02334822714328766, + 0.02374821901321411, + 0.010395116172730923, + 0.019152769818902016, + 0.0294839758425951, + 0.015622581355273724, + 0.007292915601283312, + 0.006290116347372532, + 0.010279251262545586, + 0.00967483315616846, + 0.006437411531805992, + 0.014595462940633297, + 0.017307918518781662, + 0.01160460151731968, + 0.01192530058324337, + 0.007072660140693188, + 0.012918555177748203, + 0.018137868493795395, + 0.01141181867569685, + 0.014141286723315716, + 0.02517055720090866, + 0.02948010340332985, + 0.013657588511705399, + 0.027155403047800064, + 0.01626516506075859, + 0.04443201422691345, + 0.017528004944324493, + 0.022542636841535568, + 0.02281409688293934, + 0.03732159733772278, + 0.010770706459879875, + 0.01052981335669756, + 0.013300028629601002, + 0.02771001309156418, + 0.012365642003715038, + 0.009606829844415188, + 0.03242170810699463, + 0.0234831515699625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.03369786962866783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000816131301689893 + ], + "opposite_logprob_tokens": [ + -7.125816345214844 + ], + "logprob_reward": -0.000816131301689893, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.000816131301689893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05394637957215309, + 0.08605767786502838, + 0.05375296249985695, + 0.021321654319763184, + 0.10806068778038025, + 0.13118687272071838, + 0.1320389062166214, + 0.13548655807971954, + 0.15747442841529846, + 0.12067386507987976 + ], + "credit_map": [ + 0.007068687118589878, + 0.01328953355550766, + 0.0078422324731946, + 0.0036458573304116726, + 0.004613731987774372, + 0.007707752753049135, + 0.004396607168018818, + 0.005381976254284382, + 0.007099501322954893, + 0.0032549130264669657, + 0.015775522217154503, + 0.002437253249809146, + 0.002400681609287858, + 0.03384431079030037, + 0.0024366583675146103, + 0.018808836117386818, + 0.006015221122652292, + 0.0035834398586302996, + 0.014620229601860046, + 0.005712869577109814, + 0.0022522988729178905, + 0.013278237544000149, + 0.0037723970599472523, + 0.004518270026892424, + 0.0022314870730042458, + 0.0017815455794334412, + 0.003037348622456193, + 0.003192503470927477, + 0.0014296401059255004, + 0.004268589895218611, + 0.0019687507301568985, + 0.003411786863580346, + 0.010956553742289543, + 0.01096049603074789, + 0.012590676546096802, + 0.023970792070031166, + 0.00805713515728712, + 0.01154915988445282, + 0.019513696432113647, + 0.010462178848683834, + 0.007633992005139589, + 0.011039450764656067, + 0.021885158494114876, + 0.03397214785218239, + 0.006578028202056885, + 0.018716471269726753, + 0.02253633737564087, + 0.008825277909636497, + 0.009204501286149025, + 0.006034358870238066, + 0.023787198588252068, + 0.00809394009411335, + 0.00901834573596716, + 0.04273925721645355, + 0.010709468275308609, + 0.02245182730257511, + 0.012296805158257484, + 0.005953609943389893, + 0.01936725527048111, + 0.014652760699391365, + 0.011317544616758823, + 0.03025171346962452, + 0.018382610753178596, + 0.023264259099960327, + 0.010861082933843136, + 0.015383918769657612, + 0.032686199992895126, + 0.02085920423269272, + 0.010450114496052265, + 0.01190346572548151, + 0.014854506589472294, + 0.04047594219446182, + 0.008964898996055126, + 0.009278528392314911, + 0.01370317954570055, + 0.02415752224624157, + 0.01347653940320015, + 0.014690709300339222, + 0.01608264446258545, + 0.02031984180212021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.000816131301689893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up captures a glowing paper lantern, perched precariously on the edge of a dark, weathered wooden ledge. The wood's distinct grain shimmers with a smooth, damp sheen from the humid air, capturing and reflecting the lantern's warm, golden glow. The lantern itself, made of fragile, translucent rice paper laid over a delicate bamboo frame, emanates a soft, flickering light from within. The lantern begins to slide slowly and inexorably, its base gently scraping against the damp surface. Then, the camera begins a steady, deliberate downward tilt, carefully tracking the object as it loses its grip and tumbles down the ledge. Its fragile body descends in a silent, graceful descent, plunging headfirst into a thick, swirling bank of night fog below. The warm point of light flickers in and out as it falls, casting an ethereal halo through the dense fog before being completely swallowed by the profound darkness, creating a poignant and eerie atmosphere.", + "seed": 17142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_0/index_5842_seed_17142_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": -0.011533890827558934, + "vlm_logprob": -0.011533890827558934, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.011533890827558934, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01200062409043312 + ], + "opposite_logprob_tokens": [ + -5.012000560760498 + ], + "logprob_reward": -0.01200062409043312, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01200062409043312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2304006814956665, + 0.03030308522284031, + 0.048620373010635376, + 0.1001557856798172, + 0.09538163989782333, + 0.04600272700190544, + 0.0593191534280777, + 0.1233464702963829, + 0.17949724197387695, + 0.08697286993265152 + ], + "credit_map": [ + 0.0189013984054327, + 0.01654115319252014, + 0.06986971944570541, + 0.07295618206262589, + 0.014864597469568253, + 0.013176865875720978, + 0.01157145295292139, + 0.012519295327365398, + 0.0031068723183125257, + 0.002211418468505144, + 0.0071195270866155624, + 0.007096386514604092, + 0.0035686204209923744, + 0.0026022542733699083, + 0.0022049539256840944, + 0.0023930466268211603, + 0.0037468504160642624, + 0.0030273369047790766, + 0.01491687260568142, + 0.017183085903525352, + 0.00256584445014596, + 0.002430798951536417, + 0.002319426741451025, + 0.0024301495868712664, + 0.02269289828836918, + 0.012894047424197197, + 0.019597038626670837, + 0.014812616631388664, + 0.006884315982460976, + 0.00955361220985651, + 0.004661917686462402, + 0.009059324860572815, + 0.012748554348945618, + 0.012526161968708038, + 0.017176540568470955, + 0.01040700264275074, + 0.0069244313053786755, + 0.023403225466609, + 0.00578617537394166, + 0.0064095365814864635, + 0.006377928890287876, + 0.006963549181818962, + 0.006918975152075291, + 0.005131525453180075, + 0.0035566724836826324, + 0.005048550199717283, + 0.004847299307584763, + 0.007158223073929548, + 0.006612330675125122, + 0.004752091597765684, + 0.008958492428064346, + 0.004216895438730717, + 0.005000757053494453, + 0.008884862065315247, + 0.007939583621919155, + 0.012954135425388813, + 0.013174301013350487, + 0.009115440770983696, + 0.02649112418293953, + 0.010427593253552914, + 0.012065643444657326, + 0.018346888944506645, + 0.018178105354309082, + 0.01554736215621233, + 0.01293314527720213, + 0.02044881321489811, + 0.028560681268572807, + 0.023850517347455025, + 0.013116599060595036, + 0.03970133513212204, + 0.020823903381824493, + 0.020062236115336418, + 0.010485024191439152, + 0.018711308017373085, + 0.01566561870276928, + 0.008481262251734734, + 0.0069065820425748825, + 0.012822160497307777, + 0.0066077313385903835, + 0.007293170318007469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.01200062409043312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036292895674705505 + ], + "opposite_logprob_tokens": [ + -5.628629207611084 + ], + "logprob_reward": -0.0036292895674705505, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036292895674705505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1608990728855133, + 0.07925552129745483, + 0.037015631794929504, + 0.12652453780174255, + 0.0459192618727684, + 0.08556877821683884, + 0.08106786757707596, + 0.15110445022583008, + 0.15528981387615204, + 0.07735499739646912 + ], + "credit_map": [ + 0.023192206397652626, + 0.023191453889012337, + 0.020476913079619408, + 0.02710261009633541, + 0.017973368987441063, + 0.01993650197982788, + 0.011789483018219471, + 0.017236553132534027, + 0.010700749233365059, + 0.009726066142320633, + 0.013738266192376614, + 0.018731193616986275, + 0.006393881980329752, + 0.009601199068129063, + 0.004663483705371618, + 0.005700693465769291, + 0.003685552626848221, + 0.002843120601028204, + 0.008643620647490025, + 0.00937908049672842, + 0.004197793547064066, + 0.0034094166476279497, + 0.002338837133720517, + 0.002518212655559182, + 0.025746498256921768, + 0.014970785938203335, + 0.01871444098651409, + 0.018856273964047432, + 0.009190059266984463, + 0.017267385497689247, + 0.005998751614242792, + 0.015780363231897354, + 0.0049606747925281525, + 0.006045174784958363, + 0.006268349010497332, + 0.007107699289917946, + 0.004930733237415552, + 0.004924632143229246, + 0.006200003903359175, + 0.005482000764459372, + 0.01205097883939743, + 0.012867939658463001, + 0.013286128640174866, + 0.012075556442141533, + 0.006769029423594475, + 0.007481464184820652, + 0.007288425229489803, + 0.013749264180660248, + 0.00879619363695383, + 0.00556959118694067, + 0.012466143816709518, + 0.005865959916263819, + 0.005871011409908533, + 0.017448140308260918, + 0.009457754902541637, + 0.015593087300658226, + 0.01574050262570381, + 0.012791167013347149, + 0.05117395147681236, + 0.013299557380378246, + 0.008913428522646427, + 0.01975630782544613, + 0.009976775385439396, + 0.019452771171927452, + 0.012946096248924732, + 0.020477233454585075, + 0.023135945200920105, + 0.018206922337412834, + 0.011336279101669788, + 0.03446979075670242, + 0.01777811534702778, + 0.016939450055360794, + 0.00857052393257618, + 0.01569943316280842, + 0.012160276994109154, + 0.00845526996999979, + 0.006185478065162897, + 0.011719613336026669, + 0.007193868979811668, + 0.007370540872216225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0036292895674705505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819760538637638 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.01819760538637638, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819760538637638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2346969097852707, + 0.03154229745268822, + 0.0485679991543293, + 0.09752511233091354, + 0.05931837484240532, + 0.06515973061323166, + 0.05798804387450218, + 0.17025448381900787, + 0.16078518331050873, + 0.07416193187236786 + ], + "credit_map": [ + 0.018883809447288513, + 0.010374365374445915, + 0.06730871647596359, + 0.07879379391670227, + 0.01275712437927723, + 0.02272011525928974, + 0.010300904512405396, + 0.013558054342865944, + 0.0032863107044249773, + 0.0032582064159214497, + 0.004914127755910158, + 0.006044701207429171, + 0.004216534551233053, + 0.0037355772219598293, + 0.0029864178504794836, + 0.0031004208140075207, + 0.0035790118854492903, + 0.0023593069054186344, + 0.013450455851852894, + 0.01635035313665867, + 0.0036886450834572315, + 0.0037512595299631357, + 0.002749485895037651, + 0.0026394787710160017, + 0.021533450111746788, + 0.014375840313732624, + 0.014840043149888515, + 0.011857139877974987, + 0.009135340340435505, + 0.013592452742159367, + 0.004895190242677927, + 0.007295642979443073, + 0.007445312570780516, + 0.008543391712009907, + 0.005943389609456062, + 0.006003692280501127, + 0.010397757403552532, + 0.009115014225244522, + 0.006537978537380695, + 0.005331831052899361, + 0.008527439087629318, + 0.007728081662207842, + 0.011054093018174171, + 0.009870932437479496, + 0.004783510696142912, + 0.006910896394401789, + 0.006021563895046711, + 0.01026320829987526, + 0.00603404501453042, + 0.004490676801651716, + 0.008776052854955196, + 0.004020432475954294, + 0.0058040558360517025, + 0.01275260653346777, + 0.006205141544342041, + 0.009905030019581318, + 0.013453385792672634, + 0.014416363090276718, + 0.038621172308921814, + 0.017821641638875008, + 0.014260913245379925, + 0.028314664959907532, + 0.02325240895152092, + 0.020113911479711533, + 0.009714680723845959, + 0.02393784001469612, + 0.02165687456727028, + 0.020035065710544586, + 0.015321743674576283, + 0.0328865572810173, + 0.021232349798083305, + 0.016000069677829742, + 0.00818428210914135, + 0.012713794596493244, + 0.009495186619460583, + 0.008443724364042282, + 0.007004339713603258, + 0.01174075249582529, + 0.008401133120059967, + 0.0081787109375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.01819760538637638 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280736058950424 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.023280736058950424, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280736058950424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1606118232011795, + 0.03798018768429756, + 0.03794639930129051, + 0.059909895062446594, + 0.07747556269168854, + 0.11820849031209946, + 0.14956983923912048, + 0.14449457824230194, + 0.158255934715271, + 0.0555473193526268 + ], + "credit_map": [ + 0.011416986584663391, + 0.010666231624782085, + 0.04850395396351814, + 0.05120604857802391, + 0.009952426888048649, + 0.014293606393039227, + 0.006169639993458986, + 0.008402925916016102, + 0.00465403264388442, + 0.004471771419048309, + 0.004805139731615782, + 0.006482260301709175, + 0.004489878658205271, + 0.007445835508406162, + 0.0020850393921136856, + 0.0035462328232824802, + 0.004996160976588726, + 0.002900349907577038, + 0.005475816782563925, + 0.006095934193581343, + 0.005995961371809244, + 0.005209776107221842, + 0.002746212063357234, + 0.004526183474808931, + 0.008816748857498169, + 0.010804370976984501, + 0.009358236566185951, + 0.010345150716602802, + 0.0043847872875630856, + 0.0072048986330628395, + 0.003974837251007557, + 0.005020866636186838, + 0.013865569606423378, + 0.008412613533437252, + 0.01186396460980177, + 0.007891103625297546, + 0.006807758007198572, + 0.012435194104909897, + 0.007816866971552372, + 0.008382490836083889, + 0.016162429004907608, + 0.014546876773238182, + 0.01958150416612625, + 0.01746157743036747, + 0.009339048527181149, + 0.008777424693107605, + 0.011320429854094982, + 0.021019207313656807, + 0.014399728737771511, + 0.012696400284767151, + 0.01838459074497223, + 0.017285121604800224, + 0.014158098958432674, + 0.03753731772303581, + 0.014874378219246864, + 0.020234201103448868, + 0.012535586953163147, + 0.011626395396888256, + 0.033968161791563034, + 0.01439813245087862, + 0.010467529296875, + 0.02439958229660988, + 0.01872536540031433, + 0.018373822793364525, + 0.012758896686136723, + 0.016150787472724915, + 0.02677282504737377, + 0.01974809356033802, + 0.015541262924671173, + 0.03196503594517708, + 0.014740007929503918, + 0.020579034462571144, + 0.006317972205579281, + 0.007603763602674007, + 0.009065886028110981, + 0.006520750466734171, + 0.004876062273979187, + 0.009115984663367271, + 0.006102724466472864, + 0.005944176111370325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.023280736058950424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a glowing paper lantern made of translucent rice paper over a bamboo frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lantern initially perched on the edge of a dark, weathered wooden ledge with visible grain and a damp sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lantern slide slowly down the ledge, scraping against the surface before tumbling headfirst into the fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady, deliberate downward tilt to track the lantern as it falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lantern plunge into a thick, swirling bank of night fog where its light flickers before disappearing into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005611990345641971 + ], + "opposite_logprob_tokens": [ + -7.500561237335205 + ], + "logprob_reward": -0.0005611990345641971, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005611990345641971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08333708345890045, + 0.024461377412080765, + 0.029979025945067406, + 0.10100554674863815, + 0.07201860845088959, + 0.03794180974364281, + 0.06061596795916557, + 0.23569069802761078, + 0.22449277341365814, + 0.1304570883512497 + ], + "credit_map": [ + 0.011371004395186901, + 0.005927632562816143, + 0.017922433093190193, + 0.021933797746896744, + 0.006910823751240969, + 0.007310561835765839, + 0.0056507703848183155, + 0.006310058292001486, + 0.002835652558133006, + 0.001644240808673203, + 0.005372265819460154, + 0.006279444321990013, + 0.0022375956177711487, + 0.0021881035063415766, + 0.0020909460727125406, + 0.001813129405491054, + 0.0028947102837264538, + 0.001767083304002881, + 0.005520765669643879, + 0.006999592762440443, + 0.005096618551760912, + 0.0021624353248625994, + 0.003101280191913247, + 0.0024365412537008524, + 0.0217637587338686, + 0.013114356435835361, + 0.018245531246066093, + 0.013666092418134212, + 0.00834129098802805, + 0.01049210224300623, + 0.005958382040262222, + 0.009424027986824512, + 0.007944129407405853, + 0.00796828605234623, + 0.011227973736822605, + 0.008044883608818054, + 0.007234353106468916, + 0.015609560534358025, + 0.006863642018288374, + 0.007125780452042818, + 0.00631945813074708, + 0.004557246807962656, + 0.005606113467365503, + 0.0052640400826931, + 0.0029437572229653597, + 0.0035388399846851826, + 0.004247789271175861, + 0.005464569199830294, + 0.006669011432677507, + 0.006177836563438177, + 0.008780627511441708, + 0.006227434147149324, + 0.0050862315110862255, + 0.012791816145181656, + 0.005667101126164198, + 0.00921590905636549, + 0.0226895771920681, + 0.02159474976360798, + 0.05269970744848251, + 0.025232259184122086, + 0.017251593992114067, + 0.0449538491666317, + 0.02186354249715805, + 0.02940543182194233, + 0.017287956550717354, + 0.02065233327448368, + 0.05950486660003662, + 0.023245645686984062, + 0.016636859625577927, + 0.045740414410829544, + 0.021045925095677376, + 0.020378785207867622, + 0.014124036766588688, + 0.020210476592183113, + 0.021910089999437332, + 0.01274284441024065, + 0.012355067767202854, + 0.015685930848121643, + 0.020106695592403412, + 0.013321957550942898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0005611990345641971 + } + ], + "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/0000052/group_1/index_tensor([5822, 5822])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_1/index_tensor([5822, 5822])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..695c8ab3534cb75c6298c2c6eaefe97c9e7b5ec1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_1/index_tensor([5822, 5822])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins at an extremely low angle, then tilts steadily upward as the stack of books rises, shrouding the entire image in a dreamy, soft-focus halo. Against a backdrop of profound darkness, countless heavy, ancient books suddenly rise from a smooth, dark, almost invisible ground. Their hardcovers, bound in faded leather in deep red, dark green, and dark brown, their spines adorned with worn, gilded titles, barely visible in the dim light. As they ascend, countless yellowed, fragile pages flutter frantically through the air like a swarm of moths, and the air is filled with golden dust particles illuminated by a warm, inner glow. With a magical precision that defies the laws of physics, the books swiftly and precisely align, rotate, and seamlessly stack onto one another, rapidly forming a spiraling, seemingly unstable yet miraculously balanced tower. The soft light, primarily derived from the books themselves, imbues the scene with a warm hue, creating an atmosphere of solemnity, mystery, and historical weight, as if witnessing the materialization of knowledge and time.", + "seed": 3350, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_1/index_5822_seed_3350_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": -0.09030876215547323, + "vlm_logprob": -0.09030876215547323, + "vlm_yes_no_margin": 3.324999952316284, + "vlm_token_credit": -0.09030876215547323, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2147054225206375 + ], + "opposite_logprob_tokens": [ + -1.714705467224121 + ], + "logprob_reward": -0.2147054225206375, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2147054225206375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2785071134567261, + 0.04781176149845123, + 0.10659975558519363, + 0.02862521819770336, + 0.3712047040462494, + 0.04305930435657501, + 0.01613648422062397, + 0.011982857249677181, + 0.01817224733531475, + 0.07790061086416245 + ], + "credit_map": [ + 0.016777971759438515, + 0.019755670800805092, + 0.09038150310516357, + 0.03929631784558296, + 0.023197460919618607, + 0.0381273478269577, + 0.028255252167582512, + 0.022715549916028976, + 0.0044794827699661255, + 0.006142440252006054, + 0.005503722932189703, + 0.005352895241230726, + 0.00513872317969799, + 0.009568135254085064, + 0.005005225073546171, + 0.006621137261390686, + 0.00922373402863741, + 0.010269086807966232, + 0.01218444388359785, + 0.021718576550483704, + 0.014293528161942959, + 0.010938268154859543, + 0.015940072014927864, + 0.012032035738229752, + 0.003673908533528447, + 0.004807148594409227, + 0.0034841776359826326, + 0.0032232427038252354, + 0.0030397833324968815, + 0.0029718007426708937, + 0.003564698388800025, + 0.003860454773530364, + 0.031418293714523315, + 0.048264216631650925, + 0.03792109712958336, + 0.03466085344552994, + 0.09258053451776505, + 0.03441453352570534, + 0.04550270736217499, + 0.046442463994026184, + 0.005173089448362589, + 0.005756773985922337, + 0.005126371048390865, + 0.007938351482152939, + 0.004931300412863493, + 0.0035704122856259346, + 0.004848239477723837, + 0.005714761093258858, + 0.0029056514613330364, + 0.0016570785082876682, + 0.0012822443386539817, + 0.0014425889821723104, + 0.001817830721847713, + 0.0025056907907128334, + 0.0019122238736599684, + 0.002613173332065344, + 0.0008965736487880349, + 0.0020404739771038294, + 0.0009090857929550111, + 0.0011925935978069901, + 0.0016747915651649237, + 0.0018035628600046039, + 0.0013019883772358298, + 0.0021637852769345045, + 0.0016866230871528387, + 0.0018663310911506414, + 0.0028243190608918667, + 0.0017483133124187589, + 0.0015714928740635514, + 0.002478827489539981, + 0.0022416769061237574, + 0.003754663048312068, + 0.010173245333135128, + 0.00930984877049923, + 0.009992490522563457, + 0.007708370219916105, + 0.008571269921958447, + 0.01060400903224945, + 0.006139094941318035, + 0.015402278862893581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.2147054225206375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210599198937416 + ], + "opposite_logprob_tokens": [ + -4.018210411071777 + ], + "logprob_reward": -0.018210599198937416, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018210599198937416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4499349892139435, + 0.09313178062438965, + 0.10235065221786499, + 0.021440977230668068, + 0.07889410108327866, + 0.04129011183977127, + 0.031827185302972794, + 0.03179846331477165, + 0.043367043137550354, + 0.10596470534801483 + ], + "credit_map": [ + 0.022571338340640068, + 0.019634295254945755, + 0.15818734467029572, + 0.06178567185997963, + 0.02592618204653263, + 0.06303860247135162, + 0.05434097349643707, + 0.04445056989789009, + 0.008981144055724144, + 0.01351016666740179, + 0.011654737405478954, + 0.01595746912062168, + 0.012519486248493195, + 0.012683671899139881, + 0.00908840075135231, + 0.008736710995435715, + 0.008161015808582306, + 0.006734378635883331, + 0.00779386330395937, + 0.013047409243881702, + 0.011313051916658878, + 0.015487202443182468, + 0.015752265229821205, + 0.02406146563589573, + 0.0018020402640104294, + 0.0029133190400898457, + 0.0019613178446888924, + 0.00175044359639287, + 0.002178428927436471, + 0.004253188613802195, + 0.002108287997543812, + 0.004473953042179346, + 0.006836360786110163, + 0.009676223620772362, + 0.008352491073310375, + 0.007624859921634197, + 0.018420690670609474, + 0.007601479068398476, + 0.010177711956202984, + 0.010204283520579338, + 0.00456401938572526, + 0.006620083004236221, + 0.005353234708309174, + 0.00939738005399704, + 0.003994380589574575, + 0.0029130838811397552, + 0.0037579110357910395, + 0.004690018482506275, + 0.003282519755885005, + 0.004144225735217333, + 0.0023971498012542725, + 0.00465416582301259, + 0.004438456147909164, + 0.004074595868587494, + 0.004204649943858385, + 0.004631418734788895, + 0.002669816603884101, + 0.0037577736657112837, + 0.003240054938942194, + 0.003622689051553607, + 0.004111786838620901, + 0.004473614506423473, + 0.004295241553336382, + 0.005627484526485205, + 0.004485512617975473, + 0.0052247303538024426, + 0.006490375380963087, + 0.003793605137616396, + 0.0035391945857554674, + 0.005042252130806446, + 0.00546236103400588, + 0.009329011663794518, + 0.01515384204685688, + 0.017376763746142387, + 0.010367087088525295, + 0.015005519613623619, + 0.009439065121114254, + 0.013313999399542809, + 0.010714278556406498, + 0.014594150707125664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.018210599198937416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087591759860516 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087591759860516, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087591759860516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13030719757080078, + 0.06719593703746796, + 0.061567068099975586, + 0.05833557993173599, + 0.14532507956027985, + 0.03388594463467598, + 0.022648001089692116, + 0.03923662751913071, + 0.05411691218614578, + 0.3873816132545471 + ], + "credit_map": [ + 0.009034343995153904, + 0.007526727393269539, + 0.04195805639028549, + 0.015645798295736313, + 0.010812167078256607, + 0.015367218293249607, + 0.013088135980069637, + 0.016874758526682854, + 0.007369709201157093, + 0.008631286211311817, + 0.0066812532022595406, + 0.007012063171714544, + 0.00920280534774065, + 0.012911852449178696, + 0.0060731954872608185, + 0.009313767775893211, + 0.005703076254576445, + 0.006614197511225939, + 0.0050842310301959515, + 0.009972150437533855, + 0.009598836302757263, + 0.006618008017539978, + 0.007937141694128513, + 0.010039426386356354, + 0.006763819605112076, + 0.010425945743918419, + 0.006388318724930286, + 0.005500850733369589, + 0.005502867978066206, + 0.006812802981585264, + 0.007075975649058819, + 0.009864992462098598, + 0.013802805915474892, + 0.017928218469023705, + 0.015171447768807411, + 0.015434657223522663, + 0.03364206477999687, + 0.013286330737173557, + 0.017646940425038338, + 0.01841261424124241, + 0.0029008579440414906, + 0.004811120219528675, + 0.005345664452761412, + 0.008535509929060936, + 0.0023231953382492065, + 0.0023959141690284014, + 0.004098313394933939, + 0.003475372213870287, + 0.00268373591825366, + 0.0033947466872632504, + 0.00215263944119215, + 0.003447536379098892, + 0.0032857286278158426, + 0.002673947485163808, + 0.0021948739886283875, + 0.0028147913981229067, + 0.004390708636492491, + 0.0038677756674587727, + 0.0044725872576236725, + 0.0039011253975331783, + 0.004819964524358511, + 0.005814573261886835, + 0.003920301329344511, + 0.00804959051311016, + 0.004837759304791689, + 0.008997166529297829, + 0.006704028230160475, + 0.007592060137540102, + 0.004408133681863546, + 0.005533679388463497, + 0.008947676979005337, + 0.007096409797668457, + 0.05334587022662163, + 0.06739579141139984, + 0.042087800800800323, + 0.04968171939253807, + 0.037546005100011826, + 0.04759453609585762, + 0.040588244795799255, + 0.04914165288209915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.011087591759860516 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015930414199829 + ], + "opposite_logprob_tokens": [ + -1.701593041419983 + ], + "logprob_reward": -0.2015930414199829, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015930414199829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1023973748087883, + 0.0684521347284317, + 0.09567272663116455, + 0.06529204547405243, + 0.10621944814920425, + 0.17716321349143982, + 0.04694877937436104, + 0.12449315190315247, + 0.04661847651004791, + 0.16674265265464783 + ], + "credit_map": [ + 0.006266247946768999, + 0.008089100010693073, + 0.024889597669243813, + 0.012338104657828808, + 0.011281034909188747, + 0.02018495462834835, + 0.011736875399947166, + 0.0076114642433822155, + 0.007391553372144699, + 0.010087450966238976, + 0.008446733467280865, + 0.008066551759839058, + 0.00858901347965002, + 0.011672673746943474, + 0.007100463844835758, + 0.007097690366208553, + 0.008173887617886066, + 0.0102794598788023, + 0.00911861564964056, + 0.019811643287539482, + 0.011598795652389526, + 0.010189579799771309, + 0.013337047770619392, + 0.013163691386580467, + 0.00557544594630599, + 0.007588948123157024, + 0.007491046097129583, + 0.0073593417182564735, + 0.01041993498802185, + 0.007269565481692553, + 0.008597662672400475, + 0.010990101844072342, + 0.00948577281087637, + 0.014161289669573307, + 0.011050724424421787, + 0.015613991767168045, + 0.018152395263314247, + 0.00781427975744009, + 0.015588583424687386, + 0.014352405443787575, + 0.014453657902777195, + 0.02348586916923523, + 0.01698959246277809, + 0.030493848025798798, + 0.036947235465049744, + 0.01472882553935051, + 0.018470561131834984, + 0.02159363217651844, + 0.006318859755992889, + 0.0041940961964428425, + 0.0035803462378680706, + 0.004341586027294397, + 0.004769871477037668, + 0.007746696937829256, + 0.0049771531485021114, + 0.011020171456038952, + 0.01219475269317627, + 0.01634817011654377, + 0.013631461188197136, + 0.014363542199134827, + 0.016225561499595642, + 0.016997426748275757, + 0.016442248597741127, + 0.018289992585778236, + 0.004110683221369982, + 0.009058821946382523, + 0.00692423852160573, + 0.006100564729422331, + 0.005073122680187225, + 0.004048028960824013, + 0.004709763918071985, + 0.006593252066522837, + 0.018769143149256706, + 0.027761045843362808, + 0.023319948464632034, + 0.023530011996626854, + 0.015605149790644646, + 0.018646156415343285, + 0.014492815360426903, + 0.024618379771709442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.2015930414199829 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2817261517047882, + 0.06565804034471512, + 0.042070161551237106, + 0.06576844304800034, + 0.17655588686466217, + 0.0704573318362236, + 0.04816874861717224, + 0.05104891210794449, + 0.07962311059236526, + 0.11892317980527878 + ], + "credit_map": [ + 0.014451660215854645, + 0.013883059844374657, + 0.10250794887542725, + 0.03717406839132309, + 0.019609052687883377, + 0.032289572060108185, + 0.027313340455293655, + 0.0344974584877491, + 0.00630593579262495, + 0.005857250653207302, + 0.00969111081212759, + 0.005161353386938572, + 0.006985947024077177, + 0.017508570104837418, + 0.007057793904095888, + 0.007090078666806221, + 0.004128403030335903, + 0.005681128241121769, + 0.0036104407627135515, + 0.006243192590773106, + 0.007992534898221493, + 0.004961010999977589, + 0.00556847034022212, + 0.003884977428242564, + 0.008143652230501175, + 0.008078384213149548, + 0.004791901912540197, + 0.0045349677093327045, + 0.006624474655836821, + 0.011124475859105587, + 0.008351093158125877, + 0.014119495637714863, + 0.021809076890349388, + 0.020205331966280937, + 0.0175440926104784, + 0.03104359097778797, + 0.02158760093152523, + 0.012975642457604408, + 0.02463103085756302, + 0.02675950899720192, + 0.007857455871999264, + 0.010757687501609325, + 0.006200779695063829, + 0.01143545750528574, + 0.007244125008583069, + 0.007770037744194269, + 0.010020732879638672, + 0.009171056561172009, + 0.004796196706593037, + 0.005932480096817017, + 0.005093849264085293, + 0.006435089744627476, + 0.005415250547230244, + 0.0063667865470051765, + 0.00509185716509819, + 0.009037245064973831, + 0.004881022032350302, + 0.00531502952799201, + 0.006757078226655722, + 0.005376228131353855, + 0.006096783559769392, + 0.009494992904365063, + 0.005580389406532049, + 0.007547393441200256, + 0.007540196180343628, + 0.012922997586429119, + 0.007674284745007753, + 0.011293894611299038, + 0.007081991992890835, + 0.009516624733805656, + 0.010670927353203297, + 0.012922201305627823, + 0.01283120084553957, + 0.012821055948734283, + 0.016414159908890724, + 0.012383811175823212, + 0.012776409275829792, + 0.01615297421813011, + 0.010294686071574688, + 0.025248879566788673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.005947155877947807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at an extremely low angle, then tilts steadily upward as the stack of books rises, shrouding the entire image in a dreamy, soft-focus halo. Against a backdrop of profound darkness, countless heavy, ancient books suddenly rise from a smooth, dark, almost invisible ground. Their hardcovers, bound in faded leather in deep red, dark green, and dark brown, their spines adorned with worn, gilded titles, barely visible in the dim light. As they ascend, countless yellowed, fragile pages flutter frantically through the air like a swarm of moths, and the air is filled with golden dust particles illuminated by a warm, inner glow. With a magical precision that defies the laws of physics, the books swiftly and precisely align, rotate, and seamlessly stack onto one another, rapidly forming a spiraling, seemingly unstable yet miraculously balanced tower. The soft light, primarily derived from the books themselves, imbues the scene with a warm hue, creating an atmosphere of solemnity, mystery, and historical weight, as if witnessing the materialization of knowledge and time.", + "seed": 3351, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_1/index_5822_seed_3351_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": -0.12293499726802111, + "vlm_logprob": -0.12293499726802111, + "vlm_yes_no_margin": 3.0249999880790712, + "vlm_token_credit": -0.12293499726802111, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12586478888988495 + ], + "opposite_logprob_tokens": [ + -2.3758647441864014 + ], + "logprob_reward": -0.12586478888988495, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12586478888988495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.269804984331131, + 0.03718779236078262, + 0.09587585926055908, + 0.027691083028912544, + 0.04776672646403313, + 0.09645210951566696, + 0.11135736852884293, + 0.044791776686906815, + 0.031794194132089615, + 0.2372780442237854 + ], + "credit_map": [ + 0.020485471934080124, + 0.023123322054743767, + 0.027174171060323715, + 0.043133873492479324, + 0.029572853818535805, + 0.030123768374323845, + 0.069586843252182, + 0.02660471573472023, + 0.0022914265282452106, + 0.0043632108718156815, + 0.004444217309355736, + 0.003883595811203122, + 0.006596982479095459, + 0.0068199425004422665, + 0.004992972128093243, + 0.00379545078612864, + 0.009336060844361782, + 0.01564011536538601, + 0.009434345178306103, + 0.011755900457501411, + 0.01399846002459526, + 0.013398506678640842, + 0.013024619780480862, + 0.009287871420383453, + 0.0021389699541032314, + 0.0049547771923244, + 0.0024684167001396418, + 0.004900002386420965, + 0.004907742142677307, + 0.002155986148864031, + 0.003627117956057191, + 0.0025380735751241446, + 0.005143952090293169, + 0.005535254720598459, + 0.005543371196836233, + 0.006692364811897278, + 0.005585880950093269, + 0.003496214747428894, + 0.009511005133390427, + 0.006258688401430845, + 0.015093114227056503, + 0.00476944912225008, + 0.03568495064973831, + 0.005508081521838903, + 0.002839786699041724, + 0.003249069908633828, + 0.005975073669105768, + 0.023332593962550163, + 0.01302348356693983, + 0.02623917907476425, + 0.024670589715242386, + 0.007623383775353432, + 0.011968913488090038, + 0.006839545909315348, + 0.013108504936099052, + 0.007883780635893345, + 0.0045518274419009686, + 0.004504930227994919, + 0.006118117365986109, + 0.005596199072897434, + 0.0064929393120110035, + 0.006330777425318956, + 0.005220013204962015, + 0.005976981017738581, + 0.003984927665442228, + 0.0024241330102086067, + 0.004542353563010693, + 0.0042934725061059, + 0.002563170390203595, + 0.005362060386687517, + 0.005071084480732679, + 0.003552996087819338, + 0.02313535287976265, + 0.018523236736655235, + 0.055174168199300766, + 0.01830238103866577, + 0.025759276002645493, + 0.043254267424345016, + 0.01814151555299759, + 0.03498787060379982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.12586478888988495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029827017337083817 + ], + "opposite_logprob_tokens": [ + -3.529827117919922 + ], + "logprob_reward": -0.029827017337083817, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029827017337083817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34609946608543396, + 0.06483446061611176, + 0.03784479200839996, + 0.036065954715013504, + 0.03710685297846794, + 0.0936235561966896, + 0.07256028801202774, + 0.01598174311220646, + 0.15107430517673492, + 0.1448085755109787 + ], + "credit_map": [ + 0.0332038551568985, + 0.029114389792084694, + 0.03889850527048111, + 0.03945700824260712, + 0.047282055020332336, + 0.08470191806554794, + 0.03411416709423065, + 0.039327558130025864, + 0.004453720524907112, + 0.009261910803616047, + 0.005525617860257626, + 0.006533169653266668, + 0.009727385826408863, + 0.01571204885840416, + 0.006062224507331848, + 0.007558383978903294, + 0.00313733983784914, + 0.007155898958444595, + 0.0031935374718159437, + 0.006197765935212374, + 0.005327699705958366, + 0.005024695303291082, + 0.004247802309691906, + 0.003560053650289774, + 0.002878818428143859, + 0.005532983690500259, + 0.0038782665506005287, + 0.004962362814694643, + 0.0050066811963915825, + 0.004288269206881523, + 0.00662266043946147, + 0.002895913552492857, + 0.003415052080526948, + 0.00452421186491847, + 0.00437620235607028, + 0.00654049776494503, + 0.006536987144500017, + 0.0031534109730273485, + 0.004348334856331348, + 0.004212156869471073, + 0.016201432794332504, + 0.007607687264680862, + 0.02823769301176071, + 0.009176983498036861, + 0.00787796825170517, + 0.004390394315123558, + 0.008854283019900322, + 0.011277115903794765, + 0.008932283148169518, + 0.0078247906640172, + 0.022028401494026184, + 0.004076934419572353, + 0.004999436438083649, + 0.005344386212527752, + 0.005993565544486046, + 0.013360491953790188, + 0.0025868264492601156, + 0.0011287748347967863, + 0.0039093391969799995, + 0.0012938803993165493, + 0.001305172685533762, + 0.00268495362251997, + 0.001186820911243558, + 0.0018859750125557184, + 0.01536435168236494, + 0.005647949408739805, + 0.040021978318691254, + 0.009004425257444382, + 0.006005220580846071, + 0.053213320672512054, + 0.007061399985104799, + 0.01475565042346716, + 0.013817179016768932, + 0.015765473246574402, + 0.031103570014238358, + 0.011591223999857903, + 0.019302329048514366, + 0.026027321815490723, + 0.010006863623857498, + 0.017194624990224838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.029827017337083817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023304961621761322 + ], + "opposite_logprob_tokens": [ + -3.7733049392700195 + ], + "logprob_reward": -0.023304961621761322, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304961621761322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5701571702957153, + 0.04618574306368828, + 0.09844762086868286, + 0.04022546485066414, + 0.05421293154358864, + 0.05632973834872246, + 0.03957954794168472, + 0.02961869165301323, + 0.027684304863214493, + 0.03755877539515495 + ], + "credit_map": [ + 0.03506651893258095, + 0.03341782093048096, + 0.044937245547771454, + 0.0882677510380745, + 0.05711131542921066, + 0.10851726680994034, + 0.14820899069309235, + 0.05463021248579025, + 0.003788708709180355, + 0.005551885347813368, + 0.004530781880021095, + 0.004987985361367464, + 0.006345382425934076, + 0.011039690114557743, + 0.003930048085749149, + 0.006011259742081165, + 0.00865103118121624, + 0.01176055334508419, + 0.009521751664578915, + 0.019318681210279465, + 0.011702578514814377, + 0.011390536092221737, + 0.016719525679945946, + 0.009382959455251694, + 0.00357648148201406, + 0.0054039438255131245, + 0.003799642436206341, + 0.004991642199456692, + 0.006782016716897488, + 0.004942405968904495, + 0.008031364530324936, + 0.0026979658287018538, + 0.004412706010043621, + 0.0075180912390351295, + 0.007255378644913435, + 0.01049877516925335, + 0.005882124416530132, + 0.0053033833391964436, + 0.007709568366408348, + 0.005632901564240456, + 0.012795831076800823, + 0.0040763807483017445, + 0.0073137362487614155, + 0.005028253886848688, + 0.004070068709552288, + 0.002729541389271617, + 0.007301976904273033, + 0.013013946823775768, + 0.006389042362570763, + 0.008261109702289104, + 0.013461829163134098, + 0.0021919722203165293, + 0.0022667895536869764, + 0.0013521879445761442, + 0.0033677287865430117, + 0.002288884250447154, + 0.004491414874792099, + 0.002011336851865053, + 0.008184379898011684, + 0.002590413438156247, + 0.0026707560755312443, + 0.0036891773343086243, + 0.002351899864152074, + 0.0036293098237365484, + 0.0031988907139748335, + 0.0016436289297416806, + 0.005938221700489521, + 0.0033657471649348736, + 0.0014451537281274796, + 0.004968494642525911, + 0.0035561725962907076, + 0.00356799247674644, + 0.003985878545790911, + 0.0038449850399047136, + 0.008065602742135525, + 0.002772699575871229, + 0.004008785355836153, + 0.00787337962538004, + 0.0025735516101121902, + 0.004433893132954836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.023304961621761322 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289427399635315 + ], + "opposite_logprob_tokens": [ + -1.0539426803588867 + ], + "logprob_reward": -0.4289427399635315, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4289427399635315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09750700742006302, + 0.02631368301808834, + 0.03766018897294998, + 0.029373520985245705, + 0.023839769884943962, + 0.05231911689043045, + 0.5583375096321106, + 0.006947706453502178, + 0.03276104852557182, + 0.134940505027771 + ], + "credit_map": [ + 0.005901760887354612, + 0.004923887550830841, + 0.0077873533591628075, + 0.018779033794999123, + 0.008061974309384823, + 0.01077860128134489, + 0.03227528557181358, + 0.008999094367027283, + 0.002104412531480193, + 0.0038303453475236893, + 0.0022841235622763634, + 0.0021185774821788073, + 0.004719350952655077, + 0.00536387087777257, + 0.002993329893797636, + 0.0028996714390814304, + 0.003458589082583785, + 0.006624592933803797, + 0.003356459317728877, + 0.0044489349238574505, + 0.005770909134298563, + 0.004833200015127659, + 0.005665167234838009, + 0.0035023323725908995, + 0.0029130338225513697, + 0.004538887646049261, + 0.003918790258467197, + 0.0034836172126233578, + 0.004152648616582155, + 0.0038040317595005035, + 0.0038184260483831167, + 0.0027440832927823067, + 0.002104999730363488, + 0.002203434705734253, + 0.003705279203131795, + 0.0027059996500611305, + 0.0025342183653265238, + 0.0021056337282061577, + 0.0030999411828815937, + 0.0053802598267793655, + 0.007404966279864311, + 0.002842382527887821, + 0.018012093380093575, + 0.004230698570609093, + 0.0017956736264750361, + 0.0023458825889974833, + 0.003661632537841797, + 0.012025784701108932, + 0.085581935942173, + 0.13802911341190338, + 0.185890793800354, + 0.0340987853705883, + 0.035394683480262756, + 0.01585441827774048, + 0.03726959601044655, + 0.026218142360448837, + 0.0007893788279034197, + 0.0006335829966701567, + 0.0009400300914421678, + 0.00099878606852144, + 0.0009897287236526608, + 0.0008165629697032273, + 0.0009718405781313777, + 0.0008077952079474926, + 0.0032541737891733646, + 0.0017461000243201852, + 0.009287969209253788, + 0.0020302385091781616, + 0.001885097473859787, + 0.009300666861236095, + 0.002785863121971488, + 0.0024709354620426893, + 0.013817815110087395, + 0.010280332528054714, + 0.03260275721549988, + 0.00936670321971178, + 0.013709960505366325, + 0.028490271419286728, + 0.008496075868606567, + 0.018176577985286713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.4289427399635315 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735478527843952 + ], + "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": [ + 0.06407307833433151, + 0.04951821267604828, + 0.1054646298289299, + 0.025941140949726105, + 0.013185879215598106, + 0.09692288190126419, + 0.1634662002325058, + 0.015324974432587624, + 0.011111863888800144, + 0.45499110221862793 + ], + "credit_map": [ + 0.005238494370132685, + 0.005934022832661867, + 0.006849273107945919, + 0.007052864413708448, + 0.007573586422950029, + 0.016239339485764503, + 0.0075990851037204266, + 0.00758641492575407, + 0.003617646638303995, + 0.007277204655110836, + 0.004369719419628382, + 0.006146220024675131, + 0.010481166653335094, + 0.007049383595585823, + 0.0059090713039040565, + 0.004667799454182386, + 0.01036191824823618, + 0.016185317188501358, + 0.010158691555261612, + 0.013889246620237827, + 0.014885060489177704, + 0.014372103847563267, + 0.01620464399456978, + 0.009407644160091877, + 0.0023906747810542583, + 0.004189374390989542, + 0.002583171706646681, + 0.003629327518865466, + 0.0035764682106673717, + 0.00300094042904675, + 0.00431223027408123, + 0.0022589529398828745, + 0.0013248944887891412, + 0.0018376203952357173, + 0.0013382358010858297, + 0.0020710050594061613, + 0.0012844556476920843, + 0.0014091880293563008, + 0.0017609293572604656, + 0.0021595500875264406, + 0.013687476515769958, + 0.004085785709321499, + 0.039161160588264465, + 0.00395237747579813, + 0.002717721275985241, + 0.0025333527009934187, + 0.004916076082736254, + 0.025868935510516167, + 0.02368505857884884, + 0.040270913392305374, + 0.05128851532936096, + 0.010683855973184109, + 0.01104265358299017, + 0.006052462849766016, + 0.012546337209641933, + 0.007896398194134235, + 0.001460117637179792, + 0.0011691319523379207, + 0.0020972276106476784, + 0.0026906440034508705, + 0.002082989551126957, + 0.0020197280682623386, + 0.0021835649386048317, + 0.0016215710202232003, + 0.0010574981570243835, + 0.0009980485774576664, + 0.0015389783075079322, + 0.0009393469081260264, + 0.0008693719282746315, + 0.002422996098175645, + 0.0015217340551316738, + 0.0017638903809711337, + 0.044879280030727386, + 0.03269434720277786, + 0.10686063766479492, + 0.033043019473552704, + 0.04631665349006653, + 0.09015070647001266, + 0.03451095521450043, + 0.06653552502393723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.006735478527843952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at an extremely low angle, then tilts steadily upward as the stack of books rises, shrouding the entire image in a dreamy, soft-focus halo. Against a backdrop of profound darkness, countless heavy, ancient books suddenly rise from a smooth, dark, almost invisible ground. Their hardcovers, bound in faded leather in deep red, dark green, and dark brown, their spines adorned with worn, gilded titles, barely visible in the dim light. As they ascend, countless yellowed, fragile pages flutter frantically through the air like a swarm of moths, and the air is filled with golden dust particles illuminated by a warm, inner glow. With a magical precision that defies the laws of physics, the books swiftly and precisely align, rotate, and seamlessly stack onto one another, rapidly forming a spiraling, seemingly unstable yet miraculously balanced tower. The soft light, primarily derived from the books themselves, imbues the scene with a warm hue, creating an atmosphere of solemnity, mystery, and historical weight, as if witnessing the materialization of knowledge and time.", + "seed": 3352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_1/index_5822_seed_3352_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": -0.13686908120289445, + "vlm_logprob": -0.13686908120289445, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.13686908120289445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05548515543341637 + ], + "opposite_logprob_tokens": [ + -3.180485248565674 + ], + "logprob_reward": -0.05548515543341637, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05548515543341637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1550426036119461, + 0.06915730983018875, + 0.09707414358854294, + 0.04038764536380768, + 0.08915171772241592, + 0.07764439284801483, + 0.0438413992524147, + 0.050698332488536835, + 0.3392580449581146, + 0.03774441406130791 + ], + "credit_map": [ + 0.015301969833672047, + 0.02638276293873787, + 0.019596243277192116, + 0.01820291392505169, + 0.02340712584555149, + 0.0233230572193861, + 0.014248763211071491, + 0.01457974873483181, + 0.011007379740476608, + 0.007396496366709471, + 0.008981681428849697, + 0.010190284810960293, + 0.006808091886341572, + 0.009648681618273258, + 0.007375304587185383, + 0.007749396376311779, + 0.01234047207981348, + 0.01817740499973297, + 0.007681333925575018, + 0.013247997500002384, + 0.010840124450623989, + 0.013671145774424076, + 0.01070407498627901, + 0.010411594063043594, + 0.0038861113134771585, + 0.005477814935147762, + 0.0040062968619167805, + 0.005283114034682512, + 0.006183173041790724, + 0.005447037983685732, + 0.005879604257643223, + 0.004224494099617004, + 0.012974659912288189, + 0.010914456099271774, + 0.01272727269679308, + 0.011775603517889977, + 0.009739339351654053, + 0.013684827834367752, + 0.009140221402049065, + 0.008195341564714909, + 0.011691836640238762, + 0.008477563969790936, + 0.00758978771045804, + 0.009402735158801079, + 0.01102093979716301, + 0.010745316743850708, + 0.009906396269798279, + 0.008809818886220455, + 0.00384948099963367, + 0.005488617345690727, + 0.003938066307455301, + 0.006212295964360237, + 0.007683403789997101, + 0.0058376076631248, + 0.005643317010253668, + 0.005188608076423407, + 0.004497524816542864, + 0.0058286371640861034, + 0.004510221537202597, + 0.004552227910608053, + 0.009726248681545258, + 0.007136748172342777, + 0.006726538296788931, + 0.007720185909420252, + 0.03349810838699341, + 0.0482439249753952, + 0.024525562301278114, + 0.054358042776584625, + 0.04261542856693268, + 0.04840106889605522, + 0.03991648927330971, + 0.047699421644210815, + 0.004222990013659, + 0.005126461386680603, + 0.0035195169039070606, + 0.003941819537431002, + 0.006411105394363403, + 0.004675794392824173, + 0.005795860197395086, + 0.004050868563354015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.05548515543341637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982921525835991 + ], + "opposite_logprob_tokens": [ + -3.529829263687134 + ], + "logprob_reward": -0.02982921525835991, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982921525835991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.190247043967247, + 0.0745701938867569, + 0.16672049462795258, + 0.03710244968533516, + 0.1099933311343193, + 0.18897326290607452, + 0.06349142640829086, + 0.058003053069114685, + 0.07366973906755447, + 0.0372290201485157 + ], + "credit_map": [ + 0.018801558762788773, + 0.02697405032813549, + 0.021804848685860634, + 0.02013137750327587, + 0.028470145538449287, + 0.023333752527832985, + 0.025424936786293983, + 0.025306379422545433, + 0.011904106475412846, + 0.006871133577078581, + 0.010467104613780975, + 0.011417853645980358, + 0.00919694546610117, + 0.009955087676644325, + 0.009122952818870544, + 0.005635015666484833, + 0.022537587210536003, + 0.03161884471774101, + 0.012366609647870064, + 0.018551459535956383, + 0.027302516624331474, + 0.022996405139565468, + 0.015366784296929836, + 0.01598028838634491, + 0.003968697506934404, + 0.007163460832089186, + 0.0035802850034087896, + 0.004188593477010727, + 0.004670472349971533, + 0.005556763149797916, + 0.004660359118133783, + 0.003313814988359809, + 0.014777647331357002, + 0.016024218872189522, + 0.017664339393377304, + 0.013823180459439754, + 0.012851301580667496, + 0.013140627183020115, + 0.01034412533044815, + 0.011367893777787685, + 0.025081217288970947, + 0.020808160305023193, + 0.02344277687370777, + 0.021160487085580826, + 0.02506757155060768, + 0.025277018547058105, + 0.030425747856497765, + 0.017710290849208832, + 0.0062593198381364346, + 0.006834669969975948, + 0.00644335150718689, + 0.009921283461153507, + 0.007826444692909718, + 0.00825157668441534, + 0.009672272950410843, + 0.008282510563731194, + 0.005872439593076706, + 0.005160653032362461, + 0.005649389699101448, + 0.007899601943790913, + 0.007652314379811287, + 0.00798269547522068, + 0.009793693199753761, + 0.007992267608642578, + 0.0074063995853066444, + 0.008983489125967026, + 0.008449283428490162, + 0.012564636766910553, + 0.007538733538240194, + 0.010014035739004612, + 0.008988560177385807, + 0.009724593721330166, + 0.003871713299304247, + 0.003544555976986885, + 0.004033759236335754, + 0.004196821246296167, + 0.0036291643045842648, + 0.0045149801298975945, + 0.007944091223180294, + 0.005493937525898218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.02982921525835991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221251666545868 + ], + "opposite_logprob_tokens": [ + -4.018221378326416 + ], + "logprob_reward": -0.018221251666545868, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018221251666545868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21804296970367432, + 0.0642017349600792, + 0.23464125394821167, + 0.08999428898096085, + 0.0954924076795578, + 0.06612559407949448, + 0.03482607379555702, + 0.06582734733819962, + 0.09053469449281693, + 0.04031367972493172 + ], + "credit_map": [ + 0.02133207581937313, + 0.03686331957578659, + 0.025690270587801933, + 0.030115582048892975, + 0.02533884532749653, + 0.034370582550764084, + 0.023089991882443428, + 0.021242303773760796, + 0.008482079021632671, + 0.00987268891185522, + 0.0063165719620883465, + 0.007763061672449112, + 0.006716451141983271, + 0.008698674850165844, + 0.007655084133148193, + 0.008697119541466236, + 0.03166613355278969, + 0.0407448373734951, + 0.02065308205783367, + 0.02518194355070591, + 0.03341594710946083, + 0.03144815191626549, + 0.02271413430571556, + 0.028817029669880867, + 0.008485584519803524, + 0.024480978026986122, + 0.008230145089328289, + 0.010007515549659729, + 0.010730181820690632, + 0.012511777691543102, + 0.009532746858894825, + 0.006015360821038485, + 0.012511199340224266, + 0.01199857983738184, + 0.014612145721912384, + 0.011065470054745674, + 0.011694193817675114, + 0.012731597758829594, + 0.010055125690996647, + 0.010824094526469707, + 0.008205988444387913, + 0.008099377155303955, + 0.010630621574819088, + 0.009854123927652836, + 0.007032036315649748, + 0.00834681000560522, + 0.006716309115290642, + 0.007240326143801212, + 0.003967773634940386, + 0.0032798126339912415, + 0.003484482876956463, + 0.00550101837143302, + 0.005736094433814287, + 0.005133495200425386, + 0.003283510450273752, + 0.004439887590706348, + 0.006856645457446575, + 0.004962102510035038, + 0.005847039166837931, + 0.007204863242805004, + 0.011152648366987705, + 0.009994231164455414, + 0.011179214343428612, + 0.008630598895251751, + 0.008847477845847607, + 0.012438264675438404, + 0.008631328120827675, + 0.010797658935189247, + 0.0141160162165761, + 0.011207620613276958, + 0.014377079904079437, + 0.010119243524968624, + 0.004109309520572424, + 0.00431052315980196, + 0.004326385445892811, + 0.004623071756213903, + 0.005045153200626373, + 0.004775996785610914, + 0.008942946791648865, + 0.004180291201919317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.018221251666545868 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576170802116394 + ], + "opposite_logprob_tokens": [ + -0.826170802116394 + ], + "logprob_reward": -0.576170802116394, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576170802116394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16792266070842743, + 0.0934494212269783, + 0.08112331479787827, + 0.036416035145521164, + 0.08405111730098724, + 0.14083391427993774, + 0.117320217192173, + 0.08121668547391891, + 0.13839221000671387, + 0.059274423867464066 + ], + "credit_map": [ + 0.011750005185604095, + 0.02707129344344139, + 0.023273762315511703, + 0.019420653581619263, + 0.023352032527327538, + 0.022205635905265808, + 0.02144165337085724, + 0.019407620653510094, + 0.007731648627668619, + 0.01036841981112957, + 0.010442555882036686, + 0.018675925210118294, + 0.010134323500096798, + 0.012918456457555294, + 0.01094014011323452, + 0.01223795861005783, + 0.009908335283398628, + 0.011851852759718895, + 0.006653413642197847, + 0.011550497263669968, + 0.011506013572216034, + 0.01233027782291174, + 0.008904097601771355, + 0.008418821729719639, + 0.0037153884768486023, + 0.005725538358092308, + 0.003600766183808446, + 0.0038154444191604853, + 0.006173108238726854, + 0.005265907384455204, + 0.004313783720135689, + 0.003806097898632288, + 0.012280546128749847, + 0.007273135241121054, + 0.014215591363608837, + 0.013235704973340034, + 0.007452298421412706, + 0.013572248630225658, + 0.0089340228587389, + 0.00708756735548377, + 0.016910240054130554, + 0.014713563024997711, + 0.018672268837690353, + 0.01697961799800396, + 0.021447068080306053, + 0.018943553790450096, + 0.01683887466788292, + 0.016328731551766396, + 0.01383361779153347, + 0.01693193055689335, + 0.011730905622243881, + 0.0141478655859828, + 0.01677234284579754, + 0.015166214667260647, + 0.015662042424082756, + 0.013075302354991436, + 0.006651843897998333, + 0.007279630284756422, + 0.0073091210797429085, + 0.010344967246055603, + 0.012908550910651684, + 0.011112120002508163, + 0.014171327464282513, + 0.011439123190939426, + 0.01641026698052883, + 0.018575310707092285, + 0.011814369820058346, + 0.015990177169442177, + 0.018247075378894806, + 0.02162834443151951, + 0.016043059527873993, + 0.019683603197336197, + 0.007427639793604612, + 0.007978620938956738, + 0.005157234147191048, + 0.007104094605892897, + 0.010879142209887505, + 0.007729404140263796, + 0.007702900096774101, + 0.005295387003570795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.576170802116394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463898153975606 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463898153975606, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463898153975606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18802762031555176, + 0.05542639270424843, + 0.08727455884218216, + 0.050167009234428406, + 0.17096325755119324, + 0.17506101727485657, + 0.09702666848897934, + 0.05119512230157852, + 0.07658062130212784, + 0.04827771335840225 + ], + "credit_map": [ + 0.01630162075161934, + 0.037905726581811905, + 0.018144063651561737, + 0.02347090281546116, + 0.029274508357048035, + 0.025141727179288864, + 0.021053707227110863, + 0.01673535630106926, + 0.009965689852833748, + 0.0053331791423261166, + 0.006532319355756044, + 0.006661024875938892, + 0.007974989712238312, + 0.007709315046668053, + 0.005918995477259159, + 0.005330874118953943, + 0.008943754248321056, + 0.011249514296650887, + 0.01013647299259901, + 0.006745623890310526, + 0.012920149601995945, + 0.010527770034968853, + 0.012289239093661308, + 0.014462029561400414, + 0.004829170648008585, + 0.009554135613143444, + 0.004954863805323839, + 0.005314542446285486, + 0.007410325575619936, + 0.007964501157402992, + 0.0066131772473454475, + 0.0035262913443148136, + 0.02250121906399727, + 0.022677721455693245, + 0.02664751000702381, + 0.022626152262091637, + 0.02099071443080902, + 0.022199643775820732, + 0.016158170998096466, + 0.017162110656499863, + 0.020474005490541458, + 0.01983128860592842, + 0.0260397307574749, + 0.028014173731207848, + 0.02065914496779442, + 0.021924754604697227, + 0.02255244180560112, + 0.015565465204417706, + 0.013124608434736729, + 0.011898848228156567, + 0.010032352060079575, + 0.010050580836832523, + 0.015899930149316788, + 0.01293657161295414, + 0.009688136167824268, + 0.01339563075453043, + 0.00479021156206727, + 0.006653237622231245, + 0.004619133658707142, + 0.007518355268985033, + 0.007796923164278269, + 0.006581646855920553, + 0.007570292335003614, + 0.005665317643433809, + 0.009228471666574478, + 0.009912779554724693, + 0.006662382278591394, + 0.00953899696469307, + 0.009550699964165688, + 0.01278859842568636, + 0.009276101365685463, + 0.009622586891055107, + 0.005140071269124746, + 0.007111850660294294, + 0.004492105916142464, + 0.006576984189450741, + 0.00733365910127759, + 0.006115518044680357, + 0.0062960186041891575, + 0.005211498122662306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.00463898153975606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at an extremely low angle, then tilts steadily upward as the stack of books rises, shrouding the entire image in a dreamy, soft-focus halo. Against a backdrop of profound darkness, countless heavy, ancient books suddenly rise from a smooth, dark, almost invisible ground. Their hardcovers, bound in faded leather in deep red, dark green, and dark brown, their spines adorned with worn, gilded titles, barely visible in the dim light. As they ascend, countless yellowed, fragile pages flutter frantically through the air like a swarm of moths, and the air is filled with golden dust particles illuminated by a warm, inner glow. With a magical precision that defies the laws of physics, the books swiftly and precisely align, rotate, and seamlessly stack onto one another, rapidly forming a spiraling, seemingly unstable yet miraculously balanced tower. The soft light, primarily derived from the books themselves, imbues the scene with a warm hue, creating an atmosphere of solemnity, mystery, and historical weight, as if witnessing the materialization of knowledge and time.", + "seed": 3353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_1/index_5822_seed_3353_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": -0.1402500657364726, + "vlm_logprob": -0.1402500657364726, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.1402500657364726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07293153554201126 + ], + "opposite_logprob_tokens": [ + -2.9479315280914307 + ], + "logprob_reward": -0.07293153554201126, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07293153554201126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2066199630498886, + 0.06014092639088631, + 0.06229971721768379, + 0.03167909011244774, + 0.044501323252916336, + 0.07063122093677521, + 0.3252592384815216, + 0.025973651558160782, + 0.03872283548116684, + 0.13417203724384308 + ], + "credit_map": [ + 0.013596727512776852, + 0.01166562084108591, + 0.07370281964540482, + 0.03255994990468025, + 0.007944340817630291, + 0.00971047393977642, + 0.012414826080203056, + 0.04502519965171814, + 0.006512861233204603, + 0.005342844408005476, + 0.012473464012145996, + 0.006131632253527641, + 0.005160643253475428, + 0.006067326292395592, + 0.01094510406255722, + 0.007507045287638903, + 0.006297554355114698, + 0.006125237792730331, + 0.009360750205814838, + 0.009047753177583218, + 0.007411505561321974, + 0.00847691297531128, + 0.009674938395619392, + 0.005905068013817072, + 0.0027049770578742027, + 0.003455125028267503, + 0.0051986416801810265, + 0.005689908284693956, + 0.003538411809131503, + 0.00453618448227644, + 0.0038630464114248753, + 0.0026927932631224394, + 0.0017569831106811762, + 0.0022470816038548946, + 0.01572466641664505, + 0.005911738146096468, + 0.0036602020263671875, + 0.0028106493409723043, + 0.010672187432646751, + 0.001717813080176711, + 0.004570287652313709, + 0.011085580103099346, + 0.007356648333370686, + 0.009122248739004135, + 0.011704460717737675, + 0.00951158907264471, + 0.012105814181268215, + 0.005174596793949604, + 0.02079092338681221, + 0.025020940229296684, + 0.013885502703487873, + 0.10390214622020721, + 0.09204939752817154, + 0.02166046015918255, + 0.034072455018758774, + 0.01387740671634674, + 0.0022881473414599895, + 0.003799986094236374, + 0.0013070148415863514, + 0.004046829883009195, + 0.0034165631514042616, + 0.0031427692156285048, + 0.005197584163397551, + 0.002774756168946624, + 0.0037635827902704477, + 0.003196282545104623, + 0.002092829206958413, + 0.005502298008650541, + 0.00812572706490755, + 0.004333795048296452, + 0.007749157026410103, + 0.00395916448906064, + 0.01427137479186058, + 0.012386114336550236, + 0.019222071394324303, + 0.022078577429056168, + 0.019817832857370377, + 0.011631078086793423, + 0.015597466379404068, + 0.019167527556419373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.07293153554201126 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029827943071722984 + ], + "opposite_logprob_tokens": [ + -3.529827833175659 + ], + "logprob_reward": -0.029827943071722984, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029827943071722984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35829469561576843, + 0.07622203230857849, + 0.14930014312267303, + 0.08658206462860107, + 0.048439085483551025, + 0.07363798469305038, + 0.07592214643955231, + 0.06004568189382553, + 0.029419677332043648, + 0.042136456817388535 + ], + "credit_map": [ + 0.026567621156573296, + 0.028560524806380272, + 0.11872319132089615, + 0.051958274096250534, + 0.026688558980822563, + 0.018955443054437637, + 0.02140767313539982, + 0.06543341279029846, + 0.01141539216041565, + 0.010358094237744808, + 0.010255514644086361, + 0.010120699182152748, + 0.006957930978387594, + 0.006057580932974815, + 0.014665881171822548, + 0.006390935275703669, + 0.013740223832428455, + 0.013666300103068352, + 0.022464601323008537, + 0.021268637850880623, + 0.014739951118826866, + 0.01885249838232994, + 0.03009561263024807, + 0.01447231974452734, + 0.0077798981219530106, + 0.012743410654366016, + 0.012583941221237183, + 0.01389008853584528, + 0.009775948710739613, + 0.012179011479020119, + 0.010821931064128876, + 0.006807828322052956, + 0.0021293850149959326, + 0.0026030265726149082, + 0.0156153105199337, + 0.0073514762334525585, + 0.004856140818446875, + 0.003336241003125906, + 0.00966219138354063, + 0.002885315101593733, + 0.004996903706341982, + 0.011397181078791618, + 0.008494124747812748, + 0.009205291047692299, + 0.011504486203193665, + 0.008188887499272823, + 0.01145013514906168, + 0.008400974795222282, + 0.0037437474820762873, + 0.007546055596321821, + 0.0033931932412087917, + 0.022821592167019844, + 0.019004639238119125, + 0.00524761900305748, + 0.009308824315667152, + 0.004856471437960863, + 0.0049493876285851, + 0.009520203806459904, + 0.0027361628599464893, + 0.011332494206726551, + 0.006788278464227915, + 0.005403502378612757, + 0.011398667469620705, + 0.007916986010968685, + 0.002167200669646263, + 0.0029319794848561287, + 0.0020122642163187265, + 0.0030382697004824877, + 0.007370199076831341, + 0.002550040604546666, + 0.005654751788824797, + 0.0036949720233678818, + 0.0034007683862000704, + 0.005563839338719845, + 0.00329423975199461, + 0.005196600221097469, + 0.00633730785921216, + 0.006490663625299931, + 0.005223757587373257, + 0.006629277486354113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.029827943071722984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210481852293015 + ], + "opposite_logprob_tokens": [ + -4.018210411071777 + ], + "logprob_reward": -0.018210481852293015, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018210481852293015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09170415252447128, + 0.08253856003284454, + 0.03945811465382576, + 0.12652114033699036, + 0.04908481240272522, + 0.29108723998069763, + 0.11000474542379379, + 0.0989484041929245, + 0.06353354454040527, + 0.04711930453777313 + ], + "credit_map": [ + 0.007627004757523537, + 0.009095963090658188, + 0.02512686885893345, + 0.012628503143787384, + 0.008673450909554958, + 0.00624642102047801, + 0.0063541666604578495, + 0.015951769426465034, + 0.014837674796581268, + 0.008377325721085072, + 0.0220728050917387, + 0.007066361606121063, + 0.0064923809841275215, + 0.004316109232604504, + 0.004736781120300293, + 0.014639119617640972, + 0.0029053669422864914, + 0.003976079635322094, + 0.005798612721264362, + 0.007139374036341906, + 0.003442503744736314, + 0.002525457413867116, + 0.00969487614929676, + 0.003975844942033291, + 0.01261840295046568, + 0.017133915796875954, + 0.021067293360829353, + 0.01939273439347744, + 0.010903689078986645, + 0.021088430657982826, + 0.014989172108471394, + 0.009327489882707596, + 0.0025475008878856897, + 0.004151598550379276, + 0.005545232445001602, + 0.012312706559896469, + 0.005920044612139463, + 0.005060452036559582, + 0.008925222791731358, + 0.0046220519579946995, + 0.018091654404997826, + 0.04797076806426048, + 0.03148632124066353, + 0.04058649018406868, + 0.049047213047742844, + 0.0349544957280159, + 0.05120328068733215, + 0.017746999859809875, + 0.007148879114538431, + 0.009858778677880764, + 0.005471841432154179, + 0.03487822413444519, + 0.03076392412185669, + 0.006796316709369421, + 0.010951057076454163, + 0.004135730676352978, + 0.00855506956577301, + 0.016411900520324707, + 0.004713466390967369, + 0.014530615881085396, + 0.012900209985673428, + 0.00850480142980814, + 0.023236490786075592, + 0.010095852427184582, + 0.004886699374765158, + 0.010711336508393288, + 0.005560535471886396, + 0.008737207390367985, + 0.00822604913264513, + 0.00882117822766304, + 0.00979976449161768, + 0.006790772080421448, + 0.0035127350129187107, + 0.005092943552881479, + 0.00486765056848526, + 0.005209815222769976, + 0.006971100345253944, + 0.006135276984423399, + 0.007692943327128887, + 0.007636838126927614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.018210481852293015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761845111846924 + ], + "opposite_logprob_tokens": [ + -0.8261845111846924 + ], + "logprob_reward": -0.5761845111846924, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761845111846924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17679306864738464, + 0.08799843490123749, + 0.1491672247648239, + 0.08756506443023682, + 0.0981782004237175, + 0.08431760221719742, + 0.05922761559486389, + 0.14485368132591248, + 0.06425763666629791, + 0.047641441226005554 + ], + "credit_map": [ + 0.010861396789550781, + 0.020351607352495193, + 0.02572975866496563, + 0.03316135331988335, + 0.02195647731423378, + 0.021724769845604897, + 0.028776418417692184, + 0.014231276698410511, + 0.008801436051726341, + 0.0119942557066679, + 0.015417962335050106, + 0.01046012807637453, + 0.008728966116905212, + 0.005631492007523775, + 0.016431694850325584, + 0.0105325011536479, + 0.014092099852859974, + 0.01227117981761694, + 0.02586965076625347, + 0.022332634776830673, + 0.013720780611038208, + 0.010763316415250301, + 0.035836584866046906, + 0.014280968345701694, + 0.007201286964118481, + 0.013164912350475788, + 0.015047392807900906, + 0.01306141633540392, + 0.008551463484764099, + 0.011617164127528667, + 0.010920713655650616, + 0.008000717498362064, + 0.005551489535719156, + 0.005545479711145163, + 0.035885583609342575, + 0.021405894309282303, + 0.0058456165716052055, + 0.004951109644025564, + 0.013746114447712898, + 0.005246913060545921, + 0.006377191748470068, + 0.00881179515272379, + 0.007510292809456587, + 0.011388464830815792, + 0.009071022272109985, + 0.011021220125257969, + 0.012198671698570251, + 0.017938947305083275, + 0.004010497126728296, + 0.008195706643164158, + 0.0038804742507636547, + 0.012381724081933498, + 0.011918049305677414, + 0.005205146968364716, + 0.008095941506326199, + 0.005540079902857542, + 0.011772645637392998, + 0.024515381082892418, + 0.005787436384707689, + 0.022841494530439377, + 0.020307229831814766, + 0.012288487516343594, + 0.0325351320207119, + 0.014805867336690426, + 0.003299542935565114, + 0.0054179709404706955, + 0.0032991126645356417, + 0.008436447009444237, + 0.016024142503738403, + 0.005001396872103214, + 0.0160368625074625, + 0.0067421612329781055, + 0.0042998599819839, + 0.004808441735804081, + 0.005361506715416908, + 0.00615326501429081, + 0.0074539948254823685, + 0.006969458889216185, + 0.006582056172192097, + 0.006012858357280493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.5761845111846924 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless heavy, ancient books rise from a smooth, dark, almost invisible ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the book hardcovers bound in faded leather with colors including deep red, dark green, and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do yellowed, fragile pages flutter frantically through the air like a swarm of moths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books align, rotate, and stack to form a spiraling tower with a warm, inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095857031643391 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095857031643391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095857031643391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3331475257873535, + 0.03686601296067238, + 0.054351892322301865, + 0.051566872745752335, + 0.04615548253059387, + 0.024765929207205772, + 0.3004533052444458, + 0.0627375915646553, + 0.06118783727288246, + 0.028767554089426994 + ], + "credit_map": [ + 0.02174667827785015, + 0.022026091814041138, + 0.11504416167736053, + 0.05292138829827309, + 0.018871067091822624, + 0.01919720694422722, + 0.02238408289849758, + 0.06095684692263603, + 0.004517273046076298, + 0.00601816363632679, + 0.00485987076535821, + 0.004525577183812857, + 0.00418739253655076, + 0.0030569161754101515, + 0.00611947663128376, + 0.003581342985853553, + 0.005994452629238367, + 0.005343178752809763, + 0.007970418781042099, + 0.0076815905049443245, + 0.006396500859409571, + 0.0060104490257799625, + 0.010676161386072636, + 0.004279143176972866, + 0.007929382845759392, + 0.006004598923027515, + 0.008496270515024662, + 0.005931411869823933, + 0.006818090099841356, + 0.0071962494403123856, + 0.00472326809540391, + 0.004467606544494629, + 0.0022709330078214407, + 0.002283804351463914, + 0.015585524961352348, + 0.009104916825890541, + 0.004700958728790283, + 0.00295437965542078, + 0.007289086934179068, + 0.001965878764167428, + 0.0019130116561427712, + 0.0034015364944934845, + 0.003330119652673602, + 0.004187712445855141, + 0.002953477669507265, + 0.0025014770217239857, + 0.0043446677736938, + 0.002133927308022976, + 0.018964316695928574, + 0.025133183225989342, + 0.01298111118376255, + 0.09682035446166992, + 0.08588011562824249, + 0.01996528171002865, + 0.02934221923351288, + 0.011366714723408222, + 0.003553166287019849, + 0.0087557602673769, + 0.003868639934808016, + 0.012936649844050407, + 0.007126228883862495, + 0.007887818850576878, + 0.010075846686959267, + 0.008533479645848274, + 0.0031636313069611788, + 0.005590016487985849, + 0.003134943777695298, + 0.01087978295981884, + 0.01370927132666111, + 0.004331699572503567, + 0.014835009351372719, + 0.005543483421206474, + 0.002763755153864622, + 0.0035774088464677334, + 0.0027134341653436422, + 0.0034126746468245983, + 0.00383366527967155, + 0.0045279571786522865, + 0.0036955312825739384, + 0.0042431289330124855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.004095857031643391 + } + ], + "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/0000052/group_10/index_tensor([3560, 3560])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_10/index_tensor([3560, 3560])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5a0a7704aa970807ae0a510a5f66f9f5c9c34c6a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_10/index_tensor([3560, 3560])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up shot from a slightly overhead angle focuses on an exquisite music box, slowly rotating on a dark, polished wooden table. Crafted from polished mahogany, the rectangular box's sides are adorned with delicate floral inlays of mother-of-pearl. Warmly illuminated by a single source, its surface reflects a soft sheen. Atop the music box, a porcelain ballerina doll in a layered tutu pirouettes gracefully and precisely in sync with the base's rotation. Standing on tiptoes, her arms elegantly raised overhead, tiny crystals adorning the hem of her dress catch and scatter light as she moves, creating fleeting flashes. Soft golden light streams in from the side, illuminating the edges of the music box and casting shifting, soft shadows across the tabletop. The background is a blurred, warm-toned study. The scene is shrouded in a tranquil, nostalgic, and slightly magical atmosphere, as if time has frozen in time, leaving only this eternal, graceful dance.", + "seed": 22557, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_10/index_3560_seed_22557_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": -0.01698055845918134, + "vlm_logprob": -0.01698055845918134, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.01698055845918134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.06607384979724884 + ], + "opposite_logprob_tokens": [ + -2.8160738945007324 + ], + "logprob_reward": -0.06607384979724884, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06607384979724884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2526080906391144, + 0.23620587587356567, + 0.129156693816185, + 0.07292083650827408, + 0.06286884099245071, + 0.03458501398563385, + 0.03897535055875778, + 0.10817254334688187, + 0.039610374718904495, + 0.02489635720849037 + ], + "credit_map": [ + 0.0359836108982563, + 0.028985409066081047, + 0.03852562606334686, + 0.032109905034303665, + 0.026263589039444923, + 0.03332586586475372, + 0.026272274553775787, + 0.03114182874560356, + 0.028892865404486656, + 0.00947007816284895, + 0.0701948031783104, + 0.014982875436544418, + 0.013420972973108292, + 0.07496581971645355, + 0.009114534594118595, + 0.01516393106430769, + 0.018358083441853523, + 0.015188250690698624, + 0.016273368149995804, + 0.011617643758654594, + 0.01503987517207861, + 0.026477521285414696, + 0.009320814162492752, + 0.01688113436102867, + 0.008234801702201366, + 0.007119269110262394, + 0.010713646188378334, + 0.009540713392198086, + 0.009435262531042099, + 0.013833411037921906, + 0.006580028682947159, + 0.007463705725967884, + 0.00755477137863636, + 0.007510648109018803, + 0.0075004855170845985, + 0.006224352866411209, + 0.011900371871888638, + 0.010400772094726562, + 0.006291888188570738, + 0.005485552828758955, + 0.0033880765549838543, + 0.006798462942242622, + 0.004094221163541079, + 0.003761155530810356, + 0.005972137674689293, + 0.004605463705956936, + 0.003149579046294093, + 0.0028159157373011112, + 0.005071648862212896, + 0.004008175805211067, + 0.005044770892709494, + 0.0045749712735414505, + 0.004922025837004185, + 0.006939784623682499, + 0.00395283754914999, + 0.0044611371122300625, + 0.011117542162537575, + 0.009579272009432316, + 0.009825913235545158, + 0.014634154736995697, + 0.013607269152998924, + 0.016239311546087265, + 0.02127832919359207, + 0.011890746653079987, + 0.004943544045090675, + 0.004902614280581474, + 0.0041095963679254055, + 0.002928664907813072, + 0.006950222421437502, + 0.007678180001676083, + 0.004040949046611786, + 0.004056605510413647, + 0.0039003973361104727, + 0.002164015779271722, + 0.004004199057817459, + 0.004300129599869251, + 0.002579108579084277, + 0.003158570732921362, + 0.002179047791287303, + 0.0026108864694833755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06607384979724884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.003621806390583515 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.003621806390583515, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003621806390583515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22596846520900726, + 0.21923041343688965, + 0.10868057608604431, + 0.13520506024360657, + 0.07388775050640106, + 0.02401471696794033, + 0.03367146849632263, + 0.12059016525745392, + 0.039570484310388565, + 0.019180946052074432 + ], + "credit_map": [ + 0.02962265908718109, + 0.035734470933675766, + 0.025055531412363052, + 0.02844446338713169, + 0.030039137229323387, + 0.027515215799212456, + 0.021144500002264977, + 0.02841246873140335, + 0.021652908995747566, + 0.014257698319852352, + 0.056618575006723404, + 0.014762558974325657, + 0.019176902249455452, + 0.06100686267018318, + 0.01353440061211586, + 0.018220508471131325, + 0.01497303880751133, + 0.013946881517767906, + 0.013523567467927933, + 0.009108228608965874, + 0.013939786702394485, + 0.020112790167331696, + 0.008850529789924622, + 0.01422574371099472, + 0.014352872967720032, + 0.012952026911079884, + 0.019240228459239006, + 0.019242117181420326, + 0.019804425537586212, + 0.019222861155867577, + 0.013568961061537266, + 0.016821568831801414, + 0.010641582310199738, + 0.008267643861472607, + 0.008477313444018364, + 0.007624901831150055, + 0.012265165336430073, + 0.012801224365830421, + 0.0067390273325145245, + 0.007070889696478844, + 0.002869561780244112, + 0.003432992147281766, + 0.0020918985828757286, + 0.002584305591881275, + 0.005376495886594057, + 0.002909895498305559, + 0.0030722091905772686, + 0.0016773589886724949, + 0.004982112906873226, + 0.004264236893504858, + 0.00411318801343441, + 0.0034012862015515566, + 0.005860550329089165, + 0.004834956489503384, + 0.002872833516448736, + 0.0033423039130866528, + 0.01244510617107153, + 0.008355366066098213, + 0.01077600009739399, + 0.01671978272497654, + 0.018398968502879143, + 0.016510779038071632, + 0.024926602840423584, + 0.012457558885216713, + 0.004511027596890926, + 0.005712393205612898, + 0.00399412214756012, + 0.0037302288692444563, + 0.007270285859704018, + 0.005853753071278334, + 0.004460331052541733, + 0.004038343671709299, + 0.002174173481762409, + 0.002098172903060913, + 0.0022273121867328882, + 0.002926270477473736, + 0.0019619024824351072, + 0.0025200708769261837, + 0.0028409245423972607, + 0.002432117937132716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.003621806390583515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.01254983339458704 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.01254983339458704, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254983339458704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14794649183750153, + 0.12631486356258392, + 0.3470073938369751, + 0.050503525882959366, + 0.03858574107289314, + 0.04612383246421814, + 0.026366643607616425, + 0.17169557511806488, + 0.03335605189204216, + 0.012099933810532093 + ], + "credit_map": [ + 0.019359223544597626, + 0.027111034840345383, + 0.014439714141190052, + 0.015494784340262413, + 0.019528336822986603, + 0.018725091591477394, + 0.013247833587229252, + 0.020040450617671013, + 0.013368107378482819, + 0.0051017203368246555, + 0.03775748237967491, + 0.006588135380297899, + 0.0068595330230891705, + 0.040633171796798706, + 0.0057733673602342606, + 0.010233331471681595, + 0.04775611683726311, + 0.04393863305449486, + 0.04498307406902313, + 0.026442496106028557, + 0.0419270284473896, + 0.07226470112800598, + 0.026095565408468246, + 0.04359973967075348, + 0.00736941397190094, + 0.0037939026951789856, + 0.00804648082703352, + 0.004009872209280729, + 0.006327595561742783, + 0.012054010294377804, + 0.003946040757000446, + 0.0049562049098312855, + 0.0043899123556911945, + 0.005354368127882481, + 0.005480991676449776, + 0.004437361378222704, + 0.0052626715041697025, + 0.005543042439967394, + 0.004631795920431614, + 0.0034855955746024847, + 0.00512964790686965, + 0.008016987703740597, + 0.004319268278777599, + 0.005189718212932348, + 0.008769698441028595, + 0.006533682346343994, + 0.004893917124718428, + 0.003270908957347274, + 0.004236962180584669, + 0.002598016057163477, + 0.003029915038496256, + 0.002473256317898631, + 0.004185030236840248, + 0.004146364517509937, + 0.002593942917883396, + 0.003103152848780155, + 0.01763966493308544, + 0.01890869252383709, + 0.016496172174811363, + 0.019482582807540894, + 0.03374252840876579, + 0.01969919539988041, + 0.029916150495409966, + 0.015810562297701836, + 0.004526001866906881, + 0.003240256803110242, + 0.003939467016607523, + 0.0038280203007161617, + 0.0042528011836111546, + 0.0053793941624462605, + 0.004343504086136818, + 0.003846601350232959, + 0.0019146680133417249, + 0.001365790725685656, + 0.0014360150089487433, + 0.0017327533569186926, + 0.0014646530617028475, + 0.0016990846488624811, + 0.001249131397344172, + 0.0012378365499898791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.01254983339458704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0019375376868993044 + ], + "opposite_logprob_tokens": [ + -6.251937389373779 + ], + "logprob_reward": -0.0019375376868993044, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019375376868993044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16603906452655792, + 0.13417649269104004, + 0.26238349080085754, + 0.05182831734418869, + 0.10454334318637848, + 0.07514682412147522, + 0.033041708171367645, + 0.11697053909301758, + 0.028316140174865723, + 0.027554089203476906 + ], + "credit_map": [ + 0.02040802873671055, + 0.03303423151373863, + 0.014759820885956287, + 0.01474184263497591, + 0.02600574679672718, + 0.020727073773741722, + 0.015523855574429035, + 0.02083846554160118, + 0.015853891149163246, + 0.00618070550262928, + 0.03384194150567055, + 0.007722542621195316, + 0.00890218187123537, + 0.04136531800031662, + 0.006861406844109297, + 0.013448508456349373, + 0.039794959127902985, + 0.03033844567835331, + 0.03477655351161957, + 0.02002779208123684, + 0.02956490032374859, + 0.052435752004384995, + 0.021616309881210327, + 0.03382876142859459, + 0.009478401392698288, + 0.004275111481547356, + 0.0067673553712666035, + 0.004702226724475622, + 0.007548535708338022, + 0.00948505848646164, + 0.004754186607897282, + 0.004817445296794176, + 0.017955273389816284, + 0.011856243945658207, + 0.012772214598953724, + 0.011060528457164764, + 0.017563382163643837, + 0.015539616346359253, + 0.008418588899075985, + 0.009377487935125828, + 0.008582351729273796, + 0.012242668308317661, + 0.008181060664355755, + 0.008473661728203297, + 0.013082553632557392, + 0.008599532768130302, + 0.00869895238429308, + 0.007286038715392351, + 0.004450356122106314, + 0.003415127983316779, + 0.0038209357298910618, + 0.0038352105766534805, + 0.004899122286587954, + 0.004786024801433086, + 0.003537612035870552, + 0.004297318868339062, + 0.01216733455657959, + 0.017522068694233894, + 0.01289297267794609, + 0.010669528506696224, + 0.026118747889995575, + 0.013374509289860725, + 0.014297417365014553, + 0.009927961975336075, + 0.0034824346657842398, + 0.0028234778437763453, + 0.002865321235731244, + 0.003740441519767046, + 0.003109766636043787, + 0.006056881509721279, + 0.0033111071679741144, + 0.0029267093632370234, + 0.0040248092263937, + 0.002552454359829426, + 0.0032852289732545614, + 0.004682464059442282, + 0.003228863002732396, + 0.003647505771368742, + 0.002635110402479768, + 0.0034976541064679623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0019375376868993044 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0007197650265879929 + ], + "opposite_logprob_tokens": [ + -7.2507195472717285 + ], + "logprob_reward": -0.0007197650265879929, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007197650265879929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12423650175333023, + 0.0782468244433403, + 0.37562644481658936, + 0.047052137553691864, + 0.09805136919021606, + 0.04956487566232681, + 0.07446151971817017, + 0.09121935814619064, + 0.024190323427319527, + 0.03735065087676048 + ], + "credit_map": [ + 0.014363314025104046, + 0.028541652485728264, + 0.014721767976880074, + 0.010320086032152176, + 0.0199518334120512, + 0.014162582345306873, + 0.011409775353968143, + 0.010765494778752327, + 0.008583464659750462, + 0.0034604347310960293, + 0.02360428497195244, + 0.005064334720373154, + 0.004431361798197031, + 0.0234196987003088, + 0.003843233920633793, + 0.00584000488743186, + 0.05626106262207031, + 0.044037844985723495, + 0.05011855810880661, + 0.027314525097608566, + 0.04152563214302063, + 0.07913409173488617, + 0.02975711040198803, + 0.0474775992333889, + 0.006407386623322964, + 0.005290532018989325, + 0.007532824762165546, + 0.005328844301402569, + 0.005615788511931896, + 0.00750654935836792, + 0.005657578352838755, + 0.0037126289680600166, + 0.016390008851885796, + 0.009829477407038212, + 0.00880538672208786, + 0.01218343060463667, + 0.013890477828681469, + 0.01636010967195034, + 0.011050115339457989, + 0.009542359039187431, + 0.006272810511291027, + 0.005163401365280151, + 0.0050402432680130005, + 0.00467224745079875, + 0.0139999371021986, + 0.006165111903101206, + 0.004897198174148798, + 0.0033539272844791412, + 0.011311575770378113, + 0.005506345070898533, + 0.010212434455752373, + 0.009700773283839226, + 0.008113551884889603, + 0.013786885887384415, + 0.00834646075963974, + 0.0074834879487752914, + 0.009969168342649937, + 0.013254073448479176, + 0.010638078674674034, + 0.00827746745198965, + 0.019970538094639778, + 0.010200288146734238, + 0.010890398174524307, + 0.008019338361918926, + 0.0038782430347055197, + 0.002725983504205942, + 0.003102741204202175, + 0.0030969411600381136, + 0.003891553031280637, + 0.002948618959635496, + 0.0018435685196891427, + 0.002702672965824604, + 0.0055497996509075165, + 0.002858017571270466, + 0.005286138039082289, + 0.006105029955506325, + 0.003752621589228511, + 0.005445007234811783, + 0.004227920435369015, + 0.004126113839447498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0007197650265879929 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle focuses on an exquisite music box, slowly rotating on a dark, polished wooden table. Crafted from polished mahogany, the rectangular box's sides are adorned with delicate floral inlays of mother-of-pearl. Warmly illuminated by a single source, its surface reflects a soft sheen. Atop the music box, a porcelain ballerina doll in a layered tutu pirouettes gracefully and precisely in sync with the base's rotation. Standing on tiptoes, her arms elegantly raised overhead, tiny crystals adorning the hem of her dress catch and scatter light as she moves, creating fleeting flashes. Soft golden light streams in from the side, illuminating the edges of the music box and casting shifting, soft shadows across the tabletop. The background is a blurred, warm-toned study. The scene is shrouded in a tranquil, nostalgic, and slightly magical atmosphere, as if time has frozen in time, leaving only this eternal, graceful dance.", + "seed": 22558, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_10/index_3560_seed_22558_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": -0.12199256988242269, + "vlm_logprob": -0.12199256988242269, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.12199256988242269, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.5819672346115112 + ], + "opposite_logprob_tokens": [ + -0.8319672346115112 + ], + "logprob_reward": -0.5819672346115112, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5819672346115112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22836896777153015, + 0.058193523436784744, + 0.07622504234313965, + 0.1682060807943344, + 0.1497993767261505, + 0.060713935643434525, + 0.06692612916231155, + 0.11403783410787582, + 0.05287330225110054, + 0.024655800312757492 + ], + "credit_map": [ + 0.020632101222872734, + 0.02871806174516678, + 0.024905767291784286, + 0.034761957824230194, + 0.03159963712096214, + 0.0317249596118927, + 0.032994288951158524, + 0.023032186552882195, + 0.0047632912173867226, + 0.006407619453966618, + 0.005034483969211578, + 0.008032052777707577, + 0.0066197155974805355, + 0.010261936113238335, + 0.010266146622598171, + 0.0068082790821790695, + 0.006412414833903313, + 0.007164402864873409, + 0.01628190092742443, + 0.008209273219108582, + 0.00542608043178916, + 0.01729097031056881, + 0.007378287147730589, + 0.008061712607741356, + 0.019986774772405624, + 0.015886273235082626, + 0.01752706989645958, + 0.017766283825039864, + 0.018108196556568146, + 0.0326007716357708, + 0.02538292109966278, + 0.020947784185409546, + 0.014544972218573093, + 0.019323201850056648, + 0.01620367169380188, + 0.017655886709690094, + 0.02315565012395382, + 0.025835588574409485, + 0.018725916743278503, + 0.014354481361806393, + 0.005673073697835207, + 0.007066773250699043, + 0.0090513089671731, + 0.007586454041302204, + 0.0071755386888980865, + 0.008892359212040901, + 0.007698616478592157, + 0.007569809909909964, + 0.008536871522665024, + 0.005628081038594246, + 0.007617539260536432, + 0.007016556803137064, + 0.007133119273930788, + 0.011293480172753334, + 0.006466928869485855, + 0.013233548030257225, + 0.016120048239827156, + 0.0126727856695652, + 0.011695590801537037, + 0.008093424141407013, + 0.014310103841125965, + 0.017817405983805656, + 0.011529715731739998, + 0.021798765286803246, + 0.006209016777575016, + 0.0054317619651556015, + 0.005360928364098072, + 0.004663130268454552, + 0.007031201850622892, + 0.006843096110969782, + 0.010819846764206886, + 0.00651431642472744, + 0.0017912532202899456, + 0.0025141662918031216, + 0.0027090145740658045, + 0.0033554458059370518, + 0.0027847690507769585, + 0.003211271483451128, + 0.005042968317866325, + 0.0032469138968735933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.5819672346115112 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0052617397159338 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.0052617397159338, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052617397159338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24926920235157013, + 0.06547722965478897, + 0.08601781725883484, + 0.12683433294296265, + 0.10914367437362671, + 0.07859154790639877, + 0.07077760249376297, + 0.15024133026599884, + 0.04033361002802849, + 0.023313704878091812 + ], + "credit_map": [ + 0.02200574241578579, + 0.030859757214784622, + 0.0280152540653944, + 0.03293406218290329, + 0.036237187683582306, + 0.039987873286008835, + 0.032975003123283386, + 0.026254327967762947, + 0.006886791903525591, + 0.005950265564024448, + 0.005810711998492479, + 0.009208518080413342, + 0.005528598558157682, + 0.011748346500098705, + 0.011374996043741703, + 0.00896899588406086, + 0.004856604617089033, + 0.004739745985716581, + 0.022293297573924065, + 0.007072774693369865, + 0.00405302457511425, + 0.027079716324806213, + 0.007266314700245857, + 0.008656339719891548, + 0.01629655808210373, + 0.014486201107501984, + 0.013821429572999477, + 0.012429974041879177, + 0.01398499682545662, + 0.024183563888072968, + 0.01587652415037155, + 0.01575508899986744, + 0.010718672536313534, + 0.015545775182545185, + 0.012492499314248562, + 0.014548879116773605, + 0.016499685123562813, + 0.015830758959054947, + 0.012532933615148067, + 0.010974470525979996, + 0.006863815244287252, + 0.008572718128561974, + 0.010157561860978603, + 0.009462893940508366, + 0.009695090353488922, + 0.012665238231420517, + 0.011832171119749546, + 0.009342058561742306, + 0.010150983929634094, + 0.006569775752723217, + 0.008423367515206337, + 0.008594687096774578, + 0.005623513367027044, + 0.012081485241651535, + 0.0066057052463293076, + 0.012728085741400719, + 0.023284129798412323, + 0.016548361629247665, + 0.01583651825785637, + 0.011287329718470573, + 0.018614258617162704, + 0.023376356810331345, + 0.014389479532837868, + 0.026904894039034843, + 0.004598591011017561, + 0.003927973564714193, + 0.004501314368098974, + 0.003314921399578452, + 0.00590847572311759, + 0.005570733454078436, + 0.007747022435069084, + 0.004764578305184841, + 0.0019713416695594788, + 0.002529504243284464, + 0.003353496314957738, + 0.0023967318702489138, + 0.0027335360646247864, + 0.0029405474197119474, + 0.004161450080573559, + 0.0032270969823002815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0052617397159338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.018202170729637146 + ], + "opposite_logprob_tokens": [ + -4.018202304840088 + ], + "logprob_reward": -0.018202170729637146, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018202170729637146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16218414902687073, + 0.04136742278933525, + 0.09328114986419678, + 0.06337007135152817, + 0.11084192246198654, + 0.1014082282781601, + 0.10315155982971191, + 0.24787980318069458, + 0.03461030498147011, + 0.04190536588430405 + ], + "credit_map": [ + 0.01755751110613346, + 0.016436628997325897, + 0.021649278700351715, + 0.016264062374830246, + 0.020544936880469322, + 0.02852480113506317, + 0.017800210043787956, + 0.02340671978890896, + 0.004426563624292612, + 0.0035670632496476173, + 0.004017592873424292, + 0.005836175289005041, + 0.004510167054831982, + 0.006039924453943968, + 0.006790813058614731, + 0.00617912458255887, + 0.008724146522581577, + 0.005484162364155054, + 0.030024806037545204, + 0.0052537210285663605, + 0.004998805001378059, + 0.028295500203967094, + 0.004249771125614643, + 0.006250235252082348, + 0.009336954914033413, + 0.0071165780536830425, + 0.008002219721674919, + 0.00660528102889657, + 0.006679299287497997, + 0.009714126586914062, + 0.00825752317905426, + 0.007658093236386776, + 0.012721218168735504, + 0.01229033526033163, + 0.013092569075524807, + 0.015896419063210487, + 0.00926230102777481, + 0.017038175836205482, + 0.014333110302686691, + 0.01620779000222683, + 0.009873801842331886, + 0.01613316684961319, + 0.011476896703243256, + 0.011688637547194958, + 0.015999116003513336, + 0.014262940734624863, + 0.011571375653147697, + 0.010402290150523186, + 0.009658992290496826, + 0.013998244889080524, + 0.009657524526119232, + 0.010790889151394367, + 0.014954030513763428, + 0.015864795073866844, + 0.01623564586043358, + 0.01199143286794424, + 0.03999820724129677, + 0.026583025231957436, + 0.026949195191264153, + 0.020670199766755104, + 0.029524419456720352, + 0.03839189559221268, + 0.02346929907798767, + 0.042293548583984375, + 0.003104874398559332, + 0.004180192481726408, + 0.003592538880184293, + 0.0032273554243147373, + 0.005884225945919752, + 0.005108977667987347, + 0.006239333655685186, + 0.0032728060614317656, + 0.0045000240206718445, + 0.004945940803736448, + 0.007813389413058758, + 0.005085143726319075, + 0.0036951557267457247, + 0.006753265392035246, + 0.0043653021566569805, + 0.0047471411526203156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.018202170729637146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.003610759973526001 + ], + "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": [ + 0.11010242998600006, + 0.11635605990886688, + 0.06580369919538498, + 0.05845828354358673, + 0.11020238697528839, + 0.05757301300764084, + 0.16301822662353516, + 0.2748912274837494, + 0.024680431932210922, + 0.01891420967876911 + ], + "credit_map": [ + 0.012372266501188278, + 0.013220380060374737, + 0.01317343208938837, + 0.013062352314591408, + 0.015379790216684341, + 0.014885945245623589, + 0.015365221537649632, + 0.012643052265048027, + 0.010661711916327477, + 0.01076237577944994, + 0.01178720686584711, + 0.015937458723783493, + 0.009460845030844212, + 0.023109108209609985, + 0.019133131951093674, + 0.015504223294556141, + 0.006063633132725954, + 0.0055173770524561405, + 0.016134504228830338, + 0.007028877269476652, + 0.0061318832449615, + 0.012834245339035988, + 0.00657140277326107, + 0.005521778948605061, + 0.00877749640494585, + 0.005818003322929144, + 0.010599516332149506, + 0.005557647906243801, + 0.005640299990773201, + 0.00791397038847208, + 0.0071982210502028465, + 0.006953136529773474, + 0.01048782654106617, + 0.012819310650229454, + 0.01157346647232771, + 0.012340819463133812, + 0.013195407576858997, + 0.017060933634638786, + 0.011917110532522202, + 0.020807528868317604, + 0.0062051573768258095, + 0.007694265339523554, + 0.007854530587792397, + 0.006823759526014328, + 0.004601660650223494, + 0.009209796786308289, + 0.005029320251196623, + 0.010154525749385357, + 0.02198534645140171, + 0.013405974954366684, + 0.017330696806311607, + 0.01699282042682171, + 0.01666337065398693, + 0.02745785191655159, + 0.01850934699177742, + 0.030672820284962654, + 0.04025137796998024, + 0.031216293573379517, + 0.028941486030817032, + 0.0210406556725502, + 0.03674769029021263, + 0.04167952388525009, + 0.02626221813261509, + 0.048752013593912125, + 0.00317731243558228, + 0.002431010128930211, + 0.0024398609530180693, + 0.002204510848969221, + 0.003355775959789753, + 0.0033051520586013794, + 0.004867611452937126, + 0.0028991999570280313, + 0.001970703713595867, + 0.0018070529913529754, + 0.0024341782554984093, + 0.002602493856102228, + 0.001613991567865014, + 0.002425235928967595, + 0.003630179213359952, + 0.0024303742684423923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.003610759973526001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0009209443815052509 + ], + "opposite_logprob_tokens": [ + -7.00092077255249 + ], + "logprob_reward": -0.0009209443815052509, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009209443815052509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1131352111697197, + 0.05066986754536629, + 0.2229176163673401, + 0.0920317992568016, + 0.06353463232517242, + 0.033315692096948624, + 0.14747585356235504, + 0.22464214265346527, + 0.0267946794629097, + 0.025482507422566414 + ], + "credit_map": [ + 0.012046788819134235, + 0.013891172595322132, + 0.0131619768217206, + 0.013989350758492947, + 0.01470259390771389, + 0.015038627199828625, + 0.016664493829011917, + 0.013640200719237328, + 0.004015892278403044, + 0.005297383759170771, + 0.004743332043290138, + 0.006195657420903444, + 0.004787664860486984, + 0.010892953723669052, + 0.007635782007128, + 0.007101203314960003, + 0.01668342389166355, + 0.008702665567398071, + 0.06825591623783112, + 0.014329502359032631, + 0.008675467222929, + 0.07401619851589203, + 0.01238059252500534, + 0.019873863086104393, + 0.010266582481563091, + 0.008604311384260654, + 0.008565196767449379, + 0.010544934310019016, + 0.009529435075819492, + 0.017599469050765038, + 0.01536649651825428, + 0.011555375531315804, + 0.006194030866026878, + 0.008540179580450058, + 0.007469565141946077, + 0.007556957192718983, + 0.009416108950972557, + 0.01047410536557436, + 0.007220594212412834, + 0.006663091480731964, + 0.003519135992974043, + 0.0030746692791581154, + 0.004782830830663443, + 0.004790659062564373, + 0.003264546860009432, + 0.005820888560265303, + 0.004392586182802916, + 0.0036703748628497124, + 0.02069234661757946, + 0.012268529273569584, + 0.016659988090395927, + 0.016137460246682167, + 0.013182508759200573, + 0.02488287165760994, + 0.016841135919094086, + 0.026811009272933006, + 0.03566449135541916, + 0.023502811789512634, + 0.024705730378627777, + 0.01844094693660736, + 0.026231709867715836, + 0.03424844890832901, + 0.023148981854319572, + 0.03869900852441788, + 0.0033835198264569044, + 0.0028612015303224325, + 0.003202299354597926, + 0.002054440788924694, + 0.0036359119694679976, + 0.00373079557903111, + 0.0045866817235946655, + 0.003339831018820405, + 0.0030664231162518263, + 0.0021643326617777348, + 0.005096582230180502, + 0.0026443656533956528, + 0.0018506132764741778, + 0.0037288989406079054, + 0.0026309886015951633, + 0.00430030282586813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.0009209443815052509 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle focuses on an exquisite music box, slowly rotating on a dark, polished wooden table. Crafted from polished mahogany, the rectangular box's sides are adorned with delicate floral inlays of mother-of-pearl. Warmly illuminated by a single source, its surface reflects a soft sheen. Atop the music box, a porcelain ballerina doll in a layered tutu pirouettes gracefully and precisely in sync with the base's rotation. Standing on tiptoes, her arms elegantly raised overhead, tiny crystals adorning the hem of her dress catch and scatter light as she moves, creating fleeting flashes. Soft golden light streams in from the side, illuminating the edges of the music box and casting shifting, soft shadows across the tabletop. The background is a blurred, warm-toned study. The scene is shrouded in a tranquil, nostalgic, and slightly magical atmosphere, as if time has frozen in time, leaving only this eternal, graceful dance.", + "seed": 22559, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_10/index_3560_seed_22559_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": -0.018774877104442568, + "vlm_logprob": -0.018774877104442568, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.018774877104442568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0625554621219635 + ], + "opposite_logprob_tokens": [ + -2.9375555515289307 + ], + "logprob_reward": -0.0625554621219635, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0625554621219635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33581265807151794, + 0.07361775636672974, + 0.0671340599656105, + 0.04138670489192009, + 0.03761596977710724, + 0.09219113737344742, + 0.1325669139623642, + 0.09665989875793457, + 0.040795184671878815, + 0.082219697535038 + ], + "credit_map": [ + 0.04663117229938507, + 0.04753068834543228, + 0.035117410123348236, + 0.033312879502773285, + 0.07118786126375198, + 0.04892692342400551, + 0.023788433521986008, + 0.029317306354641914, + 0.010489486157894135, + 0.011861837469041348, + 0.010650834068655968, + 0.007164023816585541, + 0.009775028564035892, + 0.012087409384548664, + 0.005352862179279327, + 0.006236271467059851, + 0.012165339663624763, + 0.006203480996191502, + 0.013036886230111122, + 0.0041437167674303055, + 0.006594189442694187, + 0.017864404246211052, + 0.0030450092162936926, + 0.004081035498529673, + 0.006025118287652731, + 0.006018894258886576, + 0.0055996752344071865, + 0.003885477315634489, + 0.006567798554897308, + 0.005607122555375099, + 0.0044509610161185265, + 0.0032316558063030243, + 0.00745746586471796, + 0.0038692911621183157, + 0.00546422740444541, + 0.003712150501087308, + 0.0043728770688176155, + 0.006653548218309879, + 0.0030004584696143866, + 0.003085952950641513, + 0.014465422369539738, + 0.009564992971718311, + 0.010850831866264343, + 0.011830915696918964, + 0.011516593396663666, + 0.017241572961211205, + 0.008787180297076702, + 0.007933630608022213, + 0.018192030489444733, + 0.022532213479280472, + 0.013293118216097355, + 0.01362995058298111, + 0.022951534017920494, + 0.01579553820192814, + 0.016384262591600418, + 0.009788274765014648, + 0.015500977635383606, + 0.00971053633838892, + 0.015590379014611244, + 0.009329908527433872, + 0.015166692435741425, + 0.01653408072888851, + 0.00733414338901639, + 0.00749318627640605, + 0.005868880543857813, + 0.005231888499110937, + 0.0048851799219846725, + 0.005988650489598513, + 0.004777852445840836, + 0.0045813871547579765, + 0.0037343010772019625, + 0.005727046635001898, + 0.010338917374610901, + 0.010826046578586102, + 0.014026055112481117, + 0.005500469356775284, + 0.014012601226568222, + 0.018194716423749924, + 0.00521641643717885, + 0.0041044787503778934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0625554621219635 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.005262451246380806 + ], + "opposite_logprob_tokens": [ + -5.25526237487793 + ], + "logprob_reward": -0.005262451246380806, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262451246380806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25365471839904785, + 0.12137983739376068, + 0.08655223250389099, + 0.11436032503843307, + 0.07534101605415344, + 0.059233080595731735, + 0.08666587620973587, + 0.10376900434494019, + 0.042284827679395676, + 0.05675903335213661 + ], + "credit_map": [ + 0.03736557811498642, + 0.03028048388659954, + 0.030918266624212265, + 0.028181960806250572, + 0.0504395067691803, + 0.03292649984359741, + 0.021664611995220184, + 0.02187783271074295, + 0.01454803068190813, + 0.014534804970026016, + 0.014478602446615696, + 0.012394062243402004, + 0.023382708430290222, + 0.017350580543279648, + 0.013625940307974815, + 0.01106511615216732, + 0.010855884291231632, + 0.012688523158431053, + 0.013227912597358227, + 0.007641290780156851, + 0.013232128694653511, + 0.016893170773983, + 0.0062034414149820805, + 0.0058098952285945415, + 0.012963560409843922, + 0.013123002834618092, + 0.015450193546712399, + 0.009387394413352013, + 0.02234114520251751, + 0.020586632192134857, + 0.01187989767640829, + 0.008628510870039463, + 0.013416944071650505, + 0.00928632915019989, + 0.010103183798491955, + 0.007352720480412245, + 0.013758225366473198, + 0.009612451307475567, + 0.005425945855677128, + 0.006385220680385828, + 0.010853295214474201, + 0.006633560638874769, + 0.0036873205099254847, + 0.005735521204769611, + 0.012380896136164665, + 0.009005778469145298, + 0.006384741514921188, + 0.004551973193883896, + 0.011938745155930519, + 0.012750783003866673, + 0.00964295957237482, + 0.008255419321358204, + 0.0152359027415514, + 0.014246753416955471, + 0.009096489287912846, + 0.005498834885656834, + 0.01764611154794693, + 0.014768952503800392, + 0.015358117409050465, + 0.00918255839496851, + 0.01349590066820383, + 0.016136864200234413, + 0.007703667506575584, + 0.009476841427385807, + 0.007140697445720434, + 0.006680397316813469, + 0.005503370892256498, + 0.005005230661481619, + 0.0052560679614543915, + 0.004594280384480953, + 0.004664904437959194, + 0.0034398825373500586, + 0.006917044520378113, + 0.00771512882784009, + 0.010867505334317684, + 0.004784313030540943, + 0.0054371552541852, + 0.011569985188543797, + 0.004144405014812946, + 0.005323497578501701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.005262451246380806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.023305542767047882 + ], + "opposite_logprob_tokens": [ + -3.773305654525757 + ], + "logprob_reward": -0.023305542767047882, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305542767047882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3046554625034332, + 0.16813091933727264, + 0.0735677182674408, + 0.1429145485162735, + 0.0551692433655262, + 0.06435506045818329, + 0.03609923645853996, + 0.04605167731642723, + 0.044099822640419006, + 0.06495637446641922 + ], + "credit_map": [ + 0.052021775394678116, + 0.029823342338204384, + 0.04028850048780441, + 0.029054559767246246, + 0.05011779069900513, + 0.04821101203560829, + 0.029472287744283676, + 0.02566619962453842, + 0.017852138727903366, + 0.023037685081362724, + 0.019403457641601562, + 0.016849759966135025, + 0.029192592948675156, + 0.024013075977563858, + 0.02022860012948513, + 0.017553599551320076, + 0.012346740812063217, + 0.010266563855111599, + 0.009013266302645206, + 0.0057921563275158405, + 0.01323645655065775, + 0.010828105732798576, + 0.006767872255295515, + 0.0053165508434176445, + 0.01720712147653103, + 0.014706293120980263, + 0.020930903032422066, + 0.011282123625278473, + 0.02104261890053749, + 0.031276099383831024, + 0.014222709462046623, + 0.012246675789356232, + 0.008012112230062485, + 0.006798463873565197, + 0.006719609722495079, + 0.0046652015298604965, + 0.0101995300501585, + 0.009836572222411633, + 0.005140477791428566, + 0.0037972740828990936, + 0.007759159430861473, + 0.007612837478518486, + 0.006531230639666319, + 0.006380212493240833, + 0.013292680494487286, + 0.00994396023452282, + 0.007169574964791536, + 0.0056654042564332485, + 0.006407597567886114, + 0.0046851178631186485, + 0.003750530071556568, + 0.0037570002023130655, + 0.005447940900921822, + 0.005029292311519384, + 0.004373657051473856, + 0.0026481004897505045, + 0.0060524302534759045, + 0.005304770078510046, + 0.007371990475803614, + 0.004914241377264261, + 0.00641216104850173, + 0.006265196017920971, + 0.004880817607045174, + 0.004850069526582956, + 0.005783316679298878, + 0.00675542838871479, + 0.004532067105174065, + 0.00577044952660799, + 0.006478351540863514, + 0.005004874896258116, + 0.005065044388175011, + 0.00471029058098793, + 0.011085618287324905, + 0.009828019887208939, + 0.008121405728161335, + 0.00504306610673666, + 0.0102217523381114, + 0.009614030830562115, + 0.005074516870081425, + 0.005967957898974419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023305542767047882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0019375376868993044 + ], + "opposite_logprob_tokens": [ + -6.251937389373779 + ], + "logprob_reward": -0.0019375376868993044, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019375376868993044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34266430139541626, + 0.11873869597911835, + 0.1283942312002182, + 0.0768515020608902, + 0.04484380781650543, + 0.06433553248643875, + 0.06685963273048401, + 0.04301854595541954, + 0.053730420768260956, + 0.06056337431073189 + ], + "credit_map": [ + 0.06808388978242874, + 0.03508454188704491, + 0.048149462789297104, + 0.025107458233833313, + 0.04460188373923302, + 0.07040699571371078, + 0.02628312073647976, + 0.0249469093978405, + 0.02033485285937786, + 0.012256679125130177, + 0.013840043917298317, + 0.012545903213322163, + 0.014097818173468113, + 0.02370879240334034, + 0.008681714534759521, + 0.013272878713905811, + 0.0277531947940588, + 0.025671545416116714, + 0.01808358170092106, + 0.01098715327680111, + 0.012287615798413754, + 0.017019569873809814, + 0.0066762929782271385, + 0.009915267117321491, + 0.013092161156237125, + 0.011743936687707901, + 0.009710966609418392, + 0.007302071433514357, + 0.010656778700649738, + 0.012240183539688587, + 0.005954933352768421, + 0.006150467786937952, + 0.0075568994507193565, + 0.006223712116479874, + 0.005501134321093559, + 0.005013364367187023, + 0.006781059317290783, + 0.006161858327686787, + 0.0038763354532420635, + 0.0037294444628059864, + 0.010580608621239662, + 0.006281157489866018, + 0.010724427178502083, + 0.009107066318392754, + 0.006901460699737072, + 0.007361845578998327, + 0.00565950945019722, + 0.0077194455079734325, + 0.009330768138170242, + 0.007351408712565899, + 0.00737532926723361, + 0.007939904928207397, + 0.010578935034573078, + 0.00821833685040474, + 0.0093603590503335, + 0.006704588420689106, + 0.006377241108566523, + 0.005528874229639769, + 0.004899272229522467, + 0.006262489128857851, + 0.004973637405782938, + 0.005035297945141792, + 0.004310993477702141, + 0.005630736704915762, + 0.012168703600764275, + 0.007937448099255562, + 0.007863584905862808, + 0.0055589680559933186, + 0.005262718070298433, + 0.006337499711662531, + 0.004564862232655287, + 0.004036630503833294, + 0.011989254504442215, + 0.009494897909462452, + 0.010806072503328323, + 0.0035639945417642593, + 0.00478141475468874, + 0.011595566757023335, + 0.003321356838569045, + 0.005010808818042278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0019375376868993044 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0008133916999213398 + ], + "opposite_logprob_tokens": [ + -7.1258134841918945 + ], + "logprob_reward": -0.0008133916999213398, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008133916999213398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29101109504699707, + 0.10095374286174774, + 0.053725406527519226, + 0.08656387776136398, + 0.04570503532886505, + 0.055834438651800156, + 0.10842174291610718, + 0.09641069173812866, + 0.1058041900396347, + 0.055569760501384735 + ], + "credit_map": [ + 0.04003068804740906, + 0.04105425253510475, + 0.03439304605126381, + 0.022024596109986305, + 0.05015974119305611, + 0.055080827325582504, + 0.026577774435281754, + 0.021690186113119125, + 0.01942593976855278, + 0.01984729804098606, + 0.011519396677613258, + 0.010958763770759106, + 0.011455336585640907, + 0.013694395311176777, + 0.006357990205287933, + 0.0076946294866502285, + 0.010057562030851841, + 0.006651426199823618, + 0.005560626275837421, + 0.00450810557231307, + 0.008861420676112175, + 0.0074654193595051765, + 0.005175842437893152, + 0.005445003043860197, + 0.014058663509786129, + 0.012168975546956062, + 0.012586694210767746, + 0.009663090109825134, + 0.011202567256987095, + 0.012976327911019325, + 0.006370905786752701, + 0.007536657154560089, + 0.006364976521581411, + 0.005710927303880453, + 0.0061730979941785336, + 0.004600488115102053, + 0.007534539792686701, + 0.007056650705635548, + 0.0041910698637366295, + 0.0040732864290475845, + 0.008587406948208809, + 0.007013504859060049, + 0.006633422337472439, + 0.006057687569409609, + 0.006558684166520834, + 0.010151704773306847, + 0.004399307072162628, + 0.006432723719626665, + 0.01355994213372469, + 0.011971808038651943, + 0.01880640536546707, + 0.00755050964653492, + 0.015556399710476398, + 0.02171868272125721, + 0.01139823254197836, + 0.007859769277274609, + 0.012250860221683979, + 0.0077613480389118195, + 0.011575707234442234, + 0.00937388651072979, + 0.01925445720553398, + 0.017018267884850502, + 0.012739362195134163, + 0.0064368052408099174, + 0.019689835608005524, + 0.013459490612149239, + 0.017462490126490593, + 0.009548117406666279, + 0.01286265254020691, + 0.01446750108152628, + 0.009806348942220211, + 0.008507753722369671, + 0.00960969552397728, + 0.010486303828656673, + 0.007318310439586639, + 0.004635457415133715, + 0.006339200772345066, + 0.007643496617674828, + 0.004328824579715729, + 0.005208469927310944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0008133916999213398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle focuses on an exquisite music box, slowly rotating on a dark, polished wooden table. Crafted from polished mahogany, the rectangular box's sides are adorned with delicate floral inlays of mother-of-pearl. Warmly illuminated by a single source, its surface reflects a soft sheen. Atop the music box, a porcelain ballerina doll in a layered tutu pirouettes gracefully and precisely in sync with the base's rotation. Standing on tiptoes, her arms elegantly raised overhead, tiny crystals adorning the hem of her dress catch and scatter light as she moves, creating fleeting flashes. Soft golden light streams in from the side, illuminating the edges of the music box and casting shifting, soft shadows across the tabletop. The background is a blurred, warm-toned study. The scene is shrouded in a tranquil, nostalgic, and slightly magical atmosphere, as if time has frozen in time, leaving only this eternal, graceful dance.", + "seed": 22560, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_10/index_3560_seed_22560_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": -0.04564285373780876, + "vlm_logprob": -0.04564285373780876, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.04564285373780876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.18611231446266174 + ], + "opposite_logprob_tokens": [ + -1.8111122846603394 + ], + "logprob_reward": -0.18611231446266174, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18611231446266174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17574156820774078, + 0.08054111897945404, + 0.05704766884446144, + 0.04221462458372116, + 0.22193266451358795, + 0.17156250774860382, + 0.14034605026245117, + 0.02405674010515213, + 0.025318104773759842, + 0.06123894453048706 + ], + "credit_map": [ + 0.021051321178674698, + 0.016839217394590378, + 0.029592975974082947, + 0.02684023790061474, + 0.017040977254509926, + 0.018677692860364914, + 0.026115799322724342, + 0.019583355635404587, + 0.007285004947334528, + 0.008482336066663265, + 0.013874072581529617, + 0.011215165257453918, + 0.012407548725605011, + 0.008612548932433128, + 0.008765405975282192, + 0.009899042546749115, + 0.005159751512110233, + 0.005315145943313837, + 0.007549948524683714, + 0.009826822206377983, + 0.007149118930101395, + 0.006096441764384508, + 0.00992076937109232, + 0.0060296752490103245, + 0.0038349300157278776, + 0.0036422645207494497, + 0.006952069234102964, + 0.00555089395493269, + 0.004514321219176054, + 0.004714245442301035, + 0.007503867615014315, + 0.005502036772668362, + 0.019398635253310204, + 0.02174992673099041, + 0.033198945224285126, + 0.027933934703469276, + 0.024513594806194305, + 0.025493813678622246, + 0.04003128036856651, + 0.029612554237246513, + 0.01724673993885517, + 0.011749510653316975, + 0.030187658965587616, + 0.03257158398628235, + 0.018941326066851616, + 0.019893130287528038, + 0.024607941508293152, + 0.01636464148759842, + 0.020691480487585068, + 0.009585096500813961, + 0.019127827137708664, + 0.01984858140349388, + 0.016230057924985886, + 0.01620691828429699, + 0.02282065711915493, + 0.015835432335734367, + 0.0026121740229427814, + 0.0023570030461996794, + 0.0028309535700827837, + 0.003009095322340727, + 0.004755697678774595, + 0.0026521484833210707, + 0.0033039706759154797, + 0.002535699401050806, + 0.0024461247958242893, + 0.002800185466185212, + 0.0037650989834219217, + 0.002642520237714052, + 0.0051542771980166435, + 0.0028413909021764994, + 0.003171063493937254, + 0.002497443463653326, + 0.008355467580258846, + 0.005727103911340237, + 0.015102776698768139, + 0.005868901964277029, + 0.006656670942902565, + 0.007445291616022587, + 0.005314499139785767, + 0.006768235936760902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.18611231446266174 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.005263044033199549 + ], + "opposite_logprob_tokens": [ + -5.255262851715088 + ], + "logprob_reward": -0.005263044033199549, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263044033199549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32323309779167175, + 0.09099749475717545, + 0.06485868990421295, + 0.06700906157493591, + 0.11969222873449326, + 0.08078542351722717, + 0.13441868126392365, + 0.04025561735033989, + 0.01695164479315281, + 0.06179812550544739 + ], + "credit_map": [ + 0.039939023554325104, + 0.032189078629016876, + 0.053047072142362595, + 0.04988843947649002, + 0.03132586553692818, + 0.03221035748720169, + 0.03975214436650276, + 0.04488111659884453, + 0.008362486027181149, + 0.010152257978916168, + 0.012039819732308388, + 0.01018589362502098, + 0.015703042969107628, + 0.010623791255056858, + 0.01260401401668787, + 0.011326184496283531, + 0.007222182583063841, + 0.005973772145807743, + 0.009607546962797642, + 0.008536022156476974, + 0.011724330484867096, + 0.006931796204298735, + 0.0072142137214541435, + 0.007648826111108065, + 0.008573233149945736, + 0.0060996077954769135, + 0.014558178372681141, + 0.007121536880731583, + 0.007342244032770395, + 0.007647686637938023, + 0.008703761734068394, + 0.0069628129713237286, + 0.01095068734139204, + 0.010651788674294949, + 0.01663329266011715, + 0.017712291330099106, + 0.012527178972959518, + 0.013903376646339893, + 0.02198357693850994, + 0.01533003244549036, + 0.006824451964348555, + 0.005916766822338104, + 0.01574515923857689, + 0.01623362861573696, + 0.008522740565240383, + 0.009273025207221508, + 0.011461739428341389, + 0.006807905156165361, + 0.019560448825359344, + 0.01053357869386673, + 0.017693201079964638, + 0.01641371101140976, + 0.017066625878214836, + 0.015522109344601631, + 0.021304305642843246, + 0.01632470265030861, + 0.004461462143808603, + 0.004087634850293398, + 0.005579885561019182, + 0.00539968628436327, + 0.007393693085759878, + 0.0047960844822227955, + 0.0047550927847623825, + 0.0037820786237716675, + 0.0017524291761219501, + 0.001868339371867478, + 0.0020684560295194387, + 0.002495847176760435, + 0.002723697340115905, + 0.0019757039844989777, + 0.0021132321562618017, + 0.001953938975930214, + 0.008822265081107616, + 0.005862937308847904, + 0.014592415653169155, + 0.007039121817797422, + 0.005503340158611536, + 0.007840263657271862, + 0.004874694161117077, + 0.007263083476573229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.005263044033199549 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.033716656267642975 + ], + "opposite_logprob_tokens": [ + -3.4087166786193848 + ], + "logprob_reward": -0.033716656267642975, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033716656267642975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36783894896507263, + 0.07365438342094421, + 0.055382754653692245, + 0.09785638004541397, + 0.0874897763133049, + 0.07864735275506973, + 0.0884098932147026, + 0.06240725517272949, + 0.03320189565420151, + 0.05511133372783661 + ], + "credit_map": [ + 0.045273132622241974, + 0.030492868274450302, + 0.06607472896575928, + 0.048978716135025024, + 0.03656638413667679, + 0.03983504697680473, + 0.040707461535930634, + 0.059910621494054794, + 0.007439511362463236, + 0.00721803680062294, + 0.010122494772076607, + 0.008276913315057755, + 0.011151784099638462, + 0.008876769803464413, + 0.00968018639832735, + 0.01088869571685791, + 0.004518519621342421, + 0.006497221067547798, + 0.008520728908479214, + 0.008056798949837685, + 0.008047137409448624, + 0.005800532177090645, + 0.007608836516737938, + 0.006332981400191784, + 0.01450849324464798, + 0.008446276187896729, + 0.020462803542613983, + 0.009985088370740414, + 0.010106023401021957, + 0.011576991528272629, + 0.012028978206217289, + 0.010741733945906162, + 0.006653310265392065, + 0.0078085302375257015, + 0.013422900810837746, + 0.013622009195387363, + 0.01041042897850275, + 0.010020021349191666, + 0.014390295371413231, + 0.01116228848695755, + 0.006792533677071333, + 0.007362863514572382, + 0.012156402692198753, + 0.011840185150504112, + 0.012422660365700722, + 0.009152596816420555, + 0.0093033192679286, + 0.009616791270673275, + 0.010551383718848228, + 0.009028575383126736, + 0.010818379931151867, + 0.011073743924498558, + 0.012962070293724537, + 0.009952272288501263, + 0.0134152602404356, + 0.010608199052512646, + 0.010504418052732944, + 0.005521461367607117, + 0.006586202420294285, + 0.007711838465183973, + 0.00793902762234211, + 0.0073439073748886585, + 0.007584028411656618, + 0.009216377511620522, + 0.003221233608201146, + 0.0040245153941214085, + 0.004609778523445129, + 0.003532653907313943, + 0.006928960792720318, + 0.0038687200285494328, + 0.003651710692793131, + 0.0033643266651779413, + 0.007232291158288717, + 0.005360797513276339, + 0.012082122266292572, + 0.005419627763330936, + 0.007763091940432787, + 0.005933702923357487, + 0.005550874397158623, + 0.005768831819295883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.033716656267642975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0024856405798345804 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024856405798345804, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024856405798345804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22881178557872772, + 0.06419919431209564, + 0.07369639724493027, + 0.060525234788656235, + 0.1265728622674942, + 0.07026050239801407, + 0.1611870676279068, + 0.03957243636250496, + 0.03473319485783577, + 0.14044125378131866 + ], + "credit_map": [ + 0.0222189798951149, + 0.024079112336039543, + 0.039203718304634094, + 0.042083218693733215, + 0.026468021795153618, + 0.025818519294261932, + 0.024374855682253838, + 0.024565374478697777, + 0.007087474688887596, + 0.006676536984741688, + 0.0112466961145401, + 0.010658420622348785, + 0.008007063530385494, + 0.007079198025166988, + 0.006722329184412956, + 0.006721476558595896, + 0.00926845520734787, + 0.006244407966732979, + 0.013691389001905918, + 0.011515798047184944, + 0.008144620805978775, + 0.008268775418400764, + 0.008010498248040676, + 0.008552461862564087, + 0.005885078106075525, + 0.007217009551823139, + 0.009722310118377209, + 0.008915839716792107, + 0.009258582256734371, + 0.0058906301856040955, + 0.0061246776022017, + 0.007511115167289972, + 0.010647621937096119, + 0.012727468274533749, + 0.018686898052692413, + 0.01801842637360096, + 0.01397086214274168, + 0.014468814246356487, + 0.021764781326055527, + 0.016287999227643013, + 0.007250909227877855, + 0.005773462355136871, + 0.012446246109902859, + 0.01156341191381216, + 0.007112463004887104, + 0.007760727312415838, + 0.008488924242556095, + 0.009864370338618755, + 0.02640640363097191, + 0.015434622764587402, + 0.023277951404452324, + 0.01956298016011715, + 0.02013811469078064, + 0.018720805644989014, + 0.021861962974071503, + 0.015784237533807755, + 0.0068140714429318905, + 0.003389715449884534, + 0.0038281383458524942, + 0.006657224614173174, + 0.004861457273364067, + 0.004965426865965128, + 0.00469958083704114, + 0.00435682525858283, + 0.003632820211350918, + 0.0037690899334847927, + 0.006739526055753231, + 0.0036381578538566828, + 0.00473226560279727, + 0.0040836455300450325, + 0.003382174065336585, + 0.004755518864840269, + 0.017523322254419327, + 0.01103720162063837, + 0.028581440448760986, + 0.019296109676361084, + 0.014166378416121006, + 0.015861736610531807, + 0.01385424006730318, + 0.020120825618505478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0024856405798345804 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rectangular music box crafted from polished mahogany with mother-of-pearl floral inlays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a porcelain ballerina doll wearing a layered tutu standing on top of the music box?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ballerina doll pirouetting with arms raised overhead while tiny crystals on her dress scatter light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dark, polished wooden table with a blurred, warm-toned study background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from a slightly overhead angle with soft golden lighting casting shifting 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": [ + -0.0006366133457049727 + ], + "opposite_logprob_tokens": [ + -7.375636577606201 + ], + "logprob_reward": -0.0006366133457049727, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006366133457049727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26339253783226013, + 0.0834292620420456, + 0.052083857357501984, + 0.06075090169906616, + 0.16093143820762634, + 0.13234473764896393, + 0.09664607048034668, + 0.048101089894771576, + 0.04590151458978653, + 0.056418586522340775 + ], + "credit_map": [ + 0.023179765790700912, + 0.026313690468668938, + 0.04550983011722565, + 0.04249970614910126, + 0.03215375542640686, + 0.031570855528116226, + 0.03279520198702812, + 0.02936975099146366, + 0.007973217405378819, + 0.008398134261369705, + 0.013207770884037018, + 0.010294460691511631, + 0.016024315729737282, + 0.00964776799082756, + 0.008291325531899929, + 0.009592270478606224, + 0.004923433531075716, + 0.005325619596987963, + 0.007848566398024559, + 0.007498830556869507, + 0.007806952111423016, + 0.00548520078882575, + 0.007517592515796423, + 0.005677661392837763, + 0.007410718128085136, + 0.005202779080718756, + 0.011046850122511387, + 0.008426085114479065, + 0.006653286051005125, + 0.006526114419102669, + 0.008533400483429432, + 0.006951671559363604, + 0.01731509529054165, + 0.015041124075651169, + 0.026089701801538467, + 0.019491303712129593, + 0.01687854342162609, + 0.017718851566314697, + 0.02743065543472767, + 0.02096615917980671, + 0.012834339402616024, + 0.008691942319273949, + 0.030428007245063782, + 0.024174675345420837, + 0.014064216054975986, + 0.014469260349869728, + 0.015621758066117764, + 0.012060534209012985, + 0.008052310906350613, + 0.012224960140883923, + 0.013851944357156754, + 0.013819515705108643, + 0.01990741863846779, + 0.010120507329702377, + 0.009992328472435474, + 0.00867708120495081, + 0.005143012385815382, + 0.003530343296006322, + 0.007250282447785139, + 0.005706751719117165, + 0.0074664936400949955, + 0.0056023006327450275, + 0.0068264976143836975, + 0.006575408391654491, + 0.004895901773124933, + 0.0048105474561452866, + 0.007858901284635067, + 0.0049019018188118935, + 0.008193142712116241, + 0.004953354597091675, + 0.00614527240395546, + 0.004142492078244686, + 0.0056612128391861916, + 0.0050226496532559395, + 0.008786352351307869, + 0.005691846366971731, + 0.01158191915601492, + 0.00678204745054245, + 0.006713840179145336, + 0.006178714334964752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0006366133457049727 + } + ], + "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/0000052/group_11/index_tensor([1218, 1218])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_11/index_tensor([1218, 1218])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1c266540291e0d3b3bcdbfa534f964f821901bee --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_11/index_tensor([1218, 1218])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, captured in a highly detailed 4K vertical format, uses a nearly static, horizontal, handheld perspective to transport the viewer into the vibrant atmosphere of a Japanese tavern. Warm afternoon sunlight streams through a wooden-lattice window on one side of the frame, creating a distinct Tyndall-effect beam in the air, illuminating floating dust particles and an indigo-colored fabric tavern flag suspended from the ceiling, gently fluttering in the air currents. At the center of the frame is a young male musician with slightly tousled dark brown curls. Wearing a washed-blue denim jacket and a white cotton T-shirt, he sits comfortably on a high wooden stool, cradling a natural-wood acoustic guitar. His eyes are slightly closed, a smile of utter immersion on his face. His fingers steadily and deliberately pluck and press the strings, not playing a set piece but engaging in a melodious, spontaneous improvisation. Around him, several friends sat scattered around a rough wooden table, holding glasses of amber-colored fruit wine, tiny droplets of condensation clinging to the sides. Their faces were filled with relaxed smiles, their bodies swaying gently to the rhythm of the music, and they chatted quietly. The entire scene was dominated by warm wood and brick red tones, with the golden highlights of sunlight illuminating the picture, creating an extremely casual, warm, and joyful atmosphere, perfectly capturing a genuine and beautiful afternoon social moment.", + "seed": 3910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_11/index_1218_seed_3910_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": -0.07328981403261423, + "vlm_logprob": -0.07328981403261423, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.07328981403261423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18313288688659668 + ], + "opposite_logprob_tokens": [ + -1.8081328868865967 + ], + "logprob_reward": -0.18313288688659668, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18313288688659668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32270845770835876, + 0.08683460205793381, + 0.17669343948364258, + 0.06627233326435089, + 0.1024957224726677, + 0.1098700687289238, + 0.05337574705481529, + 0.0470416396856308, + 0.015685537829995155, + 0.019022472202777863 + ], + "credit_map": [ + 0.03264777362346649, + 0.04589506983757019, + 0.03600306436419487, + 0.035725172609090805, + 0.031816814094781876, + 0.04310417175292969, + 0.04087020829319954, + 0.056646160781383514, + 0.006192373111844063, + 0.0075638494454324245, + 0.012019412592053413, + 0.012084998190402985, + 0.013200468383729458, + 0.006865473929792643, + 0.013181941583752632, + 0.015726080164313316, + 0.021968133747577667, + 0.01708776131272316, + 0.019764330238103867, + 0.044878143817186356, + 0.012230063788592815, + 0.009691968560218811, + 0.023151395842432976, + 0.027921630069613457, + 0.004464900586754084, + 0.004632778465747833, + 0.008096137084066868, + 0.010149107314646244, + 0.008711584843695164, + 0.007655885070562363, + 0.010730904527008533, + 0.011831028386950493, + 0.011979642324149609, + 0.013562195003032684, + 0.011427991092205048, + 0.009598600678145885, + 0.009410805068910122, + 0.01178696658462286, + 0.01200698409229517, + 0.022722523659467697, + 0.007015731651335955, + 0.00773728359490633, + 0.016508428379893303, + 0.013956885784864426, + 0.013184848241508007, + 0.009835975244641304, + 0.024338020011782646, + 0.017292886972427368, + 0.005750114098191261, + 0.004052397795021534, + 0.007559001911431551, + 0.007122730370610952, + 0.005112417973577976, + 0.004930166061967611, + 0.00991696584969759, + 0.008931947872042656, + 0.0033926903270184994, + 0.005365286488085985, + 0.005479208659380674, + 0.006515876855701208, + 0.005349001381546259, + 0.006606061942875385, + 0.007876596413552761, + 0.0064569138921797276, + 0.0013731771614402533, + 0.0017505766591057181, + 0.0017994402442127466, + 0.0024828778114169836, + 0.001559560652822256, + 0.0018423105357214808, + 0.002336766105145216, + 0.0025408274959772825, + 0.0019176389323547482, + 0.00323276873677969, + 0.002359948819503188, + 0.002905501052737236, + 0.00236823339946568, + 0.002099729375913739, + 0.00164409214630723, + 0.0024945586919784546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.18313288688659668 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077570728957653 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077570728957653, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077570728957653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21865811944007874, + 0.11271559447050095, + 0.28863194584846497, + 0.04807397723197937, + 0.07016571611166, + 0.05680548772215843, + 0.07494485378265381, + 0.033143069595098495, + 0.03874972462654114, + 0.05811145901679993 + ], + "credit_map": [ + 0.011920972727239132, + 0.018145296722650528, + 0.030515894293785095, + 0.05596495792269707, + 0.019702119752764702, + 0.014977185986936092, + 0.03582214564085007, + 0.03160954266786575, + 0.006942537613213062, + 0.00839470699429512, + 0.018673041835427284, + 0.027457501739263535, + 0.0088789202272892, + 0.0076989284716546535, + 0.014861058443784714, + 0.019808894023299217, + 0.017947262153029442, + 0.030169423669576645, + 0.0428386814892292, + 0.04144173488020897, + 0.05601188540458679, + 0.017976948991417885, + 0.04380359500646591, + 0.03844241052865982, + 0.00339947990141809, + 0.003932557068765163, + 0.006150334607809782, + 0.00799139216542244, + 0.006881507113575935, + 0.004786214325577021, + 0.0069353533908724785, + 0.007997140288352966, + 0.0056423163041472435, + 0.00638195127248764, + 0.00845553632825613, + 0.010190674103796482, + 0.011299340054392815, + 0.004605851136147976, + 0.012105735950171947, + 0.011484317481517792, + 0.004520038608461618, + 0.004020208027213812, + 0.010823464952409267, + 0.008199931122362614, + 0.005569371860474348, + 0.004697357304394245, + 0.010848321951925755, + 0.008126791566610336, + 0.005140264984220266, + 0.004955044947564602, + 0.011987963691353798, + 0.008497782982885838, + 0.009330137632787228, + 0.008014057762920856, + 0.01476660929620266, + 0.0122529910877347, + 0.0031737990211695433, + 0.0031902205664664507, + 0.004386612679809332, + 0.003955658990889788, + 0.004237210378050804, + 0.002484762342646718, + 0.006290697492659092, + 0.00542410695925355, + 0.00393662927672267, + 0.003594903275370598, + 0.006793055217713118, + 0.005886014550924301, + 0.002968203043565154, + 0.003718046937137842, + 0.004871339537203312, + 0.006981529761105776, + 0.004651615861803293, + 0.007870621979236603, + 0.007708971854299307, + 0.014199299737811089, + 0.004112941678613424, + 0.0029590572230517864, + 0.007201287895441055, + 0.009407666511833668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.011077570728957653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329017035663128 + ], + "opposite_logprob_tokens": [ + -3.7732901573181152 + ], + "logprob_reward": -0.02329017035663128, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329017035663128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2415241152048111, + 0.19019852578639984, + 0.25221383571624756, + 0.06571043282747269, + 0.04457363486289978, + 0.05820700153708458, + 0.039557334035634995, + 0.02473379299044609, + 0.04855763167142868, + 0.034723661839962006 + ], + "credit_map": [ + 0.01254142913967371, + 0.018322400748729706, + 0.03107527829706669, + 0.06496044993400574, + 0.025097282603383064, + 0.01325386855751276, + 0.04098924994468689, + 0.0352841392159462, + 0.009954029694199562, + 0.013469083234667778, + 0.028205506503582, + 0.05122506618499756, + 0.01889977417886257, + 0.008813505992293358, + 0.028704514726996422, + 0.030927054584026337, + 0.028816841542720795, + 0.04006314277648926, + 0.026607464998960495, + 0.051786504685878754, + 0.019182972609996796, + 0.01235588826239109, + 0.03403448686003685, + 0.039366547018289566, + 0.004025757312774658, + 0.004821574781090021, + 0.008676995523273945, + 0.014960912987589836, + 0.00466431537643075, + 0.006528101861476898, + 0.010686869733035564, + 0.01134591456502676, + 0.003631126368418336, + 0.003664483316242695, + 0.006015590392053127, + 0.009580140002071857, + 0.004344554152339697, + 0.004796179942786694, + 0.004994256887584925, + 0.007547303568571806, + 0.005156687460839748, + 0.004165107849985361, + 0.011953468434512615, + 0.007933450862765312, + 0.005544624757021666, + 0.004939033649861813, + 0.010078033432364464, + 0.008436594158411026, + 0.005373658146709204, + 0.004997420124709606, + 0.004861156921833754, + 0.005619235802441835, + 0.00391212897375226, + 0.0042989207431674, + 0.005198350176215172, + 0.005296461749821901, + 0.001928636571392417, + 0.0028323011938482523, + 0.00379289872944355, + 0.0027439258992671967, + 0.0031399233266711235, + 0.002659458201378584, + 0.003882728749886155, + 0.003753920551389456, + 0.004384983796626329, + 0.0037922379560768604, + 0.00871310941874981, + 0.007825320586562157, + 0.005618889816105366, + 0.004770845174789429, + 0.006383876781910658, + 0.0070683653466403484, + 0.0030352522153407335, + 0.003737062681466341, + 0.004306668881326914, + 0.007756615057587624, + 0.003091891063377261, + 0.0024842913262546062, + 0.004773358348757029, + 0.005538521334528923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02329017035663128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030331462621689 + ], + "opposite_logprob_tokens": [ + -2.3503034114837646 + ], + "logprob_reward": -0.10030331462621689, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030331462621689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.319053053855896, + 0.08829295635223389, + 0.11383106559515, + 0.09188687056303024, + 0.08036211133003235, + 0.066172294318676, + 0.08948180079460144, + 0.09765855222940445, + 0.03233128786087036, + 0.02092999964952469 + ], + "credit_map": [ + 0.02217739447951317, + 0.027479689568281174, + 0.04247748479247093, + 0.07014188170433044, + 0.02380274422466755, + 0.026081575080752373, + 0.0553155355155468, + 0.05157678946852684, + 0.004722438752651215, + 0.005429147742688656, + 0.014178181067109108, + 0.020389262586832047, + 0.008620000444352627, + 0.005464745685458183, + 0.015283276326954365, + 0.014205912128090858, + 0.019365202635526657, + 0.019534002989530563, + 0.01222254615277052, + 0.017859753221273422, + 0.010967615991830826, + 0.008020314387977123, + 0.011675943620502949, + 0.014185687527060509, + 0.007988646626472473, + 0.014483131468296051, + 0.010144690982997417, + 0.018624048680067062, + 0.008844084106385708, + 0.007049516774713993, + 0.013058946467936039, + 0.011693811975419521, + 0.008157278411090374, + 0.007916880771517754, + 0.010019876062870026, + 0.01513082068413496, + 0.008331711404025555, + 0.006057546474039555, + 0.012064625509083271, + 0.012683375738561153, + 0.009339549578726292, + 0.004688046872615814, + 0.009439210407435894, + 0.00949984323233366, + 0.006740759126842022, + 0.004817579872906208, + 0.012734174728393555, + 0.008913136087357998, + 0.009793124161660671, + 0.008897203952074051, + 0.015717821195721626, + 0.00906274002045393, + 0.00904006790369749, + 0.00677590724080801, + 0.018991999328136444, + 0.011202939786016941, + 0.006442110054194927, + 0.007192353717982769, + 0.011386515572667122, + 0.016104746609926224, + 0.012868897058069706, + 0.01026653591543436, + 0.01523082796484232, + 0.018166573718190193, + 0.0020475799683481455, + 0.002907631453126669, + 0.0051402803510427475, + 0.006266999524086714, + 0.0041931369341909885, + 0.002632823772728443, + 0.004429192282259464, + 0.004713646601885557, + 0.002094054827466607, + 0.002998025855049491, + 0.002201674971729517, + 0.003697037696838379, + 0.0023511708714067936, + 0.0015918051358312368, + 0.0029016579501330853, + 0.003094573738053441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.10030331462621689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048645127564668655 + ], + "opposite_logprob_tokens": [ + -3.04864501953125 + ], + "logprob_reward": -0.048645127564668655, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048645127564668655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21106049418449402, + 0.20526733994483948, + 0.06153741106390953, + 0.1363913118839264, + 0.06669919937849045, + 0.07673028111457825, + 0.0711265429854393, + 0.07225826382637024, + 0.04501882195472717, + 0.05391030013561249 + ], + "credit_map": [ + 0.014351117424666882, + 0.023026395589113235, + 0.02617643028497696, + 0.038708191365003586, + 0.016245802864432335, + 0.02264406718313694, + 0.03650385141372681, + 0.0334046296775341, + 0.013795332051813602, + 0.028557490557432175, + 0.029657788574695587, + 0.026369787752628326, + 0.023877613246440887, + 0.023560907691717148, + 0.021961286664009094, + 0.037487130612134933, + 0.00835833977907896, + 0.013693713583052158, + 0.005256607662886381, + 0.00858053658157587, + 0.007611855398863554, + 0.0041286274790763855, + 0.0063813370652496815, + 0.00752638652920723, + 0.009098675101995468, + 0.012408025562763214, + 0.016307814046740532, + 0.01795106753706932, + 0.01673933118581772, + 0.01767883263528347, + 0.018825886771082878, + 0.02738167904317379, + 0.008577256463468075, + 0.011727157048881054, + 0.007070140913128853, + 0.006156155839562416, + 0.0075862593948841095, + 0.008409970439970493, + 0.007405212614685297, + 0.009767047129571438, + 0.006198323797434568, + 0.006142669357359409, + 0.013670118525624275, + 0.008966336026787758, + 0.008288159966468811, + 0.006472496781498194, + 0.01642110012471676, + 0.010571076534688473, + 0.008406397886574268, + 0.005525471642613411, + 0.011414701119065285, + 0.006843831390142441, + 0.006882602348923683, + 0.005840046796947718, + 0.015951214358210564, + 0.010262275114655495, + 0.005329330451786518, + 0.010974040254950523, + 0.008634411729872227, + 0.008032206445932388, + 0.00725279888138175, + 0.00824310164898634, + 0.012513100169599056, + 0.011279268190264702, + 0.004263849463313818, + 0.007996993139386177, + 0.00478782132267952, + 0.004581317771226168, + 0.005381748080253601, + 0.006058079190552235, + 0.006420854013413191, + 0.005528157576918602, + 0.0045700883492827415, + 0.00725985923781991, + 0.005555687006562948, + 0.005972355138510466, + 0.005571182817220688, + 0.009648644365370274, + 0.008060828782618046, + 0.007271654903888702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.048645127564668655 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured in a highly detailed 4K vertical format, uses a nearly static, horizontal, handheld perspective to transport the viewer into the vibrant atmosphere of a Japanese tavern. Warm afternoon sunlight streams through a wooden-lattice window on one side of the frame, creating a distinct Tyndall-effect beam in the air, illuminating floating dust particles and an indigo-colored fabric tavern flag suspended from the ceiling, gently fluttering in the air currents. At the center of the frame is a young male musician with slightly tousled dark brown curls. Wearing a washed-blue denim jacket and a white cotton T-shirt, he sits comfortably on a high wooden stool, cradling a natural-wood acoustic guitar. His eyes are slightly closed, a smile of utter immersion on his face. His fingers steadily and deliberately pluck and press the strings, not playing a set piece but engaging in a melodious, spontaneous improvisation. Around him, several friends sat scattered around a rough wooden table, holding glasses of amber-colored fruit wine, tiny droplets of condensation clinging to the sides. Their faces were filled with relaxed smiles, their bodies swaying gently to the rhythm of the music, and they chatted quietly. The entire scene was dominated by warm wood and brick red tones, with the golden highlights of sunlight illuminating the picture, creating an extremely casual, warm, and joyful atmosphere, perfectly capturing a genuine and beautiful afternoon social moment.", + "seed": 3911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_11/index_1218_seed_3911_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": -0.07709804140031337, + "vlm_logprob": -0.07709804140031337, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.07709804140031337, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07044186443090439 + ], + "opposite_logprob_tokens": [ + -2.820441961288452 + ], + "logprob_reward": -0.07044186443090439, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07044186443090439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2985767424106598, + 0.0867636501789093, + 0.04719731584191322, + 0.030949698761105537, + 0.12010509520769119, + 0.05059380084276199, + 0.02595955692231655, + 0.04664978012442589, + 0.22610753774642944, + 0.06709688901901245 + ], + "credit_map": [ + 0.03100365400314331, + 0.04081340506672859, + 0.033184848725795746, + 0.037968266755342484, + 0.034302081912755966, + 0.03400309383869171, + 0.049908507615327835, + 0.03739288076758385, + 0.007685074117034674, + 0.009690498933196068, + 0.007556670345366001, + 0.016290519386529922, + 0.014366035349667072, + 0.008991130627691746, + 0.012502788566052914, + 0.009680929593741894, + 0.003280957695096731, + 0.004592489916831255, + 0.00601860461756587, + 0.008132890798151493, + 0.004354138392955065, + 0.004083195235580206, + 0.007223454304039478, + 0.009511583484709263, + 0.0022168324794620275, + 0.003667430253699422, + 0.0031785094179213047, + 0.003944051451981068, + 0.0062490347772836685, + 0.0033453910145908594, + 0.005017342045903206, + 0.0033311056904494762, + 0.004730375483632088, + 0.009145746938884258, + 0.008041811175644398, + 0.020148402079939842, + 0.005857651587575674, + 0.021512674167752266, + 0.042576469480991364, + 0.008091953583061695, + 0.004111868795007467, + 0.007655766326934099, + 0.0048614791594445705, + 0.007925204001367092, + 0.006071912124752998, + 0.005896709859371185, + 0.007954470813274384, + 0.006116383243352175, + 0.002005031332373619, + 0.004135061055421829, + 0.002431100932881236, + 0.0037610274739563465, + 0.0027491110377013683, + 0.003113635815680027, + 0.003878421848639846, + 0.0038861653301864862, + 0.003238068427890539, + 0.0054089585319161415, + 0.005514717195183039, + 0.008745481260120869, + 0.006367129739373922, + 0.004082455765455961, + 0.008261164650321007, + 0.005031798500567675, + 0.01201603189110756, + 0.020089834928512573, + 0.01695871911942959, + 0.032210979610681534, + 0.033660709857940674, + 0.04339442402124405, + 0.030531195923686028, + 0.03724560886621475, + 0.003564612939953804, + 0.004069156479090452, + 0.019469957798719406, + 0.007516081910580397, + 0.004708600230515003, + 0.011357298120856285, + 0.012705429457128048, + 0.003705749986693263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.07044186443090439 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352349668741226 + ], + "opposite_logprob_tokens": [ + -3.6513524055480957 + ], + "logprob_reward": -0.026352349668741226, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352349668741226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3890650272369385, + 0.06253848969936371, + 0.07922434061765671, + 0.11854185163974762, + 0.06815102696418762, + 0.041839417070150375, + 0.0540442131459713, + 0.046903349459171295, + 0.10414187610149384, + 0.0355503186583519 + ], + "credit_map": [ + 0.03307342529296875, + 0.03467009961605072, + 0.05112907662987709, + 0.06041102856397629, + 0.03769956901669502, + 0.03651117533445358, + 0.06864143162965775, + 0.06692925095558167, + 0.005696165841072798, + 0.008023134432733059, + 0.009908958338201046, + 0.008809998631477356, + 0.006997482385486364, + 0.006846908014267683, + 0.008814087137579918, + 0.007441757712513208, + 0.007345317862927914, + 0.011733664199709892, + 0.009657114744186401, + 0.010961621068418026, + 0.0071978941559791565, + 0.006652327254414558, + 0.014612099155783653, + 0.011064304038882256, + 0.007711346261203289, + 0.012518721632659435, + 0.01780780963599682, + 0.01651039719581604, + 0.022085949778556824, + 0.008406894281506538, + 0.01995011419057846, + 0.013550627045333385, + 0.0029706861823797226, + 0.005574048962444067, + 0.005243336781859398, + 0.0106381606310606, + 0.004475789610296488, + 0.010304818861186504, + 0.021026263013482094, + 0.007917925715446472, + 0.0036518701817840338, + 0.0039891465567052364, + 0.0042643314227461815, + 0.006290595978498459, + 0.004282452631741762, + 0.00416660588234663, + 0.008338043466210365, + 0.006856375839561224, + 0.003564164275303483, + 0.00726658059284091, + 0.008316874504089355, + 0.0077849747613072395, + 0.007877929136157036, + 0.006208162289112806, + 0.0072198123671114445, + 0.005805722437798977, + 0.003311695996671915, + 0.004256722051650286, + 0.00555633706972003, + 0.007125884294509888, + 0.006660267245024443, + 0.004483913537114859, + 0.008054647594690323, + 0.007453883066773415, + 0.005975101143121719, + 0.012051920406520367, + 0.009859335608780384, + 0.02466776967048645, + 0.013924864120781422, + 0.008149439468979836, + 0.014822992496192455, + 0.014690456911921501, + 0.001985671930015087, + 0.0041510299779474735, + 0.003328390419483185, + 0.006612613797187805, + 0.004582697059959173, + 0.0031226801220327616, + 0.008575951680541039, + 0.0031912836711853743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.026352349668741226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822157695889473 + ], + "opposite_logprob_tokens": [ + -3.5298221111297607 + ], + "logprob_reward": -0.029822157695889473, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822157695889473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2546362280845642, + 0.050201013684272766, + 0.051050182431936264, + 0.10934440046548843, + 0.15956330299377441, + 0.029668327420949936, + 0.02701650932431221, + 0.03875485062599182, + 0.1974756270647049, + 0.08228958398103714 + ], + "credit_map": [ + 0.01925874687731266, + 0.02576417662203312, + 0.029277849942445755, + 0.0405614972114563, + 0.02605700120329857, + 0.03352726623415947, + 0.03253389894962311, + 0.04765576496720314, + 0.004512066952884197, + 0.006577374413609505, + 0.00783684104681015, + 0.006865737494081259, + 0.006132324226200581, + 0.004156302660703659, + 0.006849315948784351, + 0.007271043956279755, + 0.005206930451095104, + 0.005485464818775654, + 0.0054168920032680035, + 0.006685951258987188, + 0.0040794601663947105, + 0.005167357157915831, + 0.008444971404969692, + 0.010563146322965622, + 0.005436760373413563, + 0.012117309495806694, + 0.013389968313276768, + 0.014918426051735878, + 0.024171818047761917, + 0.00719566410407424, + 0.020859602838754654, + 0.01125484611839056, + 0.006503247190266848, + 0.013213342055678368, + 0.010126971639692783, + 0.026217153295874596, + 0.008769012056291103, + 0.02808062545955181, + 0.054690808057785034, + 0.011962125077843666, + 0.00229465845040977, + 0.003518028650432825, + 0.002890867181122303, + 0.0046756211668252945, + 0.0029602565336972475, + 0.003263271413743496, + 0.006197675596922636, + 0.0038679472636431456, + 0.001831735367886722, + 0.003862005891278386, + 0.003763791173696518, + 0.004185707774013281, + 0.0029497314244508743, + 0.0029242797754704952, + 0.0038132499903440475, + 0.003686003852635622, + 0.003680557943880558, + 0.005308160558342934, + 0.004070436116307974, + 0.004878210835158825, + 0.005819548387080431, + 0.0038005246315151453, + 0.004887823481112719, + 0.006309584714472294, + 0.009619774296879768, + 0.01709151081740856, + 0.016408098861575127, + 0.028773976489901543, + 0.03003634884953499, + 0.03405248746275902, + 0.02567802555859089, + 0.035815395414829254, + 0.00428909994661808, + 0.0034478853922337294, + 0.02594735473394394, + 0.008716466836631298, + 0.005233416799455881, + 0.012849792838096619, + 0.0174274705350399, + 0.004378089681267738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.029822157695889473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898139953613281 + ], + "opposite_logprob_tokens": [ + -2.578981399536133 + ], + "logprob_reward": -0.07898139953613281, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898139953613281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28967052698135376, + 0.10988769680261612, + 0.044493481516838074, + 0.06738836318254471, + 0.15781202912330627, + 0.05154085531830788, + 0.04325108602643013, + 0.04839576780796051, + 0.10089877992868423, + 0.0866614356637001 + ], + "credit_map": [ + 0.02184119261801243, + 0.026443706825375557, + 0.03731235861778259, + 0.045643363147974014, + 0.029773732647299767, + 0.03574119880795479, + 0.043428290635347366, + 0.0494866669178009, + 0.010521754622459412, + 0.01119404099881649, + 0.011228044517338276, + 0.018817823380231857, + 0.008035860024392605, + 0.013191135600209236, + 0.016759617254137993, + 0.0201394222676754, + 0.0038676790427416563, + 0.004181761294603348, + 0.005397439002990723, + 0.006519549526274204, + 0.004317101556807756, + 0.004443737678229809, + 0.009163913317024708, + 0.006602300796657801, + 0.0053963325917720795, + 0.006056748330593109, + 0.009319817647337914, + 0.009146871976554394, + 0.0066213891841471195, + 0.00706172501668334, + 0.013229490257799625, + 0.010555993765592575, + 0.007149748504161835, + 0.013446099124848843, + 0.01233275793492794, + 0.025633301585912704, + 0.007627896964550018, + 0.026690438389778137, + 0.05208185315132141, + 0.012849925085902214, + 0.00375725026242435, + 0.0045672268606722355, + 0.008123846724629402, + 0.008681515231728554, + 0.005232509225606918, + 0.004775682929903269, + 0.009122397750616074, + 0.007280423771589994, + 0.003746062284335494, + 0.004152844659984112, + 0.004884888883680105, + 0.006754768546670675, + 0.00537128048017621, + 0.004658205900341272, + 0.005854472517967224, + 0.007828562520444393, + 0.0030372675973922014, + 0.005069424398243427, + 0.006432900205254555, + 0.00881851278245449, + 0.0062604318372905254, + 0.003755618818104267, + 0.009428552351891994, + 0.005593061447143555, + 0.005039492156356573, + 0.009480629116296768, + 0.007834856398403645, + 0.01786111481487751, + 0.01729564368724823, + 0.013730878010392189, + 0.01437344029545784, + 0.015282724052667618, + 0.004832567181438208, + 0.003871399210765958, + 0.027299003675580025, + 0.009188896045088768, + 0.004707569722086191, + 0.016287220641970634, + 0.015312112867832184, + 0.0051626646891236305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07898139953613281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.179892435669899 + ], + "opposite_logprob_tokens": [ + -1.8048924207687378 + ], + "logprob_reward": -0.179892435669899, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.179892435669899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11919647455215454, + 0.05679840222001076, + 0.04030315950512886, + 0.01309890951961279, + 0.03882114961743355, + 0.03182686120271683, + 0.021784763783216476, + 0.12288495153188705, + 0.5386177897453308, + 0.01666748896241188 + ], + "credit_map": [ + 0.011262000538408756, + 0.012794191017746925, + 0.012747574597597122, + 0.02163993939757347, + 0.012511594220995903, + 0.013422343879938126, + 0.022152315825223923, + 0.012666520662605762, + 0.004572989419102669, + 0.00491607328876853, + 0.005764174275100231, + 0.010889850556850433, + 0.009198608808219433, + 0.005356417503207922, + 0.009690357372164726, + 0.006409930530935526, + 0.004322119057178497, + 0.003826525527983904, + 0.005792420357465744, + 0.006746590603142977, + 0.00350444158539176, + 0.0032266988418996334, + 0.0070545244961977005, + 0.00582983810454607, + 0.0008702930063009262, + 0.0017133356304839253, + 0.0009921527234837413, + 0.0015427047619596124, + 0.002832806436344981, + 0.0012249327264726162, + 0.0021312995813786983, + 0.0017913845367729664, + 0.001794725190848112, + 0.003487383248284459, + 0.003422854468226433, + 0.006374371238052845, + 0.00338681205175817, + 0.005545478779822588, + 0.01228361576795578, + 0.0025259072426706553, + 0.00380692002363503, + 0.004360526334494352, + 0.003536219010129571, + 0.004688429180532694, + 0.0029427320696413517, + 0.0035529297310858965, + 0.004507442004978657, + 0.004431664012372494, + 0.001603886834345758, + 0.0032054707407951355, + 0.0016187670407816768, + 0.0026244856417179108, + 0.004123343620449305, + 0.002221310045570135, + 0.0030928028281778097, + 0.0032946979627013206, + 0.007863462902605534, + 0.015488880686461926, + 0.014192602597177029, + 0.02343188412487507, + 0.01491209864616394, + 0.011839184910058975, + 0.021076736971735954, + 0.0140801090747118, + 0.0272529236972332, + 0.04653153195977211, + 0.043625470250844955, + 0.0782361701130867, + 0.08289512991905212, + 0.09585651755332947, + 0.07086963951587677, + 0.09335038810968399, + 0.0007482970831915736, + 0.0018043387681245804, + 0.003539163386449218, + 0.00241889082826674, + 0.0013018224854022264, + 0.001161480089649558, + 0.00447811046615243, + 0.0012153862044215202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.179892435669899 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured in a highly detailed 4K vertical format, uses a nearly static, horizontal, handheld perspective to transport the viewer into the vibrant atmosphere of a Japanese tavern. Warm afternoon sunlight streams through a wooden-lattice window on one side of the frame, creating a distinct Tyndall-effect beam in the air, illuminating floating dust particles and an indigo-colored fabric tavern flag suspended from the ceiling, gently fluttering in the air currents. At the center of the frame is a young male musician with slightly tousled dark brown curls. Wearing a washed-blue denim jacket and a white cotton T-shirt, he sits comfortably on a high wooden stool, cradling a natural-wood acoustic guitar. His eyes are slightly closed, a smile of utter immersion on his face. His fingers steadily and deliberately pluck and press the strings, not playing a set piece but engaging in a melodious, spontaneous improvisation. Around him, several friends sat scattered around a rough wooden table, holding glasses of amber-colored fruit wine, tiny droplets of condensation clinging to the sides. Their faces were filled with relaxed smiles, their bodies swaying gently to the rhythm of the music, and they chatted quietly. The entire scene was dominated by warm wood and brick red tones, with the golden highlights of sunlight illuminating the picture, creating an extremely casual, warm, and joyful atmosphere, perfectly capturing a genuine and beautiful afternoon social moment.", + "seed": 3912, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_11/index_1218_seed_3912_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": -0.051184645108878615, + "vlm_logprob": -0.051184645108878615, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.051184645108878615, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04224675893783569 + ], + "opposite_logprob_tokens": [ + -3.4172468185424805 + ], + "logprob_reward": -0.04224675893783569, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04224675893783569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33601298928260803, + 0.1368236243724823, + 0.0926547721028328, + 0.055326737463474274, + 0.07432059198617935, + 0.04621706157922745, + 0.06666142493486404, + 0.06152450665831566, + 0.08057156205177307, + 0.04988674819469452 + ], + "credit_map": [ + 0.03164858743548393, + 0.0408964566886425, + 0.04881741479039192, + 0.04809243232011795, + 0.04740726575255394, + 0.044144853949546814, + 0.03159220144152641, + 0.043413758277893066, + 0.012883736751973629, + 0.020510077476501465, + 0.02280438505113125, + 0.012174206785857677, + 0.021981345489621162, + 0.016110366210341454, + 0.014026602730154991, + 0.016332903876900673, + 0.009166471660137177, + 0.015011980198323727, + 0.014601759612560272, + 0.0077279009856283665, + 0.014049301855266094, + 0.013137176632881165, + 0.008562911301851273, + 0.010397268459200859, + 0.0072003984823822975, + 0.011722901836037636, + 0.006593802478164434, + 0.005234923213720322, + 0.006667163223028183, + 0.007806858979165554, + 0.004965062253177166, + 0.005135627929121256, + 0.008424601517617702, + 0.014238577336072922, + 0.007505933754146099, + 0.006129428278654814, + 0.01036563515663147, + 0.009653784334659576, + 0.0098917530849576, + 0.008110874332487583, + 0.006119769532233477, + 0.006517066620290279, + 0.0043830955401062965, + 0.005663957912474871, + 0.007285499945282936, + 0.00571564631536603, + 0.004583362024277449, + 0.005948660895228386, + 0.007318098563700914, + 0.010179574601352215, + 0.007550555747002363, + 0.007547604385763407, + 0.009054490365087986, + 0.009976079687476158, + 0.006705805193632841, + 0.008329210802912712, + 0.007878033444285393, + 0.010575123131275177, + 0.006223635748028755, + 0.006213838700205088, + 0.009648614563047886, + 0.008453561924397945, + 0.006189353298395872, + 0.006342347711324692, + 0.011302385479211807, + 0.010933637619018555, + 0.009240125305950642, + 0.007988153956830502, + 0.014583186246454716, + 0.010392061434686184, + 0.007915651425719261, + 0.00821636151522398, + 0.0070495991967618465, + 0.006438633427023888, + 0.006118326913565397, + 0.004894237965345383, + 0.00836747046560049, + 0.0065784459002316, + 0.004095022566616535, + 0.006345014087855816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04224675893783569 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090303771197796 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090303771197796, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090303771197796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33960577845573425, + 0.16130001842975616, + 0.12458628416061401, + 0.0457056425511837, + 0.06387870013713837, + 0.049085307866334915, + 0.04655032604932785, + 0.0631314292550087, + 0.06278713792562485, + 0.043369416147470474 + ], + "credit_map": [ + 0.020379643887281418, + 0.024884933605790138, + 0.02633574977517128, + 0.032310374081134796, + 0.06219819188117981, + 0.060656771063804626, + 0.04059334099292755, + 0.07224678248167038, + 0.009772132150828838, + 0.016840346157550812, + 0.010617847554385662, + 0.016260534524917603, + 0.04358287528157234, + 0.020909510552883148, + 0.028172936290502548, + 0.015143846161663532, + 0.008904119953513145, + 0.008589664474129677, + 0.011954805813729763, + 0.016459891572594643, + 0.02702914923429489, + 0.015335232950747013, + 0.025708651170134544, + 0.010604766197502613, + 0.0038150588516145945, + 0.0035659633576869965, + 0.004208064638078213, + 0.004653462208807468, + 0.010470123961567879, + 0.006637176964432001, + 0.0066146343015134335, + 0.005741157103329897, + 0.005152862053364515, + 0.006084676366299391, + 0.00635686656460166, + 0.005998688284307718, + 0.012616610154509544, + 0.00812788587063551, + 0.009795747697353363, + 0.009745362214744091, + 0.003865218022838235, + 0.004526535514742136, + 0.0048099965788424015, + 0.006279736757278442, + 0.008713964372873306, + 0.00581106124445796, + 0.008644666522741318, + 0.006434125825762749, + 0.005090742837637663, + 0.0057115876115858555, + 0.004273124970495701, + 0.004034651443362236, + 0.008305083960294724, + 0.0066955192014575005, + 0.00527874706313014, + 0.007160869427025318, + 0.004457654897123575, + 0.004745709244161844, + 0.0060641029849648476, + 0.006699242629110813, + 0.01335913222283125, + 0.007272553164511919, + 0.01089518703520298, + 0.009637849405407906, + 0.004899537190794945, + 0.004786663223057985, + 0.00537409633398056, + 0.007189157884567976, + 0.015844061970710754, + 0.008939797058701515, + 0.009445423260331154, + 0.006308399606496096, + 0.004552402067929506, + 0.0035231197252869606, + 0.005980309564620256, + 0.0058423434384167194, + 0.007090821862220764, + 0.005156965926289558, + 0.005744927562773228, + 0.005478524602949619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011090303771197796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.33813804388046265, + 0.17390571534633636, + 0.0798843577504158, + 0.06953806430101395, + 0.057941775768995285, + 0.04481636360287666, + 0.03990258648991585, + 0.0541493259370327, + 0.09732562303543091, + 0.04439810663461685 + ], + "credit_map": [ + 0.021729854866862297, + 0.03330806642770767, + 0.022929444909095764, + 0.024695880711078644, + 0.058106716722249985, + 0.05590221285820007, + 0.03707423061132431, + 0.08439165353775024, + 0.010647552087903023, + 0.017661890015006065, + 0.016632448881864548, + 0.02074144408106804, + 0.037710122764110565, + 0.021263694390654564, + 0.03346208110451698, + 0.01578647270798683, + 0.005779821425676346, + 0.007698658853769302, + 0.00968841090798378, + 0.008383222855627537, + 0.01485260110348463, + 0.011045803315937519, + 0.010359952226281166, + 0.01207588892430067, + 0.005717566702514887, + 0.00884088035672903, + 0.00900681409984827, + 0.008602545596659184, + 0.01318309921771288, + 0.008258690126240253, + 0.009499135427176952, + 0.00642932765185833, + 0.006307889707386494, + 0.009239071980118752, + 0.0061159455217421055, + 0.006675108801573515, + 0.006628285627812147, + 0.006726290099322796, + 0.008927701972424984, + 0.00732148252427578, + 0.005129618104547262, + 0.005502830259501934, + 0.00527131836861372, + 0.005507966503500938, + 0.0070250993594527245, + 0.005114834755659103, + 0.005527989007532597, + 0.005736709572374821, + 0.0036026325542479753, + 0.0057053035125136375, + 0.004786756820976734, + 0.005023219157010317, + 0.005789067130535841, + 0.004252525977790356, + 0.00553326727822423, + 0.005209817551076412, + 0.005523722153156996, + 0.0057720099575817585, + 0.0056822095066308975, + 0.004797582048922777, + 0.012701944448053837, + 0.007769680581986904, + 0.006117651239037514, + 0.005784525536000729, + 0.009261766448616982, + 0.011202918365597725, + 0.009009107016026974, + 0.013795385137200356, + 0.021196873858571053, + 0.014024094678461552, + 0.010107884183526039, + 0.008727598004043102, + 0.0064146434888243675, + 0.0055836644023656845, + 0.007249812595546246, + 0.005042273085564375, + 0.005488449241966009, + 0.0049425289034843445, + 0.0044503239914774895, + 0.005226412322372198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "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 scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898569852113724 + ], + "opposite_logprob_tokens": [ + -2.5789856910705566 + ], + "logprob_reward": -0.07898569852113724, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898569852113724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3145548701286316, + 0.1233554407954216, + 0.07345597445964813, + 0.036492712795734406, + 0.056850366294384, + 0.0629892349243164, + 0.070979043841362, + 0.13742303848266602, + 0.08355209231376648, + 0.04034724831581116 + ], + "credit_map": [ + 0.020366590470075607, + 0.0390419065952301, + 0.052365273237228394, + 0.05329331383109093, + 0.03386275842785835, + 0.04118505120277405, + 0.02401934377849102, + 0.050420645624399185, + 0.011390232481062412, + 0.02439357526600361, + 0.016606701537966728, + 0.016269609332084656, + 0.014855138026177883, + 0.015122161246836185, + 0.010857767425477505, + 0.013860260136425495, + 0.0068012624979019165, + 0.011548129841685295, + 0.008482784032821655, + 0.007735001854598522, + 0.010142939165234566, + 0.011450862511992455, + 0.006372974719852209, + 0.010922028683125973, + 0.00298473984003067, + 0.004589923657476902, + 0.0036890325136482716, + 0.004959451500326395, + 0.004486872814595699, + 0.00520638283342123, + 0.0049176462925970554, + 0.005658664274960756, + 0.004141880664974451, + 0.0073969378136098385, + 0.006310384254902601, + 0.00607421575114131, + 0.010522156022489071, + 0.006900576874613762, + 0.008101201616227627, + 0.00740301888436079, + 0.00798396859318018, + 0.009463069029152393, + 0.00588497007265687, + 0.006434394977986813, + 0.010553695261478424, + 0.007776518817991018, + 0.005624899175018072, + 0.009267718531191349, + 0.007695595733821392, + 0.010466487146914005, + 0.011455381289124489, + 0.008533954620361328, + 0.007775989361107349, + 0.009109484031796455, + 0.006943698972463608, + 0.008998462930321693, + 0.017768943682312965, + 0.021146709099411964, + 0.014973817393183708, + 0.0270749069750309, + 0.016593899577856064, + 0.01814194768667221, + 0.010754500515758991, + 0.010968336835503578, + 0.008285959251224995, + 0.009404187090694904, + 0.008599577471613884, + 0.012478160671889782, + 0.017097707837820053, + 0.01124279573559761, + 0.00858607143163681, + 0.00785764865577221, + 0.00557996379211545, + 0.005633542779833078, + 0.0058656008914113045, + 0.006248126272112131, + 0.0040811835788190365, + 0.005023625213652849, + 0.003689657663926482, + 0.004225550219416618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.07898569852113724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030481964349747 + ], + "opposite_logprob_tokens": [ + -2.3503048419952393 + ], + "logprob_reward": -0.10030481964349747, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030481964349747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24738429486751556, + 0.10423575341701508, + 0.06741972267627716, + 0.05551105737686157, + 0.20352961122989655, + 0.06052231043577194, + 0.05186295881867409, + 0.05861639231443405, + 0.09474129974842072, + 0.05617661401629448 + ], + "credit_map": [ + 0.02507868967950344, + 0.03334709629416466, + 0.02322554402053356, + 0.02826755866408348, + 0.03243858739733696, + 0.033365294337272644, + 0.026325810700654984, + 0.04533572122454643, + 0.01262380089610815, + 0.02260919101536274, + 0.01207717601209879, + 0.008556471206247807, + 0.01420872937887907, + 0.011928399093449116, + 0.009715227410197258, + 0.012516757473349571, + 0.008086246438324451, + 0.011220518499612808, + 0.011738777160644531, + 0.00592286279425025, + 0.006562310736626387, + 0.008500619791448116, + 0.007205505855381489, + 0.008182883262634277, + 0.006989665795117617, + 0.008820165880024433, + 0.008007388561964035, + 0.006336526945233345, + 0.00771902222186327, + 0.00767390476539731, + 0.0050081112422049046, + 0.004956272896379232, + 0.02364976704120636, + 0.03214944899082184, + 0.025812799111008644, + 0.021340016275644302, + 0.03184264898300171, + 0.026132024824619293, + 0.022213252261281013, + 0.020389635115861893, + 0.007837827317416668, + 0.008859243243932724, + 0.007199411280453205, + 0.00707184337079525, + 0.010456670075654984, + 0.006936610676348209, + 0.006265703588724136, + 0.005895000416785479, + 0.005498978309333324, + 0.007986307144165039, + 0.006995172705501318, + 0.006715604104101658, + 0.007757812738418579, + 0.006225319113582373, + 0.0050787669606506824, + 0.005604995414614677, + 0.006175598595291376, + 0.007619919255375862, + 0.005469899158924818, + 0.0056324428878724575, + 0.012877513654530048, + 0.007113033439964056, + 0.0071294065564870834, + 0.006598578300327063, + 0.007525931112468243, + 0.012200416065752506, + 0.009427795186638832, + 0.012935001403093338, + 0.016841186210513115, + 0.01173508819192648, + 0.013544113375246525, + 0.010531765408813953, + 0.006802218500524759, + 0.006077389232814312, + 0.007791256997734308, + 0.005753030069172382, + 0.011256104335188866, + 0.0068773808889091015, + 0.005212173331528902, + 0.006407056003808975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.10030481964349747 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured in a highly detailed 4K vertical format, uses a nearly static, horizontal, handheld perspective to transport the viewer into the vibrant atmosphere of a Japanese tavern. Warm afternoon sunlight streams through a wooden-lattice window on one side of the frame, creating a distinct Tyndall-effect beam in the air, illuminating floating dust particles and an indigo-colored fabric tavern flag suspended from the ceiling, gently fluttering in the air currents. At the center of the frame is a young male musician with slightly tousled dark brown curls. Wearing a washed-blue denim jacket and a white cotton T-shirt, he sits comfortably on a high wooden stool, cradling a natural-wood acoustic guitar. His eyes are slightly closed, a smile of utter immersion on his face. His fingers steadily and deliberately pluck and press the strings, not playing a set piece but engaging in a melodious, spontaneous improvisation. Around him, several friends sat scattered around a rough wooden table, holding glasses of amber-colored fruit wine, tiny droplets of condensation clinging to the sides. Their faces were filled with relaxed smiles, their bodies swaying gently to the rhythm of the music, and they chatted quietly. The entire scene was dominated by warm wood and brick red tones, with the golden highlights of sunlight illuminating the picture, creating an extremely casual, warm, and joyful atmosphere, perfectly capturing a genuine and beautiful afternoon social moment.", + "seed": 3913, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_11/index_1218_seed_3913_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": -0.052451856713742015, + "vlm_logprob": -0.052451856713742015, + "vlm_yes_no_margin": 3.374999976158142, + "vlm_token_credit": -0.052451856713742015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13045819103717804 + ], + "opposite_logprob_tokens": [ + -2.130458116531372 + ], + "logprob_reward": -0.13045819103717804, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13045819103717804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.341235488653183, + 0.17726868391036987, + 0.0955401211977005, + 0.08187837898731232, + 0.01794920675456524, + 0.01500014215707779, + 0.11405064165592194, + 0.03232390061020851, + 0.015032250434160233, + 0.10972119122743607 + ], + "credit_map": [ + 0.03480326756834984, + 0.03459186851978302, + 0.04875386878848076, + 0.04771380499005318, + 0.04985497519373894, + 0.04875147342681885, + 0.03674055635929108, + 0.04002567008137703, + 0.008580966852605343, + 0.011413248255848885, + 0.020943745970726013, + 0.027690008282661438, + 0.02545282058417797, + 0.02567467838525772, + 0.04053463041782379, + 0.016978593543171883, + 0.011759103275835514, + 0.01901128701865673, + 0.011711095459759235, + 0.009851126931607723, + 0.00978196319192648, + 0.012873782776296139, + 0.010056652128696442, + 0.010495111346244812, + 0.007011460606008768, + 0.005574310664087534, + 0.007823968306183815, + 0.013334816321730614, + 0.012069597840309143, + 0.011247529648244381, + 0.014236765913665295, + 0.010579931549727917, + 0.0015925405314192176, + 0.0013459139736369252, + 0.0025779895950108767, + 0.0024530356749892235, + 0.0026503289118409157, + 0.0022297799587249756, + 0.002590929390862584, + 0.0025086896494030952, + 0.0010761705925688148, + 0.0013762791641056538, + 0.0029155013617128134, + 0.0019444598583504558, + 0.002474647480994463, + 0.0019852344412356615, + 0.0018061932642012835, + 0.0014216570416465402, + 0.009842564351856709, + 0.01514781266450882, + 0.014807919971644878, + 0.01460111141204834, + 0.01774223893880844, + 0.015313643962144852, + 0.01411664392799139, + 0.012478699907660484, + 0.0036363229155540466, + 0.0044752913527190685, + 0.0032133718486875296, + 0.003979665692895651, + 0.005421577487140894, + 0.004018895793706179, + 0.0033322079107165337, + 0.004246566444635391, + 0.0018146176589652896, + 0.0013727828627452254, + 0.0018629271071404219, + 0.0022291780915111303, + 0.0019752837251871824, + 0.002163247438147664, + 0.0018905632896348834, + 0.0017236500279977918, + 0.005505888257175684, + 0.008987928740680218, + 0.010409078560769558, + 0.016915418207645416, + 0.020314332097768784, + 0.016557103022933006, + 0.013607604429125786, + 0.0174238458275795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.13045819103717804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634864654392004 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634864654392004, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634864654392004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19896796345710754, + 0.4128839373588562, + 0.04983020946383476, + 0.05764848738908768, + 0.08420199155807495, + 0.022107718512415886, + 0.048013415187597275, + 0.03007073700428009, + 0.03373377397656441, + 0.06254180520772934 + ], + "credit_map": [ + 0.014321508817374706, + 0.01807670295238495, + 0.034514620900154114, + 0.027198228985071182, + 0.028841910883784294, + 0.024410614743828773, + 0.022902976721525192, + 0.028701387345790863, + 0.027213016524910927, + 0.02814885601401329, + 0.04723003879189491, + 0.05092918872833252, + 0.05944051593542099, + 0.05553432181477547, + 0.09867744892835617, + 0.04571051895618439, + 0.0034320540726184845, + 0.005862089805305004, + 0.00592874875292182, + 0.005938700865954161, + 0.008410525508224964, + 0.007123695686459541, + 0.006220811512321234, + 0.006913585588335991, + 0.0041670165956020355, + 0.0038367859087884426, + 0.005678150802850723, + 0.0072375512681901455, + 0.009413096122443676, + 0.007755024824291468, + 0.0115365544334054, + 0.008024308830499649, + 0.005432114936411381, + 0.010697416961193085, + 0.008317571133375168, + 0.0067936633713543415, + 0.022654524073004723, + 0.01270701177418232, + 0.01005610078573227, + 0.007543587125837803, + 0.0014170093927532434, + 0.002216554479673505, + 0.002745740581303835, + 0.003720402717590332, + 0.003556224750354886, + 0.0028490026015788317, + 0.0031623158138245344, + 0.002440469106659293, + 0.004527151584625244, + 0.00546687375754118, + 0.005179307889193296, + 0.005258933175355196, + 0.009584327228367329, + 0.0072859120555222034, + 0.004468618426471949, + 0.006242293864488602, + 0.0038081281818449497, + 0.0038494006730616093, + 0.004723507910966873, + 0.003472907468676567, + 0.0042346566915512085, + 0.003930709324777126, + 0.003363619325682521, + 0.00268780579790473, + 0.002774140564724803, + 0.005337725859135389, + 0.00426706625148654, + 0.0031231380999088287, + 0.006489674095064402, + 0.004500893875956535, + 0.003957710228860378, + 0.0032834229059517384, + 0.003756400663405657, + 0.003954761661589146, + 0.007784987334161997, + 0.011646397411823273, + 0.010447674430906773, + 0.008727217093110085, + 0.008791163563728333, + 0.007433201652020216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.007634864654392004 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.14960461854934692, + 0.1585363894701004, + 0.05336736515164375, + 0.08776328712701797, + 0.050527460873126984, + 0.08949001878499985, + 0.1068568155169487, + 0.10945148020982742, + 0.11708032339811325, + 0.07732222974300385 + ], + "credit_map": [ + 0.01349590439349413, + 0.024326687678694725, + 0.02450532652437687, + 0.01844794861972332, + 0.024472367018461227, + 0.019863873720169067, + 0.012333840131759644, + 0.012158660218119621, + 0.009165015071630478, + 0.016652824357151985, + 0.017740868031978607, + 0.02342481166124344, + 0.020489435642957687, + 0.023027554154396057, + 0.034355927258729935, + 0.013679958879947662, + 0.005030371714383364, + 0.007564738858491182, + 0.00698902178555727, + 0.0072580622509121895, + 0.00605476088821888, + 0.007355748675763607, + 0.006723897065967321, + 0.006390762981027365, + 0.00682184798642993, + 0.008108085952699184, + 0.0067376340739429, + 0.016120800748467445, + 0.016564929857850075, + 0.012130016461014748, + 0.014831719920039177, + 0.006448254454880953, + 0.005035115871578455, + 0.00899355299770832, + 0.005984197370707989, + 0.006512800231575966, + 0.008717004209756851, + 0.005850891582667828, + 0.004870112519711256, + 0.004563783295452595, + 0.0054548438638448715, + 0.0073935710825026035, + 0.011494839563965797, + 0.011418800801038742, + 0.01252599898725748, + 0.013263026252388954, + 0.010504994541406631, + 0.017433946952223778, + 0.007459086365997791, + 0.010135562159121037, + 0.017267704010009766, + 0.01405197661370039, + 0.01751413382589817, + 0.015962257981300354, + 0.015424367040395737, + 0.00904172845184803, + 0.01349398959428072, + 0.017927471548318863, + 0.012083184905350208, + 0.01117624994367361, + 0.01943773217499256, + 0.012855619192123413, + 0.009921695105731487, + 0.012555527500808239, + 0.009472384117543697, + 0.017791002988815308, + 0.011899847537279129, + 0.01563076861202717, + 0.017405929043889046, + 0.015556222759187222, + 0.013479995541274548, + 0.015844177454710007, + 0.006747247185558081, + 0.007906856015324593, + 0.009450510144233704, + 0.008148123510181904, + 0.017271453514695168, + 0.011705701239407063, + 0.007533946540206671, + 0.008558399975299835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.016069892793893814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000947743654251 + ], + "opposite_logprob_tokens": [ + -2.695009469985962 + ], + "logprob_reward": -0.07000947743654251, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000947743654251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21799428761005402, + 0.15867258608341217, + 0.07049784064292908, + 0.0702565386891365, + 0.03008248470723629, + 0.07486051321029663, + 0.11048958450555801, + 0.04872245714068413, + 0.01778886653482914, + 0.2006348967552185 + ], + "credit_map": [ + 0.017556432634592056, + 0.02021866664290428, + 0.02362009696662426, + 0.03520777449011803, + 0.03415358439087868, + 0.030152613297104836, + 0.02968248724937439, + 0.027402618899941444, + 0.012178851291537285, + 0.027111666277050972, + 0.014820181764662266, + 0.020149841904640198, + 0.026431884616613388, + 0.020049719139933586, + 0.02369757741689682, + 0.01423285435885191, + 0.009524324908852577, + 0.009473057463765144, + 0.010344298556447029, + 0.00626370171085, + 0.006793544162064791, + 0.009140321984887123, + 0.0073502869345247746, + 0.011608302593231201, + 0.005620043259114027, + 0.0059670209884643555, + 0.005694136954843998, + 0.010395647957921028, + 0.013913163915276527, + 0.009878531098365784, + 0.011397484689950943, + 0.007390513084828854, + 0.0015264350222423673, + 0.002365704160183668, + 0.0043090637773275375, + 0.004553251434117556, + 0.0036101944278925657, + 0.003955802880227566, + 0.004378186073154211, + 0.005383848212659359, + 0.0033307948615401983, + 0.006517615169286728, + 0.010590444318950176, + 0.01558842696249485, + 0.009846613742411137, + 0.009955896995961666, + 0.01096426136791706, + 0.008066453970968723, + 0.01020368654280901, + 0.01676853373646736, + 0.01728670671582222, + 0.01107439398765564, + 0.01996232382953167, + 0.01418965682387352, + 0.012979584746062756, + 0.008024693466722965, + 0.0053866165690124035, + 0.006738762836903334, + 0.005446425639092922, + 0.005527629051357508, + 0.00978309940546751, + 0.006475456524640322, + 0.004869934171438217, + 0.004494532011449337, + 0.0011746601667255163, + 0.0020890990272164345, + 0.002266754861921072, + 0.002042638836428523, + 0.003631546860560775, + 0.0023375083692371845, + 0.0022686668671667576, + 0.001977991545572877, + 0.009623772464692593, + 0.017311029136180878, + 0.022913822904229164, + 0.033779945224523544, + 0.029798533767461777, + 0.029305679723620415, + 0.027420727536082268, + 0.03048136457800865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07000947743654251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a Japanese tavern with warm wood and brick red tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young male musician with dark brown curls wearing a washed-blue denim jacket and white T-shirt visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the musician playing a natural-wood acoustic guitar with a spontaneous improvisation while smiling with eyes slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are friends sitting around a rough wooden table drinking amber-colored fruit wine with condensation droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight streaming through a wooden-lattice window creating a Tyndall-effect beam illuminating an indigo tavern flag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808685764670372 + ], + "opposite_logprob_tokens": [ + -3.2880868911743164 + ], + "logprob_reward": -0.03808685764670372, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808685764670372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26129627227783203, + 0.1301996260881424, + 0.3329482972621918, + 0.06882407516241074, + 0.021259546279907227, + 0.022631296887993813, + 0.045644018799066544, + 0.01942329667508602, + 0.013221759349107742, + 0.08455175906419754 + ], + "credit_map": [ + 0.017397437244653702, + 0.030129561200737953, + 0.02529396116733551, + 0.048915762454271317, + 0.03776194155216217, + 0.03722329065203667, + 0.035459328442811966, + 0.02911500819027424, + 0.01322291512042284, + 0.010183052159845829, + 0.012984156608581543, + 0.019354570657014847, + 0.01905699260532856, + 0.01672920025885105, + 0.02814890258014202, + 0.010519859381020069, + 0.06800806522369385, + 0.03730415552854538, + 0.05360988900065422, + 0.023264093324542046, + 0.04821513965725899, + 0.047592643648386, + 0.021655892953276634, + 0.03329843655228615, + 0.008142373524606228, + 0.004557366948574781, + 0.006397130433470011, + 0.010103084146976471, + 0.009552168659865856, + 0.009135731495916843, + 0.011459498666226864, + 0.00947672687470913, + 0.0017008959548547864, + 0.0024514098186045885, + 0.001647492521442473, + 0.0018814537907019258, + 0.005656501278281212, + 0.0032025075051933527, + 0.0021401066333055496, + 0.002579181920737028, + 0.002077460056170821, + 0.0033913033548742533, + 0.0031732432544231415, + 0.002192653249949217, + 0.0037447039503604174, + 0.0029606821481138468, + 0.0022819580044597387, + 0.0028092938009649515, + 0.0038858645129948854, + 0.004920472856611013, + 0.007101647090166807, + 0.005084612872451544, + 0.008038995787501335, + 0.006673484109342098, + 0.005575936287641525, + 0.00436300877481699, + 0.0016078762710094452, + 0.0016232982743531466, + 0.002023584209382534, + 0.002629156457260251, + 0.0029666577465832233, + 0.0025240518152713776, + 0.0024639861658215523, + 0.003584687365218997, + 0.0009543423657305539, + 0.0008896924555301666, + 0.0015242008958011866, + 0.0015139132738113403, + 0.0017892754403874278, + 0.001884567434899509, + 0.001828320906497538, + 0.0028374476823955774, + 0.008068460039794445, + 0.013000646606087685, + 0.010538776405155659, + 0.007605789229273796, + 0.014092005789279938, + 0.011229285039007664, + 0.007666966877877712, + 0.012349831871688366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.03808685764670372 + } + ], + "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/0000052/group_12/index_tensor([6908, 6908])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_12/index_tensor([6908, 6908])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3771be8f3bfb61843d8e2fa0f987735bf551507a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_12/index_tensor([6908, 6908])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sharp close-up shot, captured at a high frame rate, begins with a news anchor, sporting a sleek, dark bob and a serious, focused expression, reporting directly into the camera with a slim, futuristic microphone. The camera then steadily and deliberately pulls back to reveal the vast expanse of the scene she inhabits: the interior of a vast, multi-story, futuristic vertical farm. As the camera pulls back, the central subject gradually emerges: a sleek, self-driving transporter, quietly parked on a polished concrete floor, its matte white composite exterior standing out against the surroundings. A thick, translucent power transmission cable connects from the side of the vehicle to a charging port on the wall. A long, blue LED current indicator on the vehicle pulses smoothly, emitting a steady, rhythmic, cool glow. Its smooth, unobstructed motion underscores the clarity of the high frame rate capture. In the background, neat rows of hydroponic racks extend upwards. The lush green plants on the racks, illuminated by purple and pink plant growth lights, add a touch of life to the entire clean and efficient space, creating a futuristic atmosphere of high technology, cleanliness and tranquility.", + "seed": 18666, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_12/index_6908_seed_18666_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": -0.010666211508214474, + "vlm_logprob": -0.010666211508214474, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.010666211508214474, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016167255118489265 + ], + "opposite_logprob_tokens": [ + -4.266167163848877 + ], + "logprob_reward": -0.016167255118489265, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016167255118489265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09323327243328094, + 0.04560407996177673, + 0.11970306932926178, + 0.25808343291282654, + 0.20196533203125, + 0.05353773385286331, + 0.0884961485862732, + 0.10319071263074875, + 0.024035152047872543, + 0.012151063419878483 + ], + "credit_map": [ + 0.013797112740576267, + 0.007410210557281971, + 0.008861103095114231, + 0.015806199982762337, + 0.011696912348270416, + 0.019158178940415382, + 0.006663993000984192, + 0.009839559905230999, + 0.006060468032956123, + 0.004184626508504152, + 0.0039643109776079655, + 0.009296589531004429, + 0.00492363516241312, + 0.00810872483998537, + 0.0030071225482970476, + 0.006058600731194019, + 0.013628087937831879, + 0.012064400129020214, + 0.010177437216043472, + 0.02801959589123726, + 0.010678389109671116, + 0.01814226433634758, + 0.009268243797123432, + 0.0177246555685997, + 0.0328085757791996, + 0.02463517151772976, + 0.05274491012096405, + 0.050210870802402496, + 0.022743744775652885, + 0.018703730776906013, + 0.027386313304305077, + 0.028850123286247253, + 0.045428916811943054, + 0.01325527485460043, + 0.06901940703392029, + 0.01289452612400055, + 0.010003169998526573, + 0.012521211057901382, + 0.011546660214662552, + 0.027296148240566254, + 0.005995305720716715, + 0.005480470135807991, + 0.007737050298601389, + 0.010693500749766827, + 0.005728936288505793, + 0.004492747597396374, + 0.00804258044809103, + 0.005367144010961056, + 0.011195424012839794, + 0.008178414776921272, + 0.010779653675854206, + 0.011927272193133831, + 0.00722862733528018, + 0.010314851999282837, + 0.017216848209500313, + 0.011655058711767197, + 0.00711470190435648, + 0.011263156309723854, + 0.015456406399607658, + 0.01939469203352928, + 0.011634023860096931, + 0.007290070876479149, + 0.020131753757596016, + 0.010905908420681953, + 0.00293234852142632, + 0.0021033224184066057, + 0.003535372670739889, + 0.003203804139047861, + 0.0036606937646865845, + 0.0025992891751229763, + 0.003202620195224881, + 0.0027977002318948507, + 0.001526776934042573, + 0.0011251108953729272, + 0.0028949002735316753, + 0.001668318873271346, + 0.0013742545852437615, + 0.0009531723335385323, + 0.0013545119436457753, + 0.0012540178140625358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.016167255118489265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264348816126585 + ], + "opposite_logprob_tokens": [ + -5.2552642822265625 + ], + "logprob_reward": -0.005264348816126585, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264348816126585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0682847648859024, + 0.016300424933433533, + 0.03470173478126526, + 0.11961660534143448, + 0.4210006892681122, + 0.09426037222146988, + 0.10700777918100357, + 0.08362007141113281, + 0.040625493973493576, + 0.014582074247300625 + ], + "credit_map": [ + 0.008424436673521996, + 0.004933096934109926, + 0.003834592644125223, + 0.00903779361397028, + 0.006859134417027235, + 0.019482847303152084, + 0.003920776769518852, + 0.011792083270847797, + 0.002168305218219757, + 0.001405495684593916, + 0.0014956058003008366, + 0.0030629609245806932, + 0.0018721641972661018, + 0.0024610538966953754, + 0.0016247446183115244, + 0.002210094826295972, + 0.003931126557290554, + 0.0033902726136147976, + 0.0024264520034193993, + 0.008562805131077766, + 0.003207304049283266, + 0.005131899379193783, + 0.0024648504331707954, + 0.005587024614214897, + 0.014653529040515423, + 0.010653686709702015, + 0.023401280865073204, + 0.023796960711479187, + 0.012553077191114426, + 0.011867842637002468, + 0.009412208572030067, + 0.013278025202453136, + 0.10402864962816238, + 0.021495677530765533, + 0.15537939965724945, + 0.02394210547208786, + 0.020694280043244362, + 0.020741501823067665, + 0.020804600790143013, + 0.053914472460746765, + 0.008357702754437923, + 0.010791347362101078, + 0.010649135336279869, + 0.019778503105044365, + 0.008694899268448353, + 0.00795778539031744, + 0.019181478768587112, + 0.008849523961544037, + 0.00890860054641962, + 0.008412802591919899, + 0.010822739452123642, + 0.017425602301955223, + 0.010197245515882969, + 0.014956888742744923, + 0.018903503194451332, + 0.01738038845360279, + 0.005234981887042522, + 0.009377686306834221, + 0.011731628328561783, + 0.018687615171074867, + 0.008283481001853943, + 0.005429703742265701, + 0.016441717743873596, + 0.00843325536698103, + 0.005046986974775791, + 0.0035329933743923903, + 0.0076729003340005875, + 0.0037593133747577667, + 0.007211390882730484, + 0.00409776670858264, + 0.00472624134272337, + 0.00457790307700634, + 0.002005045535042882, + 0.0013135699555277824, + 0.0021338777150958776, + 0.0021179800387471914, + 0.001601566793397069, + 0.001218525692820549, + 0.002366489265114069, + 0.0018250191351398826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.005264348816126585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036222815979272127 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036222815979272127, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036222815979272127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06254570931196213, + 0.020263098180294037, + 0.01651216670870781, + 0.09124919027090073, + 0.08093143254518509, + 0.13841266930103302, + 0.14477331936359406, + 0.30351918935775757, + 0.03076828643679619, + 0.11102495342493057 + ], + "credit_map": [ + 0.009670594707131386, + 0.005324161611497402, + 0.004995712544769049, + 0.0101251769810915, + 0.006986652500927448, + 0.013557609170675278, + 0.004331036005169153, + 0.007554757874459028, + 0.0033007224556058645, + 0.0014748640824109316, + 0.0020204989705234766, + 0.0032705196645110846, + 0.0018628392135724425, + 0.003789284033700824, + 0.0016512287547811866, + 0.002893142169341445, + 0.002013401361182332, + 0.0009981532348319888, + 0.001231617177836597, + 0.0025669552851468325, + 0.002192810410633683, + 0.004213155247271061, + 0.0009469604701735079, + 0.0023491147439926863, + 0.010944968089461327, + 0.007746560499072075, + 0.01947472058236599, + 0.018783513456583023, + 0.00871292408555746, + 0.009512975811958313, + 0.006050904281437397, + 0.01002262532711029, + 0.011475068517029285, + 0.009466024115681648, + 0.012497354298830032, + 0.010118267498910427, + 0.009080993011593819, + 0.007192865014076233, + 0.013156074099242687, + 0.007944787852466106, + 0.00935427751392126, + 0.011567478068172932, + 0.015040832571685314, + 0.03199734166264534, + 0.013463967479765415, + 0.010925408452749252, + 0.028442583978176117, + 0.01762077957391739, + 0.013699479401111603, + 0.010644585825502872, + 0.01818886213004589, + 0.02560020051896572, + 0.0135203767567873, + 0.012636509723961353, + 0.03349420428276062, + 0.016989095136523247, + 0.025140689685940742, + 0.024805059656500816, + 0.043728191405534744, + 0.05774436146020889, + 0.023245586082339287, + 0.020724015310406685, + 0.07567476481199265, + 0.032456520944833755, + 0.002325693378224969, + 0.0037393649108707905, + 0.004218076355755329, + 0.004926865920424461, + 0.003581495024263859, + 0.0032533772755414248, + 0.004529606085270643, + 0.004193807486444712, + 0.009879282675683498, + 0.009382844902575016, + 0.013074914924800396, + 0.019775113090872765, + 0.00905876699835062, + 0.010601688176393509, + 0.020565234124660492, + 0.018687104806303978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.0036222815979272127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633725479245186 + ], + "opposite_logprob_tokens": [ + -3.651337146759033 + ], + "logprob_reward": -0.02633725479245186, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633725479245186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05295614153146744, + 0.022337818518280983, + 0.02120543271303177, + 0.10950101166963577, + 0.09062864631414413, + 0.10735750198364258, + 0.1826661080121994, + 0.20829489827156067, + 0.057394564151763916, + 0.1476578712463379 + ], + "credit_map": [ + 0.006900643464177847, + 0.003557749791070819, + 0.004046253394335508, + 0.006926659494638443, + 0.006462403107434511, + 0.015028056688606739, + 0.002898630453273654, + 0.0071357483975589275, + 0.0036601272877305746, + 0.0016164043918251991, + 0.0023538011591881514, + 0.0034060373436659575, + 0.0025061434134840965, + 0.004120911005884409, + 0.0017228213837370276, + 0.0029515738133341074, + 0.0030681181233376265, + 0.0017087047453969717, + 0.002195946639403701, + 0.003548040986061096, + 0.00256349821574986, + 0.004606589674949646, + 0.0014738006284460425, + 0.0020407340489327908, + 0.012239810079336166, + 0.00923541933298111, + 0.020906031131744385, + 0.02227615751326084, + 0.012420854531228542, + 0.012391392141580582, + 0.008464023470878601, + 0.011567329987883568, + 0.012846047058701515, + 0.011417431756854057, + 0.013364064507186413, + 0.010537884198129177, + 0.010918769054114819, + 0.010211475193500519, + 0.0143086863681674, + 0.007024282123893499, + 0.010318689048290253, + 0.010463390499353409, + 0.011204779148101807, + 0.024810515344142914, + 0.009160931222140789, + 0.01277256105095148, + 0.018957195803523064, + 0.009669434279203415, + 0.01811828278005123, + 0.014998110942542553, + 0.019045472145080566, + 0.033961471170186996, + 0.014894033782184124, + 0.01748574711382389, + 0.040991269052028656, + 0.023171715438365936, + 0.015342436730861664, + 0.026022061705589294, + 0.02729431539773941, + 0.043934863060712814, + 0.02647593431174755, + 0.014731773175299168, + 0.031822800636291504, + 0.02267073467373848, + 0.005093968473374844, + 0.00580698112025857, + 0.007065097335726023, + 0.00882555078715086, + 0.0076534440740942955, + 0.006945121102035046, + 0.006963229272514582, + 0.009041177108883858, + 0.012223929166793823, + 0.02048524096608162, + 0.026248818263411522, + 0.01712155155837536, + 0.02005133591592312, + 0.014419320039451122, + 0.015645837411284447, + 0.021461835131049156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.02633725479245186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001939917216077447 + ], + "opposite_logprob_tokens": [ + -6.25193977355957 + ], + "logprob_reward": -0.001939917216077447, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001939917216077447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07881519198417664, + 0.02855663001537323, + 0.03046487271785736, + 0.056073568761348724, + 0.29519057273864746, + 0.09640692174434662, + 0.17792977392673492, + 0.08540987968444824, + 0.11298558115959167, + 0.03816701099276543 + ], + "credit_map": [ + 0.011983665637671947, + 0.008914388716220856, + 0.005340405739843845, + 0.013069399632513523, + 0.007047342136502266, + 0.016247818246483803, + 0.006034355144947767, + 0.010177820920944214, + 0.0038493575993925333, + 0.002152340952306986, + 0.002312812954187393, + 0.0043082470074296, + 0.0027493212837725878, + 0.007260107900947332, + 0.001796803786419332, + 0.004127638880163431, + 0.005278968717902899, + 0.002070904476568103, + 0.0025280576664954424, + 0.003200290957465768, + 0.0037264146376401186, + 0.008429219014942646, + 0.0021780114620923996, + 0.0030530064832419157, + 0.004775324836373329, + 0.00505554536357522, + 0.008838235400617123, + 0.01105346530675888, + 0.00775691494345665, + 0.007507509086281061, + 0.005380089860409498, + 0.005706481635570526, + 0.07337378710508347, + 0.018203983083367348, + 0.10632891207933426, + 0.01892496459186077, + 0.013548010028898716, + 0.013808845542371273, + 0.017451275140047073, + 0.033550817519426346, + 0.008326108567416668, + 0.013611100614070892, + 0.00984897930175066, + 0.01853967271745205, + 0.010137282311916351, + 0.00942747388035059, + 0.014069711789488792, + 0.012446600012481213, + 0.02011040784418583, + 0.013280218467116356, + 0.024671532213687897, + 0.02946045994758606, + 0.013550019823014736, + 0.025768477469682693, + 0.0301214512437582, + 0.02096720226109028, + 0.008106665685772896, + 0.006840749643743038, + 0.009871874935925007, + 0.016702139750123024, + 0.008709153160452843, + 0.006655497010797262, + 0.01998380944132805, + 0.008539985865354538, + 0.010628878138959408, + 0.017751144245266914, + 0.01154877245426178, + 0.011864655651152134, + 0.021660804748535156, + 0.011210588738322258, + 0.012088300660252571, + 0.016232434660196304, + 0.005138212814927101, + 0.005425089038908482, + 0.005839062388986349, + 0.004021131433546543, + 0.005381494294852018, + 0.003561269259080291, + 0.004174103960394859, + 0.004626648500561714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.001939917216077447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp close-up shot, captured at a high frame rate, begins with a news anchor, sporting a sleek, dark bob and a serious, focused expression, reporting directly into the camera with a slim, futuristic microphone. The camera then steadily and deliberately pulls back to reveal the vast expanse of the scene she inhabits: the interior of a vast, multi-story, futuristic vertical farm. As the camera pulls back, the central subject gradually emerges: a sleek, self-driving transporter, quietly parked on a polished concrete floor, its matte white composite exterior standing out against the surroundings. A thick, translucent power transmission cable connects from the side of the vehicle to a charging port on the wall. A long, blue LED current indicator on the vehicle pulses smoothly, emitting a steady, rhythmic, cool glow. Its smooth, unobstructed motion underscores the clarity of the high frame rate capture. In the background, neat rows of hydroponic racks extend upwards. The lush green plants on the racks, illuminated by purple and pink plant growth lights, add a touch of life to the entire clean and efficient space, creating a futuristic atmosphere of high technology, cleanliness and tranquility.", + "seed": 18667, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_12/index_6908_seed_18667_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": -0.040278492495417595, + "vlm_logprob": -0.040278492495417595, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.040278492495417595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06340526044368744 + ], + "opposite_logprob_tokens": [ + -2.8134052753448486 + ], + "logprob_reward": -0.06340526044368744, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06340526044368744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24198612570762634, + 0.11290669441223145, + 0.36394843459129333, + 0.01963121071457863, + 0.08641987293958664, + 0.0583798810839653, + 0.06376583874225616, + 0.013357278890907764, + 0.024798527359962463, + 0.014806137420237064 + ], + "credit_map": [ + 0.03615398705005646, + 0.02968059480190277, + 0.03673496097326279, + 0.021144157275557518, + 0.028416002169251442, + 0.03662680834531784, + 0.032436028122901917, + 0.02079358696937561, + 0.01657017506659031, + 0.01953398995101452, + 0.016126064583659172, + 0.01398411300033331, + 0.011800224892795086, + 0.013263987377285957, + 0.010903842747211456, + 0.010724298655986786, + 0.04814962297677994, + 0.03685110807418823, + 0.0813969150185585, + 0.05733886733651161, + 0.02307445928454399, + 0.05085843801498413, + 0.02641872875392437, + 0.03986029326915741, + 0.0026689365040510893, + 0.002060515806078911, + 0.003692186903208494, + 0.002124025719240308, + 0.0022434683050960302, + 0.0028437578584998846, + 0.0016078000189736485, + 0.002390519483014941, + 0.006398078519850969, + 0.009226654656231403, + 0.007151233963668346, + 0.007371989078819752, + 0.018073294311761856, + 0.010827713645994663, + 0.019174283370375633, + 0.008196617476642132, + 0.005047048907727003, + 0.007769912946969271, + 0.007838533259928226, + 0.004806044977158308, + 0.006670593749731779, + 0.008733544498682022, + 0.0067214337177574635, + 0.010792766697704792, + 0.005030258558690548, + 0.0065368120558559895, + 0.01332029327750206, + 0.006123692728579044, + 0.005780296865850687, + 0.008360989391803741, + 0.00876765139400959, + 0.009845840744674206, + 0.0010225350270047784, + 0.002626530360430479, + 0.002339631086215377, + 0.0012565486831590533, + 0.0013935425085946918, + 0.0008227286161854863, + 0.0019088726257905364, + 0.001986891031265259, + 0.002197569701820612, + 0.002601443789899349, + 0.004885404370725155, + 0.0024684760719537735, + 0.0032324541825801134, + 0.002506626769900322, + 0.0026494760531932116, + 0.004257076885551214, + 0.0011505240108817816, + 0.002054969547316432, + 0.00331398518756032, + 0.0014112297212705016, + 0.0015367013402283192, + 0.0009804811561480165, + 0.0013700785348191857, + 0.002988168504089117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.06340526044368744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899175584316254 + ], + "opposite_logprob_tokens": [ + -2.578991651535034 + ], + "logprob_reward": -0.07899175584316254, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899175584316254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19602146744728088, + 0.11532584577798843, + 0.06101132184267044, + 0.042644768953323364, + 0.1161498948931694, + 0.15801915526390076, + 0.0919373631477356, + 0.07929785549640656, + 0.06670209765434265, + 0.0728902667760849 + ], + "credit_map": [ + 0.03433048352599144, + 0.026591641828417778, + 0.0339740589261055, + 0.015973858535289764, + 0.022933389991521835, + 0.026669006794691086, + 0.02106267400085926, + 0.014486336149275303, + 0.019649645313620567, + 0.009882435202598572, + 0.01307561807334423, + 0.010265287011861801, + 0.007802202831953764, + 0.027848491445183754, + 0.00946196261793375, + 0.01734018884599209, + 0.007333533372730017, + 0.006907320115715265, + 0.008665714412927628, + 0.007104974705725908, + 0.004727932158857584, + 0.007636280730366707, + 0.004665616899728775, + 0.013969942927360535, + 0.008625062182545662, + 0.003990943543612957, + 0.005936429835855961, + 0.003945048898458481, + 0.0052813258953392506, + 0.0071019334718585014, + 0.0026617380790412426, + 0.0051022800616919994, + 0.011144398711621761, + 0.012095670215785503, + 0.00576985627412796, + 0.009595808573067188, + 0.030328182503581047, + 0.013487882912158966, + 0.02173803001642227, + 0.011990058235824108, + 0.013468717224895954, + 0.022107718512415886, + 0.01792890392243862, + 0.013272857293486595, + 0.01806970313191414, + 0.02561519667506218, + 0.018097009509801865, + 0.029459036886692047, + 0.006606533192098141, + 0.009169216267764568, + 0.019418856129050255, + 0.009968682192265987, + 0.00776318134739995, + 0.012357247993350029, + 0.012882755137979984, + 0.013770883902907372, + 0.007194104138761759, + 0.007417883258312941, + 0.014060339890420437, + 0.006774144247174263, + 0.0062051923014223576, + 0.00934743694961071, + 0.013025897555053234, + 0.01527284923940897, + 0.005241702776402235, + 0.009346064180135727, + 0.00983999203890562, + 0.00489075668156147, + 0.010128939524292946, + 0.005970125552266836, + 0.009061435237526894, + 0.0122230751439929, + 0.0040182690136134624, + 0.008296427316963673, + 0.015468550845980644, + 0.009574727155268192, + 0.010077190585434437, + 0.004136156290769577, + 0.008097643032670021, + 0.013221295550465584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.07899175584316254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036200247704982758 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036200247704982758, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036200247704982758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19701600074768066, + 0.0587979257106781, + 0.06621839106082916, + 0.03309842199087143, + 0.14407338201999664, + 0.11254506558179855, + 0.11952120065689087, + 0.12165400385856628, + 0.06831604242324829, + 0.0787595808506012 + ], + "credit_map": [ + 0.03095719963312149, + 0.028670182451605797, + 0.03115936927497387, + 0.015820935368537903, + 0.023885192349553108, + 0.028167134150862694, + 0.022862834855914116, + 0.015493150800466537, + 0.009220425970852375, + 0.007498142309486866, + 0.010140803642570972, + 0.005936520639806986, + 0.004109579604119062, + 0.00941150076687336, + 0.005405303090810776, + 0.007075653411448002, + 0.007641526870429516, + 0.00857559498399496, + 0.01018507033586502, + 0.009876010939478874, + 0.0053891888819634914, + 0.008038755506277084, + 0.004839168395847082, + 0.011673077009618282, + 0.004103578627109528, + 0.0050453185103833675, + 0.005669829901307821, + 0.005070030689239502, + 0.0032205895986407995, + 0.004218606743961573, + 0.0028470212128013372, + 0.002923447173088789, + 0.010210065171122551, + 0.015190777368843555, + 0.010935215279459953, + 0.012380505912005901, + 0.03389839455485344, + 0.018787667155265808, + 0.02762913890182972, + 0.015041629783809185, + 0.010425117798149586, + 0.012851632200181484, + 0.015785474330186844, + 0.01096909400075674, + 0.014501010999083519, + 0.014852317050099373, + 0.014461017213761806, + 0.01869940012693405, + 0.008867152966558933, + 0.016770510002970695, + 0.01725984923541546, + 0.0160526093095541, + 0.009095500223338604, + 0.007301645819097757, + 0.015176254324615002, + 0.028997676447033882, + 0.009386606514453888, + 0.02947683446109295, + 0.013680403120815754, + 0.01356458105146885, + 0.012284478172659874, + 0.007759676314890385, + 0.01904834434390068, + 0.016453083604574203, + 0.004783577751368284, + 0.0076111480593681335, + 0.010268265381455421, + 0.0058639622293412685, + 0.01009862869977951, + 0.008834125474095345, + 0.009105931967496872, + 0.01175040565431118, + 0.004640484228730202, + 0.01224810816347599, + 0.017098428681492805, + 0.00658285291865468, + 0.009260526858270168, + 0.003963441587984562, + 0.006332640070468187, + 0.018633093684911728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0036200247704982758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864171892404556 + ], + "opposite_logprob_tokens": [ + -3.0486416816711426 + ], + "logprob_reward": -0.04864171892404556, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864171892404556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07113160938024521, + 0.06840420514345169, + 0.04980742186307907, + 0.05576471984386444, + 0.2584981918334961, + 0.09096592664718628, + 0.12054294347763062, + 0.07372669875621796, + 0.03413041681051254, + 0.17702792584896088 + ], + "credit_map": [ + 0.010631861165165901, + 0.008855867199599743, + 0.011304754763841629, + 0.0069147381000220776, + 0.007644362282007933, + 0.00907411240041256, + 0.007203218527138233, + 0.00950268842279911, + 0.010841366834938526, + 0.008721777237951756, + 0.010165985673666, + 0.00731142470613122, + 0.003502396633848548, + 0.013122977688908577, + 0.005410351790487766, + 0.009327918291091919, + 0.009789417497813702, + 0.003882127581164241, + 0.014646180905401707, + 0.004443828947842121, + 0.0027139901649206877, + 0.008957893587648869, + 0.0024441792629659176, + 0.00292979646474123, + 0.007031124550849199, + 0.005742337089031935, + 0.008799758739769459, + 0.005716032348573208, + 0.005736528895795345, + 0.007921676151454449, + 0.0037565550301223993, + 0.011060708202421665, + 0.021287059411406517, + 0.023155028000473976, + 0.023610932752490044, + 0.033061884343624115, + 0.03292866423726082, + 0.028126917779445648, + 0.056800492107868195, + 0.03952718898653984, + 0.008892961777746677, + 0.008636543527245522, + 0.014880271628499031, + 0.008230888284742832, + 0.011599389836192131, + 0.011033371090888977, + 0.015467234887182713, + 0.01222525630146265, + 0.007845153100788593, + 0.017097478732466698, + 0.01960141211748123, + 0.015079734846949577, + 0.009144654497504234, + 0.011138972826302052, + 0.014191347174346447, + 0.02644418552517891, + 0.004834587685763836, + 0.016938770189881325, + 0.011486799456179142, + 0.006450013257563114, + 0.0054851556196808815, + 0.003488485235720873, + 0.010969823226332664, + 0.014073052443563938, + 0.002230572747066617, + 0.004127194173634052, + 0.0051920050755143166, + 0.0030000770930200815, + 0.0061781019903719425, + 0.002383921528235078, + 0.004571663681417704, + 0.006446877960115671, + 0.01010311208665371, + 0.027058195322752, + 0.037400152534246445, + 0.01810142770409584, + 0.02031708136200905, + 0.009488421492278576, + 0.014764629304409027, + 0.03979488089680672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.04864171892404556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067337024956941605 + ], + "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": [ + 0.29274746775627136, + 0.08085823059082031, + 0.03855295851826668, + 0.04433516785502434, + 0.1008150577545166, + 0.15165922045707703, + 0.12532801926136017, + 0.0650559514760971, + 0.05439731478691101, + 0.046250589191913605 + ], + "credit_map": [ + 0.04795274883508682, + 0.037848100066185, + 0.04637344926595688, + 0.024974435567855835, + 0.03525207191705704, + 0.03868532180786133, + 0.04036217927932739, + 0.02129920944571495, + 0.012578776106238365, + 0.008984879590570927, + 0.008007915690541267, + 0.009609363973140717, + 0.007156158331781626, + 0.01529997680336237, + 0.008173640817403793, + 0.011047531850636005, + 0.004710468929260969, + 0.0047441208735108376, + 0.009252100251615047, + 0.006294759921729565, + 0.0034171862062066793, + 0.0045617856085300446, + 0.0029141895938664675, + 0.0026583487633615732, + 0.007632873952388763, + 0.005132008343935013, + 0.006550328806042671, + 0.004747474566102028, + 0.003881521988660097, + 0.006751393899321556, + 0.003441466949880123, + 0.0061981030739843845, + 0.010133549571037292, + 0.009597632102668285, + 0.006305711809545755, + 0.013697439804673195, + 0.012239651754498482, + 0.01687101647257805, + 0.020257167518138885, + 0.011712899431586266, + 0.012094411067664623, + 0.020055407658219337, + 0.01958867721259594, + 0.017378585413098335, + 0.016281437128782272, + 0.021495604887604713, + 0.021581260487437248, + 0.023183850571513176, + 0.009249519556760788, + 0.013478904031217098, + 0.023209812119603157, + 0.01470908708870411, + 0.011458811350166798, + 0.012662672437727451, + 0.01876327209174633, + 0.021795962005853653, + 0.005227989051491022, + 0.014848915860056877, + 0.006933765485882759, + 0.006643247324973345, + 0.006263143382966518, + 0.007538296282291412, + 0.008528821170330048, + 0.009071781300008297, + 0.004209961276501417, + 0.007977724075317383, + 0.00752607686445117, + 0.004096809774637222, + 0.008259731344878674, + 0.007966089062392712, + 0.006324558984488249, + 0.008036372251808643, + 0.0030097244307398796, + 0.005294123198837042, + 0.009622681885957718, + 0.005506705492734909, + 0.007013910915702581, + 0.003788542468100786, + 0.004560932517051697, + 0.007453973405063152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0067337024956941605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp close-up shot, captured at a high frame rate, begins with a news anchor, sporting a sleek, dark bob and a serious, focused expression, reporting directly into the camera with a slim, futuristic microphone. The camera then steadily and deliberately pulls back to reveal the vast expanse of the scene she inhabits: the interior of a vast, multi-story, futuristic vertical farm. As the camera pulls back, the central subject gradually emerges: a sleek, self-driving transporter, quietly parked on a polished concrete floor, its matte white composite exterior standing out against the surroundings. A thick, translucent power transmission cable connects from the side of the vehicle to a charging port on the wall. A long, blue LED current indicator on the vehicle pulses smoothly, emitting a steady, rhythmic, cool glow. Its smooth, unobstructed motion underscores the clarity of the high frame rate capture. In the background, neat rows of hydroponic racks extend upwards. The lush green plants on the racks, illuminated by purple and pink plant growth lights, add a touch of life to the entire clean and efficient space, creating a futuristic atmosphere of high technology, cleanliness and tranquility.", + "seed": 18668, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_12/index_6908_seed_18668_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": -0.012369537213817238, + "vlm_logprob": -0.012369537213817238, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.012369537213817238, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02434816025197506 + ], + "opposite_logprob_tokens": [ + -3.774348258972168 + ], + "logprob_reward": -0.02434816025197506, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02434816025197506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22841371595859528, + 0.11039605736732483, + 0.0764675959944725, + 0.05856133997440338, + 0.11327546834945679, + 0.054050009697675705, + 0.09450975060462952, + 0.12281923741102219, + 0.06840237230062485, + 0.07310453802347183 + ], + "credit_map": [ + 0.04552440345287323, + 0.02763292007148266, + 0.03782237693667412, + 0.01645079255104065, + 0.02892930619418621, + 0.028414350003004074, + 0.026296161115169525, + 0.017343387007713318, + 0.01062605157494545, + 0.017917340621352196, + 0.022024137899279594, + 0.011126141995191574, + 0.013441286981105804, + 0.00946821365505457, + 0.012766601517796516, + 0.013026286847889423, + 0.014240478165447712, + 0.011545438319444656, + 0.010875607840716839, + 0.012713306583464146, + 0.005831319373100996, + 0.00926230475306511, + 0.0061670816503465176, + 0.005832062102854252, + 0.007219641003757715, + 0.005844085477292538, + 0.010991689749062061, + 0.0077313813380897045, + 0.005645603407174349, + 0.007078204769641161, + 0.007891444489359856, + 0.006159284617751837, + 0.014116708189249039, + 0.015325423330068588, + 0.015059790574014187, + 0.015477467328310013, + 0.009166266769170761, + 0.013268842361867428, + 0.01018547359853983, + 0.02067549154162407, + 0.00811181589961052, + 0.008131328038871288, + 0.007121920119971037, + 0.00470868730917573, + 0.006575707346200943, + 0.005767256952822208, + 0.005501443520188332, + 0.008131850510835648, + 0.015278064645826817, + 0.013557774014770985, + 0.01221623457968235, + 0.008935082703828812, + 0.013253416866064072, + 0.010162102989852428, + 0.006490293890237808, + 0.014616780914366245, + 0.014187703840434551, + 0.021238502115011215, + 0.02016059122979641, + 0.008745395578444004, + 0.022831417620182037, + 0.011955385096371174, + 0.006958036683499813, + 0.016742203384637833, + 0.008052095770835876, + 0.009145919233560562, + 0.010604113340377808, + 0.008343874476850033, + 0.007429817691445351, + 0.01096249744296074, + 0.005699397064745426, + 0.008164660073816776, + 0.014720861800014973, + 0.011425625532865524, + 0.010468117892742157, + 0.004883196670562029, + 0.007197984494268894, + 0.009766100905835629, + 0.00547400489449501, + 0.009168645367026329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.02434816025197506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078457236289978 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.016078457236289978, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078457236289978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2551093101501465, + 0.03074389509856701, + 0.054782792925834656, + 0.059126656502485275, + 0.11236579716205597, + 0.044531382620334625, + 0.16566389799118042, + 0.08281448483467102, + 0.1099313423037529, + 0.08493048697710037 + ], + "credit_map": [ + 0.052405040711164474, + 0.022232716903090477, + 0.06202233210206032, + 0.019270330667495728, + 0.024765631183981895, + 0.02860400080680847, + 0.021174509078264236, + 0.024634728208184242, + 0.0027248898986727, + 0.003508727764710784, + 0.004044470377266407, + 0.004262578207999468, + 0.006005433388054371, + 0.0028375221882015467, + 0.004313822835683823, + 0.003046447644010186, + 0.013786958530545235, + 0.006663283798843622, + 0.0049936408177018166, + 0.006009490229189396, + 0.005647487007081509, + 0.00885037425905466, + 0.0037176511250436306, + 0.005113903433084488, + 0.005656760185956955, + 0.004812129773199558, + 0.01072462648153305, + 0.009893236681818962, + 0.005695817992091179, + 0.006825071759521961, + 0.008677562698721886, + 0.006841449066996574, + 0.01493071485310793, + 0.015374342910945415, + 0.014094889163970947, + 0.01185489259660244, + 0.009279336780309677, + 0.01461624726653099, + 0.010335730388760567, + 0.021879633888602257, + 0.006912289187312126, + 0.0065273018553853035, + 0.00586346909403801, + 0.0036383981350809336, + 0.005295955576002598, + 0.0051590269431471825, + 0.003401246853172779, + 0.007733688689768314, + 0.013592200353741646, + 0.018725326284766197, + 0.02666291780769825, + 0.017559049651026726, + 0.035486284643411636, + 0.01659511961042881, + 0.010770228691399097, + 0.026272768154740334, + 0.008933106437325478, + 0.013517452403903008, + 0.014552582986652851, + 0.006961458828300238, + 0.014933062717318535, + 0.008902951143682003, + 0.004488267470151186, + 0.010525599122047424, + 0.018321817740797997, + 0.025633012875914574, + 0.01651585102081299, + 0.009013812988996506, + 0.008700058795511723, + 0.012662860564887524, + 0.005415577907115221, + 0.013668338768184185, + 0.016886111348867416, + 0.015741338953375816, + 0.014034402556717396, + 0.007344463840126991, + 0.009852588176727295, + 0.008245447650551796, + 0.0054168375208973885, + 0.0074092913419008255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.016078457236289978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036201435141265392 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036201435141265392, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036201435141265392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07765261828899384, + 0.05148695036768913, + 0.03627832233905792, + 0.07126489281654358, + 0.16401953995227814, + 0.05045327916741371, + 0.22740331292152405, + 0.12104968726634979, + 0.1156066358089447, + 0.08478474617004395 + ], + "credit_map": [ + 0.01472771167755127, + 0.005856324452906847, + 0.01577286794781685, + 0.007566035259515047, + 0.008863469585776329, + 0.008106260560452938, + 0.010055123828351498, + 0.0067048221826553345, + 0.01198022160679102, + 0.004776927176862955, + 0.007688006851822138, + 0.0035488686989992857, + 0.0042768376879394054, + 0.01108525414019823, + 0.003628338687121868, + 0.004502493422478437, + 0.005294328089803457, + 0.0040250979363918304, + 0.003214399330317974, + 0.007511395029723644, + 0.0030880256090313196, + 0.004802507348358631, + 0.004235442262142897, + 0.0041071274317801, + 0.01000966876745224, + 0.006902281194925308, + 0.012479779310524464, + 0.009114746004343033, + 0.007267143577337265, + 0.007796349935233593, + 0.010405692271888256, + 0.007289224769920111, + 0.019522622227668762, + 0.0164823979139328, + 0.02714829333126545, + 0.016300005838274956, + 0.010864151641726494, + 0.017554834485054016, + 0.020492039620876312, + 0.035655200481414795, + 0.006697291508316994, + 0.007194523699581623, + 0.008421541191637516, + 0.005602081771939993, + 0.004858705215156078, + 0.004971541464328766, + 0.006308631040155888, + 0.006398963276296854, + 0.01629100926220417, + 0.025112686678767204, + 0.03463584557175636, + 0.030285527929663658, + 0.04484708979725838, + 0.018005114048719406, + 0.017870431765913963, + 0.04035560414195061, + 0.011679328046739101, + 0.011523199267685413, + 0.01740334928035736, + 0.013062423095107079, + 0.01740294136106968, + 0.0120572280138731, + 0.008354712277650833, + 0.029566505923867226, + 0.016505874693393707, + 0.019118526950478554, + 0.01703588105738163, + 0.013310860842466354, + 0.009984227828681469, + 0.014544258825480938, + 0.009010789915919304, + 0.016096217557787895, + 0.012865717522799969, + 0.010461857542395592, + 0.012613997794687748, + 0.010104222223162651, + 0.010362118482589722, + 0.00702578853815794, + 0.01017063669860363, + 0.011180403642356396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.0036201435141265392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186720363795757 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014186720363795757, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186720363795757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03994317352771759, + 0.03516482189297676, + 0.023887185379862785, + 0.12606197595596313, + 0.24903933703899384, + 0.08774244040250778, + 0.1819339543581009, + 0.07224662601947784, + 0.10431478917598724, + 0.0796656385064125 + ], + "credit_map": [ + 0.007460950408130884, + 0.0033591927494853735, + 0.005049719009548426, + 0.005143004935234785, + 0.004917331505566835, + 0.004727592226117849, + 0.0048677329905331135, + 0.0044176517985761166, + 0.003302937839180231, + 0.006270496640354395, + 0.005994010716676712, + 0.004013984929770231, + 0.004786635749042034, + 0.0043428088538348675, + 0.0028588303830474615, + 0.0035951139871031046, + 0.003800845704972744, + 0.002272137673571706, + 0.0030309350695461035, + 0.003413496073335409, + 0.002199274254962802, + 0.003988932818174362, + 0.0026183524169027805, + 0.002563211601227522, + 0.0203850194811821, + 0.012823903933167458, + 0.01855693943798542, + 0.01572369411587715, + 0.011923150159418583, + 0.013258029706776142, + 0.02007400617003441, + 0.013317241333425045, + 0.023834960535168648, + 0.023309459909796715, + 0.029838761314749718, + 0.03464973717927933, + 0.017312433570623398, + 0.023551736027002335, + 0.04915890470147133, + 0.047383323311805725, + 0.0091514578089118, + 0.008727732114493847, + 0.011344830505549908, + 0.01289565209299326, + 0.010552112013101578, + 0.008128990419209003, + 0.01029903069138527, + 0.016642635688185692, + 0.023144828155636787, + 0.02632622793316841, + 0.02597649209201336, + 0.012590420432388783, + 0.03181389346718788, + 0.018340853974223137, + 0.009322364814579487, + 0.034418877214193344, + 0.005860579665750265, + 0.006442342884838581, + 0.010931920260190964, + 0.008408320136368275, + 0.010959470644593239, + 0.006961342412978411, + 0.004896246362477541, + 0.01778639853000641, + 0.01575363613665104, + 0.017492391169071198, + 0.014013250358402729, + 0.010002141818404198, + 0.009508879855275154, + 0.015062355436384678, + 0.006433898117393255, + 0.016048235818743706, + 0.012274923734366894, + 0.011499278247356415, + 0.016845135018229485, + 0.0067602889612317085, + 0.00947655364871025, + 0.00702606700360775, + 0.006696665659546852, + 0.0090867280960083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.014186720363795757 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "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": [ + 0.2033848613500595, + 0.10374050587415695, + 0.060986995697021484, + 0.10213208943605423, + 0.06771327555179596, + 0.0420004278421402, + 0.11777399480342865, + 0.08307667821645737, + 0.12096551805734634, + 0.09822563081979752 + ], + "credit_map": [ + 0.04100242257118225, + 0.021191194653511047, + 0.034704484045505524, + 0.016415933147072792, + 0.025168390944600105, + 0.024778088554739952, + 0.02179827354848385, + 0.01832607202231884, + 0.02920439839363098, + 0.00839895848184824, + 0.012560331262648106, + 0.005214089062064886, + 0.008099746890366077, + 0.025836339220404625, + 0.004862494766712189, + 0.009564142674207687, + 0.009449795819818974, + 0.007460029795765877, + 0.007134275510907173, + 0.006606319919228554, + 0.006346981041133404, + 0.012681113556027412, + 0.005916397552937269, + 0.0053920806385576725, + 0.010317065753042698, + 0.010390659794211388, + 0.013375223614275455, + 0.009827072732150555, + 0.012925282120704651, + 0.02050573006272316, + 0.012982235290110111, + 0.011808820068836212, + 0.008188953623175621, + 0.00873352587223053, + 0.009428421966731548, + 0.006877491250634193, + 0.00561364134773612, + 0.008387782610952854, + 0.007395862601697445, + 0.013087601400911808, + 0.006742581259459257, + 0.006542902439832687, + 0.006013012956827879, + 0.0034349821507930756, + 0.0047090984880924225, + 0.004147466737776995, + 0.003990245051681995, + 0.006420136895030737, + 0.016896812245249748, + 0.017111225053668022, + 0.015877099707722664, + 0.007738480810075998, + 0.022245436906814575, + 0.012538192793726921, + 0.00570438290014863, + 0.019662365317344666, + 0.009704820811748505, + 0.015811072662472725, + 0.012694891542196274, + 0.0062467362731695175, + 0.014354582875967026, + 0.008111879229545593, + 0.0045348783023655415, + 0.011617821641266346, + 0.02109365165233612, + 0.018040694296360016, + 0.016269944608211517, + 0.010889766737818718, + 0.010511836037039757, + 0.01883230358362198, + 0.007244241423904896, + 0.018083078786730766, + 0.018909672275185585, + 0.026274114847183228, + 0.017277270555496216, + 0.007344154641032219, + 0.009989800862967968, + 0.007979770191013813, + 0.0040183900855481625, + 0.006432444788515568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.00361420470289886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp close-up shot, captured at a high frame rate, begins with a news anchor, sporting a sleek, dark bob and a serious, focused expression, reporting directly into the camera with a slim, futuristic microphone. The camera then steadily and deliberately pulls back to reveal the vast expanse of the scene she inhabits: the interior of a vast, multi-story, futuristic vertical farm. As the camera pulls back, the central subject gradually emerges: a sleek, self-driving transporter, quietly parked on a polished concrete floor, its matte white composite exterior standing out against the surroundings. A thick, translucent power transmission cable connects from the side of the vehicle to a charging port on the wall. A long, blue LED current indicator on the vehicle pulses smoothly, emitting a steady, rhythmic, cool glow. Its smooth, unobstructed motion underscores the clarity of the high frame rate capture. In the background, neat rows of hydroponic racks extend upwards. The lush green plants on the racks, illuminated by purple and pink plant growth lights, add a touch of life to the entire clean and efficient space, creating a futuristic atmosphere of high technology, cleanliness and tranquility.", + "seed": 18669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_12/index_6908_seed_18669_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": -0.02559486902318895, + "vlm_logprob": -0.02559486902318895, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.02559486902318895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08210693299770355 + ], + "opposite_logprob_tokens": [ + -2.582106828689575 + ], + "logprob_reward": -0.08210693299770355, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08210693299770355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2040552794933319, + 0.050086501985788345, + 0.02244844287633896, + 0.04131704196333885, + 0.05941406637430191, + 0.329439640045166, + 0.054895274341106415, + 0.029914192855358124, + 0.01604677177965641, + 0.19238276779651642 + ], + "credit_map": [ + 0.014025932177901268, + 0.02976667694747448, + 0.017462894320487976, + 0.048907097429037094, + 0.012030498124659061, + 0.03036103956401348, + 0.012250936590135098, + 0.0392502062022686, + 0.004786619450896978, + 0.009516483172774315, + 0.003981853369623423, + 0.015299683436751366, + 0.0028958963230252266, + 0.003257719101384282, + 0.003763774177059531, + 0.006584470625966787, + 0.0030891995411366224, + 0.002787391422316432, + 0.003978648688644171, + 0.003014259971678257, + 0.0026517363730818033, + 0.0027250770945101976, + 0.00206222222186625, + 0.002139908028766513, + 0.0055061038583517075, + 0.0015949769876897335, + 0.008734296075999737, + 0.00772408302873373, + 0.002003794303163886, + 0.005583138205111027, + 0.005655928049236536, + 0.004514716565608978, + 0.0063181184232234955, + 0.006316664163023233, + 0.009156230837106705, + 0.006862998008728027, + 0.0059170848689973354, + 0.0072527555748820305, + 0.01031037513166666, + 0.007279842160642147, + 0.020240159705281258, + 0.09280485659837723, + 0.015595843084156513, + 0.011354836635291576, + 0.02137642726302147, + 0.02084212191402912, + 0.04354914277791977, + 0.10367624461650848, + 0.0035289383959025145, + 0.005209187511354685, + 0.009264678694307804, + 0.009677467867732048, + 0.003148412797600031, + 0.004138296935707331, + 0.012190116569399834, + 0.007738174870610237, + 0.0034619071520864964, + 0.0029969271272420883, + 0.005456904415041208, + 0.004013186786323786, + 0.0023450267035514116, + 0.0026178706903010607, + 0.0034161731600761414, + 0.005606197286397219, + 0.0009194684098474681, + 0.0012174798175692558, + 0.002348625333979726, + 0.0027528994251042604, + 0.0014569849008694291, + 0.0012946255737915635, + 0.0027905311435461044, + 0.0032661589793860912, + 0.008032109588384628, + 0.010851720348000526, + 0.059193581342697144, + 0.0663210079073906, + 0.00644526956602931, + 0.008472525514662266, + 0.013193825259804726, + 0.019872745499014854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.08210693299770355 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583942532539368 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583942532539368, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583942532539368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041666507720947266, + 0.023375025019049644, + 0.01670972630381584, + 0.0708000659942627, + 0.03693556785583496, + 0.45074522495269775, + 0.04871084541082382, + 0.02886366844177246, + 0.02265845611691475, + 0.25953492522239685 + ], + "credit_map": [ + 0.0023464092519134283, + 0.005939309485256672, + 0.0025559677742421627, + 0.010456444695591927, + 0.0015830787597224116, + 0.006836605723947287, + 0.002334958640858531, + 0.009613736532628536, + 0.001529756118543446, + 0.0021126531064510345, + 0.0016559428768232465, + 0.0027148330118507147, + 0.004827503580600023, + 0.0046643889509141445, + 0.0017384698148816824, + 0.004131479188799858, + 0.0013903443468734622, + 0.0010493542067706585, + 0.002359790960326791, + 0.001500707701779902, + 0.0018724494148045778, + 0.0041627767495810986, + 0.0020002489909529686, + 0.002374054864048958, + 0.007279484998434782, + 0.003958574030548334, + 0.011344288475811481, + 0.012951744720339775, + 0.004244439769536257, + 0.01182154007256031, + 0.011370128020644188, + 0.007829871028661728, + 0.004458117764443159, + 0.004385787528008223, + 0.005145904142409563, + 0.003455713391304016, + 0.0039202310144901276, + 0.004574489314109087, + 0.006472044158726931, + 0.004523284267634153, + 0.027924930676817894, + 0.13172468543052673, + 0.018538588657975197, + 0.01411464437842369, + 0.031809546053409576, + 0.031136585399508476, + 0.055906202644109726, + 0.1395900547504425, + 0.004080676473677158, + 0.0027706148102879524, + 0.008824197575449944, + 0.0057185194455087185, + 0.0022674493957310915, + 0.006988423876464367, + 0.00849173590540886, + 0.009569233283400536, + 0.0025013054255396128, + 0.0025185607373714447, + 0.005760112311691046, + 0.004225703421980143, + 0.0018054309766739607, + 0.002191442996263504, + 0.00396393658593297, + 0.005897177383303642, + 0.0016235917573794723, + 0.0015356015646830201, + 0.003701274050399661, + 0.0036344591062515974, + 0.0019194649066776037, + 0.0017500108806416392, + 0.002934780204668641, + 0.005559274926781654, + 0.010970493778586388, + 0.014802373945713043, + 0.08024166524410248, + 0.09269458055496216, + 0.009098262526094913, + 0.01093797106295824, + 0.016120046377182007, + 0.02466954104602337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.020583942532539368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024927754420787096 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.0024927754420787096, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024927754420787096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09104961156845093, + 0.05239219591021538, + 0.02625187113881111, + 0.04163401201367378, + 0.04996458441019058, + 0.15784668922424316, + 0.05724388360977173, + 0.07885312288999557, + 0.06556876003742218, + 0.3791951835155487 + ], + "credit_map": [ + 0.006798942573368549, + 0.015107349492609501, + 0.010633490979671478, + 0.02466820739209652, + 0.007442318368703127, + 0.007898235693573952, + 0.006910926196724176, + 0.011590137146413326, + 0.0032077813521027565, + 0.00960239116102457, + 0.0047362735494971275, + 0.01783059909939766, + 0.0044546485878527164, + 0.004095481242984533, + 0.003333592787384987, + 0.005131430923938751, + 0.0024336394853889942, + 0.0019274954684078693, + 0.0038878254126757383, + 0.002699352102354169, + 0.003515233751386404, + 0.005953219719231129, + 0.0024428367614746094, + 0.0033922677394002676, + 0.0036031140480190516, + 0.003343819873407483, + 0.004626618232578039, + 0.004542986396700144, + 0.004063740372657776, + 0.008013417944312096, + 0.007665963377803564, + 0.005774353630840778, + 0.004479323048144579, + 0.006265395320951939, + 0.006861577276140451, + 0.004505470395088196, + 0.005726037081331015, + 0.008375915698707104, + 0.007435672916471958, + 0.006315186619758606, + 0.011573185212910175, + 0.03709246590733528, + 0.014735902659595013, + 0.007340631447732449, + 0.009875781834125519, + 0.00925357360392809, + 0.02235589362680912, + 0.045619260519742966, + 0.003826606785878539, + 0.0039866953156888485, + 0.011819214560091496, + 0.009761824272572994, + 0.0033223445061594248, + 0.004005798604339361, + 0.011604821309447289, + 0.008916576392948627, + 0.005479954648762941, + 0.005520980805158615, + 0.016835520043969154, + 0.013743149116635323, + 0.0041845086961984634, + 0.007137646898627281, + 0.009054020047187805, + 0.01689733751118183, + 0.0039277998730540276, + 0.005899473093450069, + 0.010512154549360275, + 0.01466273982077837, + 0.008078784681856632, + 0.004592330195009708, + 0.006242068950086832, + 0.011653412133455276, + 0.01602683961391449, + 0.024250147864222527, + 0.1064564660191536, + 0.12285862863063812, + 0.015614934265613556, + 0.020035238936543465, + 0.038489848375320435, + 0.03546306863427162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0024927754420787096 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016056636348366737 + ], + "opposite_logprob_tokens": [ + -4.141056537628174 + ], + "logprob_reward": -0.016056636348366737, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016056636348366737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021047132089734077, + 0.011110836640000343, + 0.008660335093736649, + 0.02575274184346199, + 0.020805368199944496, + 0.06149857118725777, + 0.12285014241933823, + 0.024092333391308784, + 0.015725400298833847, + 0.688457190990448 + ], + "credit_map": [ + 0.00243981066159904, + 0.0035954806953668594, + 0.002524446463212371, + 0.0057306974194943905, + 0.0013331050286069512, + 0.0017736406298354268, + 0.0014073844067752361, + 0.002242564456537366, + 0.0011441032402217388, + 0.0010895642917603254, + 0.001062589231878519, + 0.001836359384469688, + 0.0016480851918458939, + 0.0019897492602467537, + 0.0007475646561942995, + 0.0015928195789456367, + 0.0007824019412510097, + 0.0006596571183763444, + 0.000994649832136929, + 0.0007583311526104808, + 0.0012934787664562464, + 0.0020685777999460697, + 0.0010632005287334323, + 0.001040037488564849, + 0.0036121420562267303, + 0.001316184876486659, + 0.005799316801130772, + 0.005211406387388706, + 0.0018141998443752527, + 0.0023171622306108475, + 0.0031441617757081985, + 0.002538165310397744, + 0.00184591393917799, + 0.0017450809245929122, + 0.003101109527051449, + 0.0026624654419720173, + 0.0016001883195713162, + 0.002876649145036936, + 0.003858777694404125, + 0.0031151818111538887, + 0.004227797035127878, + 0.016212480142712593, + 0.00329919857904315, + 0.0025167374406009912, + 0.0040447586216032505, + 0.004064279142767191, + 0.008743655867874622, + 0.018389664590358734, + 0.0073713636957108974, + 0.008765444159507751, + 0.023651819676160812, + 0.022049838677048683, + 0.00818625744432211, + 0.0068208202719688416, + 0.028879091143608093, + 0.017125502228736877, + 0.0016039848560467362, + 0.001737674931064248, + 0.00443741912022233, + 0.003660900052636862, + 0.0014624661998823285, + 0.0013146381825208664, + 0.0044132317416369915, + 0.005462015513330698, + 0.0012449956266209483, + 0.001365543226711452, + 0.0022572132293134928, + 0.0031195632182061672, + 0.001099969376809895, + 0.0013865747023373842, + 0.0026014677714556456, + 0.0026500720996409655, + 0.030459679663181305, + 0.04431351274251938, + 0.20057788491249084, + 0.23783588409423828, + 0.027065347880125046, + 0.03295647352933884, + 0.05043694004416466, + 0.06481145322322845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.016056636348366737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a sharp close-up shot that begins with a news anchor sporting a dark bob haircut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily pull back to reveal a vast, multi-story futuristic vertical farm interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, matte white self-driving transporter visible parked on a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick translucent power cable connected to the vehicle with a pulsing blue LED indicator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are neat rows of hydroponic racks with green plants illuminated by purple and pink lights visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18695829808712006, + 0.05998198688030243, + 0.07081122696399689, + 0.09081581979990005, + 0.051754921674728394, + 0.06800113618373871, + 0.09628876298666, + 0.0439230278134346, + 0.03671026602387428, + 0.2947545647621155 + ], + "credit_map": [ + 0.0130826560780406, + 0.02254515327513218, + 0.020806221291422844, + 0.03520248085260391, + 0.014856597408652306, + 0.03489429131150246, + 0.012265382334589958, + 0.03330550715327263, + 0.004999420139938593, + 0.006155984476208687, + 0.0062555274926126, + 0.010018797591328621, + 0.008412313647568226, + 0.010477874428033829, + 0.0034782416187226772, + 0.010183827951550484, + 0.00469600223004818, + 0.004279887303709984, + 0.006930329371243715, + 0.004692438989877701, + 0.008531736209988594, + 0.02407708950340748, + 0.007719829678535461, + 0.009883918799459934, + 0.006974261254072189, + 0.005610935855656862, + 0.011298646219074726, + 0.013554750010371208, + 0.0074379052966833115, + 0.013982294127345085, + 0.019778510555624962, + 0.012178510427474976, + 0.004714946262538433, + 0.008645801804959774, + 0.006399064790457487, + 0.005064037628471851, + 0.006255341228097677, + 0.0063932244665920734, + 0.007099063601344824, + 0.0071834418922662735, + 0.00576220965012908, + 0.01218758150935173, + 0.008250156417489052, + 0.004152297507971525, + 0.0080012958496809, + 0.007420050911605358, + 0.012338094413280487, + 0.009889448992908001, + 0.007345413789153099, + 0.0064919814467430115, + 0.019753098487854004, + 0.009097529575228691, + 0.005879526026546955, + 0.012940232641994953, + 0.01587514579296112, + 0.01890583336353302, + 0.005928600672632456, + 0.004520760383456945, + 0.006907642353326082, + 0.005474708508700132, + 0.004698103293776512, + 0.005811918992549181, + 0.004676697310060263, + 0.005904593504965305, + 0.0024355952627956867, + 0.0030181927140802145, + 0.005786800291389227, + 0.007705165073275566, + 0.0035808065440505743, + 0.002755991881713271, + 0.00475776381790638, + 0.006669951137155294, + 0.012515095062553883, + 0.018064850941300392, + 0.08394919335842133, + 0.1001414880156517, + 0.011554577387869358, + 0.015067902393639088, + 0.01927577517926693, + 0.03418566286563873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.006734057795256376 + } + ], + "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/0000052/group_13/index_tensor([1988, 1988])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_13/index_tensor([1988, 1988])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f1aca78197cc10577da69247c4d55dcb2a2ba3ad --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_13/index_tensor([1988, 1988])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures an elegant hardcover biography as it slowly and gracefully spirals through an indeterminate, soft space. The book's cover, a deep, desaturated crimson with a matte texture, features a classic black-and-white portrait of a glamorous Hollywood actress from the 1940s. The title, embossed in a timeless serif font, is delicately non-reflective gold foil. As the book spins weightlessly on its vertical axis, its thick, cream-colored pages gently flutter and separate like a slow-motion waterfall, revealing a series of timeless black-and-white photographs from film sets and premieres. The scene is illuminated by a single, soft, diffuse light from above, creating gentle highlights on the book's spine and along the edges of the rippled pages, while the background fades into a dark, out-of-focus void. The visuals are rendered in a rich, desaturated color gradient, deepening the shadows and lending the image a nostalgic, dreamlike quality, perfectly evoking an atmosphere of tranquility and respect for fond memories.", + "seed": 27068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_13/index_1988_seed_27068_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": -0.09620927348732948, + "vlm_logprob": -0.09620927348732948, + "vlm_yes_no_margin": 2.8749999046325683, + "vlm_token_credit": -0.09620927348732948, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24614323675632477 + ], + "opposite_logprob_tokens": [ + -1.6211432218551636 + ], + "logprob_reward": -0.24614323675632477, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24614323675632477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11160136014223099, + 0.08959171175956726, + 0.049796439707279205, + 0.05393989011645317, + 0.39249831438064575, + 0.02994154952466488, + 0.13488635420799255, + 0.05956834927201271, + 0.03883902356028557, + 0.03933699429035187 + ], + "credit_map": [ + 0.013644557446241379, + 0.010926532559096813, + 0.015303830616176128, + 0.01679215393960476, + 0.01122994627803564, + 0.014719363301992416, + 0.01670452021062374, + 0.012280471622943878, + 0.011876956559717655, + 0.003485045861452818, + 0.015078944154083729, + 0.020625019446015358, + 0.005153777077794075, + 0.010100041516125202, + 0.009042451158165932, + 0.014229482039809227, + 0.005344837438315153, + 0.0029613624792546034, + 0.008857355453073978, + 0.008085129782557487, + 0.004618395585566759, + 0.006276729982346296, + 0.005675341933965683, + 0.00797729380428791, + 0.009339436888694763, + 0.0027689854614436626, + 0.009296667762100697, + 0.008059263229370117, + 0.0033346472773700953, + 0.003919631242752075, + 0.008391671814024448, + 0.008829589933156967, + 0.03964253142476082, + 0.01741216890513897, + 0.07169075310230255, + 0.06976941227912903, + 0.018178589642047882, + 0.016598977148532867, + 0.11373867094516754, + 0.04546722397208214, + 0.00675600441172719, + 0.0017105491133406758, + 0.006489995401352644, + 0.00339752365835011, + 0.0024387603625655174, + 0.0024833744391798973, + 0.0031870414968580008, + 0.003478302387520671, + 0.013641267083585262, + 0.02015654928982258, + 0.021433141082525253, + 0.020954305306077003, + 0.01610150747001171, + 0.01050131767988205, + 0.019239133223891258, + 0.012859150767326355, + 0.007721724454313517, + 0.0066289897076785564, + 0.00633517000824213, + 0.011987448669970036, + 0.006339132785797119, + 0.005365656688809395, + 0.009293978102505207, + 0.005896260030567646, + 0.004970855545252562, + 0.0028555956669151783, + 0.009048138745129108, + 0.006115837022662163, + 0.003441875334829092, + 0.0041005415841937065, + 0.003246540203690529, + 0.00505964457988739, + 0.004474669229239225, + 0.004011254757642746, + 0.0061221253126859665, + 0.0084947245195508, + 0.0033048593904823065, + 0.002872626995667815, + 0.003937247209250927, + 0.00611948873847723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.24614323675632477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042304039001465 + ], + "opposite_logprob_tokens": [ + -1.9104230403900146 + ], + "logprob_reward": -0.16042304039001465, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042304039001465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13773202896118164, + 0.14639174938201904, + 0.1217375248670578, + 0.05171971395611763, + 0.30818721652030945, + 0.03757944330573082, + 0.08682344108819962, + 0.04989086091518402, + 0.017584683373570442, + 0.042353298515081406 + ], + "credit_map": [ + 0.0065270378254354, + 0.004925941117107868, + 0.007118574343621731, + 0.008485040627419949, + 0.006005329545587301, + 0.04832646623253822, + 0.008943729102611542, + 0.04739992320537567, + 0.017835861071944237, + 0.00747698126360774, + 0.0237757358700037, + 0.032929014414548874, + 0.00718826986849308, + 0.019618511199951172, + 0.015581118874251842, + 0.021986274048686028, + 0.013586712069809437, + 0.0062717292457818985, + 0.024563178420066833, + 0.023485630750656128, + 0.007383485324680805, + 0.011275981552898884, + 0.013006914407014847, + 0.022163908928632736, + 0.006350258830934763, + 0.0027696534525603056, + 0.008638759143650532, + 0.009068114683032036, + 0.003271777182817459, + 0.0034704972058534622, + 0.010814218781888485, + 0.007336439564824104, + 0.02856134995818138, + 0.013703970238566399, + 0.05596772953867912, + 0.05530325323343277, + 0.014702868647873402, + 0.01445257943123579, + 0.08726957440376282, + 0.03822590038180351, + 0.008875961415469646, + 0.002192526124417782, + 0.006491469219326973, + 0.004787181504070759, + 0.002715596230700612, + 0.004251452628523111, + 0.003822426777333021, + 0.004442836157977581, + 0.009768735617399216, + 0.013984855264425278, + 0.012819971889257431, + 0.011888863518834114, + 0.009742612950503826, + 0.00719094043597579, + 0.012026108801364899, + 0.009401362389326096, + 0.005416606552898884, + 0.005298936739563942, + 0.0040938048623502254, + 0.012030726298689842, + 0.00469632912427187, + 0.005932914093136787, + 0.00781591609120369, + 0.004605632275342941, + 0.0020471997559070587, + 0.001348394202068448, + 0.0027277234476059675, + 0.0032928346190601587, + 0.0013228964526206255, + 0.002323382766917348, + 0.0021980053279548883, + 0.002324247732758522, + 0.005131129175424576, + 0.004637771286070347, + 0.006103023421019316, + 0.006036882754415274, + 0.0033387956209480762, + 0.005437416955828667, + 0.004800246097147465, + 0.006868042983114719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.16042304039001465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199127167463303 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.018199127167463303, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018199127167463303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1596769243478775, + 0.061934541910886765, + 0.0957147553563118, + 0.05871736258268356, + 0.07683919370174408, + 0.09155724942684174, + 0.10426194220781326, + 0.2438846379518509, + 0.04058516025543213, + 0.06682828068733215 + ], + "credit_map": [ + 0.01774485409259796, + 0.007937629707157612, + 0.018014049157500267, + 0.012447389774024487, + 0.007439012173563242, + 0.04099997505545616, + 0.0173850916326046, + 0.03770890831947327, + 0.006687820889055729, + 0.002315196208655834, + 0.011876419186592102, + 0.009535027667880058, + 0.002853366080671549, + 0.0074704475700855255, + 0.005933603271842003, + 0.015262660570442677, + 0.006428403314203024, + 0.003298202995210886, + 0.028270358219742775, + 0.010118323378264904, + 0.004960272461175919, + 0.005599550437182188, + 0.008688670583069324, + 0.02835097350180149, + 0.006464354693889618, + 0.003228353103622794, + 0.01215568371117115, + 0.008844534866511822, + 0.0042634750716388226, + 0.005797212477773428, + 0.010177208110690117, + 0.0077865407802164555, + 0.014089735224843025, + 0.005873709451407194, + 0.012533433735370636, + 0.008961506187915802, + 0.007801843341439962, + 0.006924659013748169, + 0.010141190141439438, + 0.010513119399547577, + 0.019887039437890053, + 0.005906000267714262, + 0.015109186992049217, + 0.014741560444235802, + 0.00875932164490223, + 0.008816713467240334, + 0.01007666066288948, + 0.008260764181613922, + 0.009886945597827435, + 0.007671039085835218, + 0.01191896665841341, + 0.012553668580949306, + 0.0074651846662163734, + 0.006074949633330107, + 0.02941001020371914, + 0.01928117498755455, + 0.02669178694486618, + 0.01919705979526043, + 0.024239355698227882, + 0.06993061304092407, + 0.01753896474838257, + 0.039839968085289, + 0.026513921096920967, + 0.019932975992560387, + 0.007476320024579763, + 0.0033739935606718063, + 0.007138065993785858, + 0.00624329037964344, + 0.0036945862229913473, + 0.003630691207945347, + 0.003274968359619379, + 0.005753242410719395, + 0.0070335413329303265, + 0.007975528948009014, + 0.0095954155549407, + 0.012978657148778439, + 0.0055642640218138695, + 0.00613923417404294, + 0.008446547202765942, + 0.00909509975463152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.018199127167463303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038103725761175156 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.038103725761175156, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038103725761175156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23293013870716095, + 0.07673553377389908, + 0.03903095796704292, + 0.06475505232810974, + 0.06413434445858002, + 0.03503349795937538, + 0.07422353327274323, + 0.28299084305763245, + 0.09590182453393936, + 0.03426431119441986 + ], + "credit_map": [ + 0.008304037153720856, + 0.006908467039465904, + 0.009907149709761143, + 0.011780788190662861, + 0.0075187175534665585, + 0.0909264013171196, + 0.013599341735243797, + 0.08398523926734924, + 0.009022127836942673, + 0.0030223927460610867, + 0.011537251994013786, + 0.013722172006964684, + 0.005127954762428999, + 0.011438477784395218, + 0.007722994778305292, + 0.015142162330448627, + 0.003303913166746497, + 0.002248839708045125, + 0.00941854901611805, + 0.006308934185653925, + 0.00280813779681921, + 0.002679024590179324, + 0.004261492285877466, + 0.008002066053450108, + 0.008438545279204845, + 0.004237744491547346, + 0.014883710071444511, + 0.0076172566041350365, + 0.005611969158053398, + 0.005278348922729492, + 0.009128101170063019, + 0.009559378027915955, + 0.00907081551849842, + 0.0036728077102452517, + 0.010708396323025227, + 0.008645539171993732, + 0.004062725696712732, + 0.004318587947636843, + 0.015997646376490593, + 0.007657824084162712, + 0.006089428439736366, + 0.003499771235510707, + 0.005152454599738121, + 0.004282429814338684, + 0.004573912359774113, + 0.003844973398372531, + 0.0042093112133443356, + 0.003381218295544386, + 0.007702483795583248, + 0.00772536126896739, + 0.013343566097319126, + 0.009184757247567177, + 0.00727271381765604, + 0.006467639468610287, + 0.01029162760823965, + 0.012235380709171295, + 0.028119932860136032, + 0.017835287377238274, + 0.02854274772107601, + 0.08969143778085709, + 0.020594054833054543, + 0.04772505164146423, + 0.030250729992985725, + 0.020231595262885094, + 0.012645264156162739, + 0.0074904668144881725, + 0.01837439462542534, + 0.015720244497060776, + 0.00880174059420824, + 0.009374692104756832, + 0.011510854586958885, + 0.011984165757894516, + 0.0024721194058656693, + 0.0021343871485441923, + 0.0030581566970795393, + 0.010100179351866245, + 0.002549925120547414, + 0.0036176051944494247, + 0.0039042718708515167, + 0.006427667569369078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.038103725761175156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817723736166954 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.01817723736166954, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01817723736166954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3780864179134369, + 0.14975625276565552, + 0.06634614616632462, + 0.03139882534742355, + 0.147044375538826, + 0.04402456060051918, + 0.07332738488912582, + 0.06300067901611328, + 0.01853443868458271, + 0.02848089672625065 + ], + "credit_map": [ + 0.03845827281475067, + 0.010049856267869473, + 0.03754810988903046, + 0.017170937731862068, + 0.012203504331409931, + 0.1282079815864563, + 0.020895816385746002, + 0.11355197429656982, + 0.015440959483385086, + 0.007052427623420954, + 0.02590758353471756, + 0.02647617645561695, + 0.006998312193900347, + 0.022064534947276115, + 0.018520593643188477, + 0.027295676991343498, + 0.0043808952905237675, + 0.002201646799221635, + 0.019327327609062195, + 0.007940584793686867, + 0.003109791548922658, + 0.004212487023323774, + 0.005416289437562227, + 0.01975713111460209, + 0.0026566325686872005, + 0.002154986374080181, + 0.006482100114226341, + 0.004365837667137384, + 0.002057791454717517, + 0.0021391657646745443, + 0.006711321882903576, + 0.004830989520996809, + 0.014685221947729588, + 0.007165808230638504, + 0.026115210726857185, + 0.02508477121591568, + 0.0075936405919492245, + 0.007697556633502245, + 0.040768347680568695, + 0.01793382689356804, + 0.008828653022646904, + 0.002694828435778618, + 0.007286589127033949, + 0.00809942651540041, + 0.004405486863106489, + 0.0031950846314430237, + 0.004462970420718193, + 0.005051526706665754, + 0.008945884183049202, + 0.0064384969882667065, + 0.008439905941486359, + 0.007390232756733894, + 0.006478084251284599, + 0.006326414179056883, + 0.011949737556278706, + 0.017358632758259773, + 0.009174762293696404, + 0.005843201652169228, + 0.006154154427349567, + 0.012588901445269585, + 0.006930933799594641, + 0.007612662855535746, + 0.007999680005013943, + 0.006696387659758329, + 0.002820658963173628, + 0.0024296739138662815, + 0.0014630142832174897, + 0.0029605161398649216, + 0.0020705147180706263, + 0.002255015540868044, + 0.002139826538041234, + 0.002395220100879669, + 0.002417151350528002, + 0.0031011714600026608, + 0.003653992433100939, + 0.005855987314134836, + 0.002457686932757497, + 0.0035331416875123978, + 0.0029336675070226192, + 0.004528099671006203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01817723736166954 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an elegant hardcover biography as it slowly and gracefully spirals through an indeterminate, soft space. The book's cover, a deep, desaturated crimson with a matte texture, features a classic black-and-white portrait of a glamorous Hollywood actress from the 1940s. The title, embossed in a timeless serif font, is delicately non-reflective gold foil. As the book spins weightlessly on its vertical axis, its thick, cream-colored pages gently flutter and separate like a slow-motion waterfall, revealing a series of timeless black-and-white photographs from film sets and premieres. The scene is illuminated by a single, soft, diffuse light from above, creating gentle highlights on the book's spine and along the edges of the rippled pages, while the background fades into a dark, out-of-focus void. The visuals are rendered in a rich, desaturated color gradient, deepening the shadows and lending the image a nostalgic, dreamlike quality, perfectly evoking an atmosphere of tranquility and respect for fond memories.", + "seed": 27069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_13/index_1988_seed_27069_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": -0.0825809946283698, + "vlm_logprob": -0.0825809946283698, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.0825809946283698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32701221108436584 + ], + "opposite_logprob_tokens": [ + -1.3270121812820435 + ], + "logprob_reward": -0.32701221108436584, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32701221108436584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11872997134923935, + 0.08718609809875488, + 0.052870165556669235, + 0.17340701818466187, + 0.05342070758342743, + 0.08377549797296524, + 0.07605256140232086, + 0.06614413112401962, + 0.10211258381605148, + 0.18630127608776093 + ], + "credit_map": [ + 0.010760782286524773, + 0.012440755032002926, + 0.010899529792368412, + 0.010961418971419334, + 0.020578283816576004, + 0.012036244384944439, + 0.015092860907316208, + 0.025960098952054977, + 0.00904005952179432, + 0.0076730637811124325, + 0.008138356730341911, + 0.009032883681356907, + 0.0201604925096035, + 0.010802567936480045, + 0.011579122394323349, + 0.010759560391306877, + 0.007791029755026102, + 0.005246383138000965, + 0.0058548529632389545, + 0.006176457740366459, + 0.006418967619538307, + 0.006474212743341923, + 0.006673244293779135, + 0.008235016837716103, + 0.017664210870862007, + 0.019391318783164024, + 0.014338663779199123, + 0.011837940663099289, + 0.04698202759027481, + 0.022506793960928917, + 0.01815631426870823, + 0.022529739886522293, + 0.005143644753843546, + 0.005349107552319765, + 0.005528315901756287, + 0.005669264122843742, + 0.009913388639688492, + 0.004076923709362745, + 0.010439828038215637, + 0.007300234399735928, + 0.009089842438697815, + 0.008801120333373547, + 0.009363976307213306, + 0.008814032189548016, + 0.010885982774198055, + 0.009876115247607231, + 0.013378153555095196, + 0.01356627605855465, + 0.020046988502144814, + 0.007046026643365622, + 0.013486484065651894, + 0.005934247747063637, + 0.0068933358415961266, + 0.005786614958196878, + 0.006008402444422245, + 0.010850459337234497, + 0.019807610660791397, + 0.005775886587798595, + 0.008712041191756725, + 0.0035198356490582228, + 0.004914548248052597, + 0.0038697856944054365, + 0.004200885072350502, + 0.015343541279435158, + 0.010916714556515217, + 0.007069545332342386, + 0.020479220896959305, + 0.012303591705858707, + 0.0060677072033286095, + 0.0076342117972671986, + 0.00906471535563469, + 0.02857687510550022, + 0.0513371117413044, + 0.010892748832702637, + 0.014248920604586601, + 0.011416527442634106, + 0.008620215579867363, + 0.014251419343054295, + 0.016783401370048523, + 0.05875091999769211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.32701221108436584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309845343232155 + ], + "opposite_logprob_tokens": [ + -3.1680984497070312 + ], + "logprob_reward": -0.04309845343232155, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309845343232155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20559391379356384, + 0.1901516318321228, + 0.07205332815647125, + 0.10099134594202042, + 0.0680394172668457, + 0.0519765205681324, + 0.049821704626083374, + 0.032473765313625336, + 0.1469423621892929, + 0.08195604383945465 + ], + "credit_map": [ + 0.02587527222931385, + 0.019563473761081696, + 0.023717116564512253, + 0.01800731010735035, + 0.024001609534025192, + 0.020322369411587715, + 0.02678864821791649, + 0.04731811210513115, + 0.020981207489967346, + 0.018673168495297432, + 0.016289742663502693, + 0.018982239067554474, + 0.028530627489089966, + 0.018768642097711563, + 0.027249149978160858, + 0.04067683964967728, + 0.010053197853267193, + 0.009700269438326359, + 0.009425629861652851, + 0.007595060858875513, + 0.011614590883255005, + 0.007281488738954067, + 0.0075756944715976715, + 0.008807399310171604, + 0.012689008377492428, + 0.007919460535049438, + 0.016763385385274887, + 0.012089855037629604, + 0.0136179244145751, + 0.006916744634509087, + 0.01683788001537323, + 0.014157092198729515, + 0.007764694280922413, + 0.009148411452770233, + 0.007821537554264069, + 0.005211472511291504, + 0.0163352582603693, + 0.005217152647674084, + 0.006588136777281761, + 0.00995275191962719, + 0.006071920972317457, + 0.008817332796752453, + 0.00576596986502409, + 0.005132652353495359, + 0.0058786701411008835, + 0.00704486109316349, + 0.005956682376563549, + 0.007308433763682842, + 0.009733626618981361, + 0.005444113165140152, + 0.009980601258575916, + 0.0059365578927099705, + 0.006248051300644875, + 0.004793493077158928, + 0.0037665143609046936, + 0.003918744623661041, + 0.0038252207450568676, + 0.0031974585726857185, + 0.005029754247516394, + 0.004023475106805563, + 0.003796293633058667, + 0.002502725925296545, + 0.004456899594515562, + 0.005641939118504524, + 0.048014022409915924, + 0.005429222248494625, + 0.015545693226158619, + 0.00828596018254757, + 0.004715361632406712, + 0.00928007997572422, + 0.006490311119705439, + 0.04918171465396881, + 0.017566977068781853, + 0.007269832771271467, + 0.009180056862533092, + 0.0068555232137441635, + 0.00515789957717061, + 0.006332987453788519, + 0.007260302547365427, + 0.02233247086405754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.04309845343232155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069306060671806 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069306060671806, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069306060671806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07577851414680481, + 0.09070802479982376, + 0.05110920965671539, + 0.17278142273426056, + 0.10006974637508392, + 0.057714950293302536, + 0.0686015859246254, + 0.05388329178094864, + 0.25133004784584045, + 0.07802323251962662 + ], + "credit_map": [ + 0.009792489930987358, + 0.006269416771829128, + 0.007188177667558193, + 0.01003081351518631, + 0.010117077268660069, + 0.010419301688671112, + 0.009997389279305935, + 0.011963850818574429, + 0.008965189568698406, + 0.009079190902411938, + 0.006620755884796381, + 0.010016120970249176, + 0.014971937984228134, + 0.011583024635910988, + 0.01583360694348812, + 0.013638198375701904, + 0.008251623250544071, + 0.0068962289951741695, + 0.006457961164414883, + 0.005268565844744444, + 0.006637695711106062, + 0.004992972128093243, + 0.005881092045456171, + 0.0067230709828436375, + 0.014434434473514557, + 0.019694147631525993, + 0.011815294623374939, + 0.0114889619871974, + 0.05179303511977196, + 0.024444784969091415, + 0.01779702492058277, + 0.021313738077878952, + 0.007243345957249403, + 0.01228734478354454, + 0.006416134536266327, + 0.007060244679450989, + 0.028625940904021263, + 0.011953206732869148, + 0.010680380277335644, + 0.01580314338207245, + 0.005555674433708191, + 0.0060505676083266735, + 0.006764305289834738, + 0.006691773422062397, + 0.006283378228545189, + 0.006928771734237671, + 0.012473173439502716, + 0.006967303808778524, + 0.008908553048968315, + 0.007499550934880972, + 0.007129704114049673, + 0.01012993324548006, + 0.007791927084326744, + 0.006002354435622692, + 0.008253132924437523, + 0.012886431068181992, + 0.01036966685205698, + 0.005698472261428833, + 0.0062652346678078175, + 0.004366267006844282, + 0.009646564722061157, + 0.004010202828794718, + 0.00550834322348237, + 0.00801854208111763, + 0.06810091435909271, + 0.010305939242243767, + 0.027261883020401, + 0.01570933498442173, + 0.009884548373520374, + 0.016847047954797745, + 0.012860212475061417, + 0.09036015719175339, + 0.021327480673789978, + 0.004923802334815264, + 0.005670579615980387, + 0.005138327367603779, + 0.0036645943764597178, + 0.0057897381484508514, + 0.0061846040189266205, + 0.025324108079075813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.016069306060671806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195535331964493 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195535331964493, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195535331964493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08067772537469864, + 0.04019274190068245, + 0.0902126133441925, + 0.19827230274677277, + 0.04663078114390373, + 0.06712467223405838, + 0.05581844970583916, + 0.035828180611133575, + 0.06466527283191681, + 0.3205772340297699 + ], + "credit_map": [ + 0.010229048319160938, + 0.005979557987302542, + 0.006699500139802694, + 0.00826751533895731, + 0.012294299900531769, + 0.008710217662155628, + 0.009301085956394672, + 0.01919650472700596, + 0.005074222106486559, + 0.004129607696086168, + 0.0045703439973294735, + 0.004453937988728285, + 0.0067723034881055355, + 0.0062162503600120544, + 0.004197244066745043, + 0.004778830800205469, + 0.01412905938923359, + 0.00992570910602808, + 0.012105750851333141, + 0.008540225215256214, + 0.015431847423315048, + 0.009904861450195312, + 0.00870512705296278, + 0.011470025405287743, + 0.018374312669038773, + 0.02320585586130619, + 0.013080194592475891, + 0.01566130295395851, + 0.055793456733226776, + 0.027168329805135727, + 0.01770777814090252, + 0.027281058952212334, + 0.0060892547480762005, + 0.004969537258148193, + 0.005504686385393143, + 0.003081555012613535, + 0.010196262039244175, + 0.00391909247264266, + 0.004863325972110033, + 0.008007066324353218, + 0.006753265392035246, + 0.012623102404177189, + 0.008539612405002117, + 0.007250932045280933, + 0.007391049992293119, + 0.008979164063930511, + 0.005945488344877958, + 0.009642058983445168, + 0.0051594078540802, + 0.008112570270895958, + 0.00835471972823143, + 0.006133050657808781, + 0.006201078649610281, + 0.0065706027671694756, + 0.006116699427366257, + 0.009170320816338062, + 0.009702920913696289, + 0.00226280908100307, + 0.00545246247202158, + 0.0033141556195914745, + 0.0018259077332913876, + 0.002788524143397808, + 0.003437109524384141, + 0.007044290192425251, + 0.014496969059109688, + 0.0036855717189610004, + 0.005835695192217827, + 0.00682140002027154, + 0.004397468641400337, + 0.006205842364579439, + 0.004412321373820305, + 0.018810002133250237, + 0.09997722506523132, + 0.017639225348830223, + 0.02205582708120346, + 0.015145406126976013, + 0.008474315516650677, + 0.02222081460058689, + 0.016280492767691612, + 0.11878394335508347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.014195535331964493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012529467232525349 + ], + "opposite_logprob_tokens": [ + -4.387529373168945 + ], + "logprob_reward": -0.012529467232525349, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012529467232525349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0797792449593544, + 0.048186562955379486, + 0.07139386981725693, + 0.06347940117120743, + 0.08337511867284775, + 0.023335760459303856, + 0.02847200445830822, + 0.026335570961236954, + 0.05456775426864624, + 0.5210745930671692 + ], + "credit_map": [ + 0.010457668453454971, + 0.007310985121876001, + 0.008690052665770054, + 0.00936773233115673, + 0.01184767298400402, + 0.009498285129666328, + 0.010105762630701065, + 0.012501087039709091, + 0.004871937446296215, + 0.005436243023723364, + 0.004571524448692799, + 0.005035133101046085, + 0.0071784378960728645, + 0.005839499644935131, + 0.00861605629324913, + 0.00663773063570261, + 0.012948712334036827, + 0.006936602294445038, + 0.01150756236165762, + 0.007885909639298916, + 0.006883350200951099, + 0.006059356965124607, + 0.0089659309014678, + 0.010206449776887894, + 0.009585866704583168, + 0.005420240107923746, + 0.010906388983130455, + 0.007481266278773546, + 0.006680588703602552, + 0.006029745098203421, + 0.007681975606828928, + 0.009693321771919727, + 0.012231487780809402, + 0.005298870149999857, + 0.013037220574915409, + 0.009420560672879219, + 0.011200131848454475, + 0.007317209150642157, + 0.013692951761186123, + 0.011176690459251404, + 0.002889652969315648, + 0.0031697845552116632, + 0.0028933887369930744, + 0.002304608002305031, + 0.0028798840939998627, + 0.002630592091009021, + 0.0026322114281356335, + 0.003935637883841991, + 0.0061728633008897305, + 0.0026231103111058474, + 0.005787424277514219, + 0.0028856524731963873, + 0.003237079596146941, + 0.0017513810889795423, + 0.0030413034837692976, + 0.0029731900431215763, + 0.007341694552451372, + 0.0015719853108748794, + 0.004281800705939531, + 0.0020320445764809847, + 0.0015238865744322538, + 0.0016700526466593146, + 0.0020215464755892754, + 0.005892558954656124, + 0.006154886446893215, + 0.0040446072816848755, + 0.009161717258393764, + 0.00535904336720705, + 0.006417152471840382, + 0.006157307419925928, + 0.003841150552034378, + 0.013431888073682785, + 0.16316665709018707, + 0.027455339208245277, + 0.03391817584633827, + 0.020001491531729698, + 0.010647201910614967, + 0.03694756701588631, + 0.023878242820501328, + 0.2050599306821823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.012529467232525349 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an elegant hardcover biography as it slowly and gracefully spirals through an indeterminate, soft space. The book's cover, a deep, desaturated crimson with a matte texture, features a classic black-and-white portrait of a glamorous Hollywood actress from the 1940s. The title, embossed in a timeless serif font, is delicately non-reflective gold foil. As the book spins weightlessly on its vertical axis, its thick, cream-colored pages gently flutter and separate like a slow-motion waterfall, revealing a series of timeless black-and-white photographs from film sets and premieres. The scene is illuminated by a single, soft, diffuse light from above, creating gentle highlights on the book's spine and along the edges of the rippled pages, while the background fades into a dark, out-of-focus void. The visuals are rendered in a rich, desaturated color gradient, deepening the shadows and lending the image a nostalgic, dreamlike quality, perfectly evoking an atmosphere of tranquility and respect for fond memories.", + "seed": 27070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_13/index_1988_seed_27070_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": -0.11054475679993629, + "vlm_logprob": -0.11054475679993629, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.11054475679993629, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3357352316379547 + ], + "opposite_logprob_tokens": [ + -1.3357352018356323 + ], + "logprob_reward": -0.3357352316379547, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3357352316379547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0858578085899353, + 0.15814730525016785, + 0.10851390659809113, + 0.03416202589869499, + 0.03923143818974495, + 0.3284654915332794, + 0.13802024722099304, + 0.04876184090971947, + 0.03296814486384392, + 0.025871751829981804 + ], + "credit_map": [ + 0.007526069413870573, + 0.007561094127595425, + 0.008286725729703903, + 0.010746320709586143, + 0.008680082857608795, + 0.014966343529522419, + 0.01381524559110403, + 0.0142759308218956, + 0.00940224900841713, + 0.010422350838780403, + 0.010980564169585705, + 0.026364104822278023, + 0.010921661742031574, + 0.03070034459233284, + 0.030962686985731125, + 0.028393344953656197, + 0.009091363288462162, + 0.0097430981695652, + 0.015489703975617886, + 0.018251288682222366, + 0.008149727247655392, + 0.01532647293061018, + 0.016575351357460022, + 0.01588689535856247, + 0.002164517529308796, + 0.002622314728796482, + 0.004496380686759949, + 0.007298035081475973, + 0.0024684180971235037, + 0.004182044416666031, + 0.005946329329162836, + 0.004983988124877214, + 0.004313701298087835, + 0.004172483459115028, + 0.0064252037554979324, + 0.004182625561952591, + 0.00338023342192173, + 0.005019172560423613, + 0.005775582045316696, + 0.0059624346904456615, + 0.022596344351768494, + 0.05918809026479721, + 0.03938939422369003, + 0.05438222736120224, + 0.04392128437757492, + 0.0197067242115736, + 0.05239082872867584, + 0.03689058497548103, + 0.010175586678087711, + 0.014646683819591999, + 0.02230316959321499, + 0.035020384937524796, + 0.01577284187078476, + 0.012155052274465561, + 0.014256692491471767, + 0.013689819723367691, + 0.004090092144906521, + 0.00506433192640543, + 0.009184978902339935, + 0.009405667893588543, + 0.005343507044017315, + 0.003300481243059039, + 0.006910716649144888, + 0.00546206533908844, + 0.003170125652104616, + 0.003200902370736003, + 0.005351623520255089, + 0.004863099195063114, + 0.003927815705537796, + 0.0029148419853299856, + 0.004976291675120592, + 0.004563446156680584, + 0.002677694195881486, + 0.0021949035581201315, + 0.003790410468354821, + 0.004807155579328537, + 0.00306978402659297, + 0.0022525768727064133, + 0.003236526856198907, + 0.0038427007384598255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.3357352316379547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17996086180210114 + ], + "opposite_logprob_tokens": [ + -1.80496084690094 + ], + "logprob_reward": -0.17996086180210114, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17996086180210114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040264442563056946, + 0.13105125725269318, + 0.13276566565036774, + 0.02826036885380745, + 0.09010543674230576, + 0.4055686295032501, + 0.053901854902505875, + 0.06864246726036072, + 0.0404355451464653, + 0.009004313498735428 + ], + "credit_map": [ + 0.0025201495736837387, + 0.005214971490204334, + 0.004696196410804987, + 0.005149370990693569, + 0.004024663474410772, + 0.005685740616172552, + 0.00641272496432066, + 0.006560624577105045, + 0.011586694978177547, + 0.010936611331999302, + 0.00944262184202671, + 0.020381953567266464, + 0.01007413025945425, + 0.0254922304302454, + 0.024116257205605507, + 0.01902075484395027, + 0.008489575237035751, + 0.009722170419991016, + 0.017181631177663803, + 0.02293793484568596, + 0.009727059863507748, + 0.021792450919747353, + 0.024102268740534782, + 0.018812566995620728, + 0.0018345394637435675, + 0.0015473697567358613, + 0.003194964723661542, + 0.005458126775920391, + 0.002136329421773553, + 0.004513376392424107, + 0.005015363916754723, + 0.004560296889394522, + 0.007467543240636587, + 0.010091411881148815, + 0.014091203920543194, + 0.01052895002067089, + 0.010753198526799679, + 0.00929368007928133, + 0.010316060855984688, + 0.01756339706480503, + 0.02752213552594185, + 0.06960799545049667, + 0.04864571616053581, + 0.07026928663253784, + 0.05565682053565979, + 0.02413497306406498, + 0.06595925241708755, + 0.043772436678409576, + 0.0034085793886333704, + 0.004937728401273489, + 0.007976215332746506, + 0.012985496781766415, + 0.006019785534590483, + 0.004249551799148321, + 0.005714070983231068, + 0.008610427379608154, + 0.00536405760794878, + 0.005319265648722649, + 0.017866602167487144, + 0.011860768310725689, + 0.006872706115245819, + 0.004343006759881973, + 0.006840886548161507, + 0.010175174102187157, + 0.004091737326234579, + 0.004296055529266596, + 0.007527845911681652, + 0.00473199924454093, + 0.005320514552295208, + 0.0031591474544256926, + 0.005631635896861553, + 0.005676609929651022, + 0.0014975315425544977, + 0.0008057413506321609, + 0.0012867163168266416, + 0.0016187474830076098, + 0.0007740723085589707, + 0.0008295544539578259, + 0.000910010130610317, + 0.001281940727494657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.17996086180210114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212576672434807 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212576672434807, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212576672434807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16119854152202606, + 0.10212109982967377, + 0.14349094033241272, + 0.12994560599327087, + 0.054895564913749695, + 0.1764477789402008, + 0.08997738361358643, + 0.06906099617481232, + 0.03937341272830963, + 0.03348866477608681 + ], + "credit_map": [ + 0.01717558689415455, + 0.011716940440237522, + 0.03647715225815773, + 0.02047412469983101, + 0.008800454437732697, + 0.020051896572113037, + 0.017867323011159897, + 0.028635047376155853, + 0.006232606712728739, + 0.011080033145844936, + 0.008151648566126823, + 0.018024122342467308, + 0.009029987268149853, + 0.014490713365375996, + 0.019255898892879486, + 0.01585608348250389, + 0.010328393429517746, + 0.011693627573549747, + 0.019346933811903, + 0.027624284848570824, + 0.013179274275898933, + 0.02011364884674549, + 0.022525951266288757, + 0.0186788197606802, + 0.013294290751218796, + 0.012791832908987999, + 0.01607336290180683, + 0.0280670877546072, + 0.013123742304742336, + 0.0146957291290164, + 0.015356200747191906, + 0.01654335856437683, + 0.004858839325606823, + 0.005225285887718201, + 0.007339203730225563, + 0.006803074851632118, + 0.004986999556422234, + 0.007451328914612532, + 0.005774076096713543, + 0.012456757016479969, + 0.011411036364734173, + 0.023740416392683983, + 0.022780677303671837, + 0.03501743823289871, + 0.023884085938334465, + 0.010880228132009506, + 0.030678601935505867, + 0.018055295571684837, + 0.0063344556838274, + 0.008784446865320206, + 0.011235903948545456, + 0.023536639288067818, + 0.008493649773299694, + 0.006211564876139164, + 0.008898301050066948, + 0.016482418403029442, + 0.006524129770696163, + 0.00804936047643423, + 0.011057044379413128, + 0.010861048474907875, + 0.007223169319331646, + 0.004761440679430962, + 0.013743212446570396, + 0.00684159342199564, + 0.004332439508289099, + 0.00284273037686944, + 0.006701881531625986, + 0.006404087878763676, + 0.004218933172523975, + 0.003691488178446889, + 0.0057851048186421394, + 0.005396748427301645, + 0.003039777744561434, + 0.002892924239858985, + 0.00404973654076457, + 0.006058685481548309, + 0.004079445265233517, + 0.0027155422139912844, + 0.005229767877608538, + 0.005422784946858883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.014212576672434807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078924760222435 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016078924760222435, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078924760222435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10242275893688202, + 0.1203426718711853, + 0.11474230885505676, + 0.29923927783966064, + 0.07555154711008072, + 0.0917348712682724, + 0.05967022478580475, + 0.06212511658668518, + 0.04407971724867821, + 0.03009144216775894 + ], + "credit_map": [ + 0.010240040719509125, + 0.007192335091531277, + 0.022217942401766777, + 0.013851599767804146, + 0.005785832181572914, + 0.014334202744066715, + 0.01205686666071415, + 0.016743941232562065, + 0.009395747445523739, + 0.01150275394320488, + 0.007720147259533405, + 0.01877092383801937, + 0.009169572964310646, + 0.02442076802253723, + 0.020890414714813232, + 0.01847233809530735, + 0.008951368741691113, + 0.011500784195959568, + 0.0119063351303339, + 0.021551091223955154, + 0.009785301052033901, + 0.013471802696585655, + 0.020447956398129463, + 0.017127662897109985, + 0.026531362906098366, + 0.02810831181704998, + 0.04524748772382736, + 0.06282776594161987, + 0.025999609380960464, + 0.03626056760549545, + 0.035638708621263504, + 0.03862546756863594, + 0.007361487951129675, + 0.007829003036022186, + 0.010570589452981949, + 0.011659938842058182, + 0.006188708357512951, + 0.008785919286310673, + 0.011014321818947792, + 0.0121415751054883, + 0.005662459880113602, + 0.0160725936293602, + 0.01340263057500124, + 0.012490159831941128, + 0.01080778893083334, + 0.006562425289303064, + 0.01330504845827818, + 0.013431758619844913, + 0.0051725078374147415, + 0.006941211875528097, + 0.007691820617765188, + 0.01004892960190773, + 0.0076303440146148205, + 0.005503731779754162, + 0.009356682188808918, + 0.007324998266994953, + 0.004696205258369446, + 0.006020307540893555, + 0.011584389954805374, + 0.011014712043106556, + 0.006148244254291058, + 0.005359077360481024, + 0.0077389380894601345, + 0.009563242085278034, + 0.004252629354596138, + 0.0042151715606451035, + 0.0062391795217990875, + 0.00686420314013958, + 0.005228618159890175, + 0.00450653163716197, + 0.006967248395085335, + 0.005806136876344681, + 0.002672062022611499, + 0.0021464352030307055, + 0.005662452429533005, + 0.004274495877325535, + 0.0026377474423497915, + 0.0027379195671528578, + 0.003918744623661041, + 0.006041587796062231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.016078924760222435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736189126968384 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736189126968384, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736189126968384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09202420711517334, + 0.14055025577545166, + 0.1086016446352005, + 0.1082523837685585, + 0.09592217206954956, + 0.15515613555908203, + 0.11325440555810928, + 0.08111290633678436, + 0.06122630462050438, + 0.0438995398581028 + ], + "credit_map": [ + 0.009854698553681374, + 0.005779606755822897, + 0.02069678343832493, + 0.011864453554153442, + 0.005409263074398041, + 0.013397451490163803, + 0.009984582662582397, + 0.015037380158901215, + 0.007080292794853449, + 0.011606385931372643, + 0.008813023567199707, + 0.01975674368441105, + 0.011381876654922962, + 0.028108326718211174, + 0.02960299700498581, + 0.024200614541769028, + 0.004862263333052397, + 0.0072924550622701645, + 0.00901180412620306, + 0.01900322176516056, + 0.007467950228601694, + 0.017351076006889343, + 0.0249400045722723, + 0.0186728797852993, + 0.005757906939834356, + 0.005582884419709444, + 0.010362301953136921, + 0.019814543426036835, + 0.008502868935465813, + 0.01533439476042986, + 0.024324357509613037, + 0.018573133274912834, + 0.010270334780216217, + 0.011471773497760296, + 0.01463910099118948, + 0.011159581132233143, + 0.009593706578016281, + 0.012147593311965466, + 0.010500500909984112, + 0.016139589250087738, + 0.009948676452040672, + 0.023652009665966034, + 0.02063676156103611, + 0.029071055352687836, + 0.01907048560678959, + 0.009197400882840157, + 0.02614978514611721, + 0.017429981380701065, + 0.011593499220907688, + 0.012452785857021809, + 0.018653875216841698, + 0.022259293124079704, + 0.010948900133371353, + 0.011032173410058022, + 0.016126461327075958, + 0.010187423788011074, + 0.006603474263101816, + 0.0071929204277694225, + 0.014035135507583618, + 0.014110415242612362, + 0.008686088025569916, + 0.006218648981302977, + 0.013652253895998001, + 0.010613977909088135, + 0.0072479150258004665, + 0.006651573348790407, + 0.011239641346037388, + 0.007720345631241798, + 0.007651778869330883, + 0.00457422249019146, + 0.008467800915241241, + 0.0076730335131287575, + 0.005592760164290667, + 0.003847719170153141, + 0.007032996974885464, + 0.008281762711703777, + 0.004580425564199686, + 0.003370906924828887, + 0.004866895731538534, + 0.006326078437268734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.006736189126968384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an elegant hardcover biography as it slowly and gracefully spirals through an indeterminate, soft space. The book's cover, a deep, desaturated crimson with a matte texture, features a classic black-and-white portrait of a glamorous Hollywood actress from the 1940s. The title, embossed in a timeless serif font, is delicately non-reflective gold foil. As the book spins weightlessly on its vertical axis, its thick, cream-colored pages gently flutter and separate like a slow-motion waterfall, revealing a series of timeless black-and-white photographs from film sets and premieres. The scene is illuminated by a single, soft, diffuse light from above, creating gentle highlights on the book's spine and along the edges of the rippled pages, while the background fades into a dark, out-of-focus void. The visuals are rendered in a rich, desaturated color gradient, deepening the shadows and lending the image a nostalgic, dreamlike quality, perfectly evoking an atmosphere of tranquility and respect for fond memories.", + "seed": 27071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_13/index_1988_seed_27071_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": -0.057194477692246434, + "vlm_logprob": -0.057194477692246434, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.057194477692246434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20375485718250275 + ], + "opposite_logprob_tokens": [ + -1.8287549018859863 + ], + "logprob_reward": -0.20375485718250275, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.20375485718250275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1687929481267929, + 0.06769528985023499, + 0.08459824323654175, + 0.03940685838460922, + 0.1085711345076561, + 0.10343242436647415, + 0.22054259479045868, + 0.09336395561695099, + 0.06389398872852325, + 0.049702562391757965 + ], + "credit_map": [ + 0.01629737764596939, + 0.022848112508654594, + 0.02009844221174717, + 0.012106476351618767, + 0.01361048873513937, + 0.027087120339274406, + 0.016195837408304214, + 0.040549084544181824, + 0.005946412216871977, + 0.007778109982609749, + 0.0065485406666994095, + 0.008418123237788677, + 0.006786275655031204, + 0.011120686307549477, + 0.007936984300613403, + 0.013160152360796928, + 0.005705142393708229, + 0.00586712034419179, + 0.006968147587031126, + 0.006081932689994574, + 0.007963623851537704, + 0.023316765204072, + 0.008069461211562157, + 0.020626038312911987, + 0.003602617885917425, + 0.004549113102257252, + 0.004581643734127283, + 0.002434144262224436, + 0.0036248830147087574, + 0.009787782095372677, + 0.00339304655790329, + 0.007433623541146517, + 0.010898938402533531, + 0.00892591942101717, + 0.018738439306616783, + 0.014801821671426296, + 0.008366771042346954, + 0.022198446094989777, + 0.013831700198352337, + 0.0108090965077281, + 0.00974219385534525, + 0.011217203922569752, + 0.01405512448400259, + 0.02587013877928257, + 0.008849259465932846, + 0.013730504550039768, + 0.010920464061200619, + 0.009047524072229862, + 0.017650051042437553, + 0.0228184275329113, + 0.04759053885936737, + 0.03786658123135567, + 0.011699034832417965, + 0.013921177946031094, + 0.04023837298154831, + 0.02875838242471218, + 0.011042348109185696, + 0.0076607889495790005, + 0.011032717302441597, + 0.007805336732417345, + 0.005164996720850468, + 0.01241624541580677, + 0.01526673324406147, + 0.022974779829382896, + 0.009574447758495808, + 0.005566523410379887, + 0.0081446198746562, + 0.006846417672932148, + 0.005633849184960127, + 0.010087382048368454, + 0.007573415525257587, + 0.010467332787811756, + 0.008238338865339756, + 0.005899020470678806, + 0.004501438699662685, + 0.003991235513240099, + 0.0046144574880599976, + 0.008375006727874279, + 0.005128609016537666, + 0.008954452350735664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.20375485718250275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309936612844467 + ], + "opposite_logprob_tokens": [ + -3.1680994033813477 + ], + "logprob_reward": -0.04309936612844467, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309936612844467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05698930099606514, + 0.09644439816474915, + 0.08810875564813614, + 0.17264406383037567, + 0.1835281401872635, + 0.044127028435468674, + 0.11889869719743729, + 0.11019038408994675, + 0.08834186941385269, + 0.04072744771838188 + ], + "credit_map": [ + 0.005410175304859877, + 0.007953841239213943, + 0.0062409136444330215, + 0.006933972239494324, + 0.004355935379862785, + 0.007374860346317291, + 0.009614055044949055, + 0.00910553615540266, + 0.009949282743036747, + 0.011505159549415112, + 0.014145721681416035, + 0.010092953220009804, + 0.011190668679773808, + 0.014884141273796558, + 0.007300431374460459, + 0.01737602986395359, + 0.00535815954208374, + 0.004405565559864044, + 0.0054079387336969376, + 0.004106331150978804, + 0.007366659119725227, + 0.03454110771417618, + 0.004080857615917921, + 0.022842124104499817, + 0.010700861923396587, + 0.007990486919879913, + 0.018839126452803612, + 0.005126239266246557, + 0.007440092507749796, + 0.05802595615386963, + 0.00456950394436717, + 0.059951793402433395, + 0.011091930791735649, + 0.007854841649532318, + 0.03282850235700607, + 0.008946757763624191, + 0.014436702243983746, + 0.06416811048984528, + 0.014393927529454231, + 0.029807347804307938, + 0.005389955826103687, + 0.004206736106425524, + 0.005657307803630829, + 0.009175647981464863, + 0.003973830956965685, + 0.006807034835219383, + 0.0037284193094819784, + 0.00518809026107192, + 0.012157699093222618, + 0.013207897543907166, + 0.0246827881783247, + 0.01742628961801529, + 0.00661090575158596, + 0.014879828318953514, + 0.016904233023524284, + 0.013029040768742561, + 0.012465931475162506, + 0.007109822239726782, + 0.010528852231800556, + 0.011101346462965012, + 0.010403167456388474, + 0.01607317477464676, + 0.020503101870417595, + 0.022004973143339157, + 0.012882181443274021, + 0.007973629981279373, + 0.009569019079208374, + 0.01078607328236103, + 0.006404304411262274, + 0.013328107073903084, + 0.011674505658447742, + 0.015724046155810356, + 0.00690347608178854, + 0.005200628191232681, + 0.0037311373744159937, + 0.004026207607239485, + 0.0031457487493753433, + 0.00706011475995183, + 0.0038350929971784353, + 0.006825042888522148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.04309936612844467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190816044807434 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018190816044807434, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190816044807434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380731165409088, + 0.0715637132525444, + 0.09086281806230545, + 0.11190927028656006, + 0.10960202664136887, + 0.09605495631694794, + 0.06278204917907715, + 0.06638701260089874, + 0.17327533662319183, + 0.07948975265026093 + ], + "credit_map": [ + 0.012945739552378654, + 0.017834564670920372, + 0.021100878715515137, + 0.010546439327299595, + 0.012187150306999683, + 0.022092774510383606, + 0.013903495855629444, + 0.027462052181363106, + 0.009071262553334236, + 0.00949232280254364, + 0.0116419093683362, + 0.006479929201304913, + 0.007447372190654278, + 0.010896727442741394, + 0.005299445241689682, + 0.011234736070036888, + 0.004891699180006981, + 0.006105648819357157, + 0.006152802612632513, + 0.005218781530857086, + 0.008290876634418964, + 0.031163375824689865, + 0.006556931417435408, + 0.02248268388211727, + 0.006070561241358519, + 0.0031670627649873495, + 0.0051292357966303825, + 0.0023413216695189476, + 0.0034500150941312313, + 0.04304695501923561, + 0.003575329901650548, + 0.045128777623176575, + 0.0111868716776371, + 0.008902653120458126, + 0.01591777242720127, + 0.015308820642530918, + 0.008292386308312416, + 0.018307305872440338, + 0.014932116493582726, + 0.016754090785980225, + 0.013125213794410229, + 0.010084935463964939, + 0.014452370814979076, + 0.015283693559467793, + 0.00810548011213541, + 0.013611690141260624, + 0.008618615567684174, + 0.012772949412465096, + 0.006763382814824581, + 0.009572065435349941, + 0.011646059341728687, + 0.008336926810443401, + 0.0038347169756889343, + 0.005719831679016352, + 0.008016183972358704, + 0.008892872370779514, + 0.00810055248439312, + 0.005329633131623268, + 0.008775894530117512, + 0.004104356747120619, + 0.003294396447017789, + 0.008885201998054981, + 0.01113738864660263, + 0.01675957627594471, + 0.02234000898897648, + 0.014699478633701801, + 0.024287626147270203, + 0.02142420783638954, + 0.013611851260066032, + 0.026661640033125877, + 0.02289211004972458, + 0.027358409017324448, + 0.012135743163526058, + 0.00876239687204361, + 0.005442822352051735, + 0.00813598744571209, + 0.00821616966277361, + 0.015178313478827477, + 0.009759565815329552, + 0.01185875479131937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.018190816044807434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194712042808533 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194712042808533, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194712042808533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09523813426494598, + 0.021559938788414, + 0.066591776907444, + 0.06436052173376083, + 0.299004465341568, + 0.05944472551345825, + 0.14097735285758972, + 0.07787298411130905, + 0.14000733196735382, + 0.03494272008538246 + ], + "credit_map": [ + 0.008724935352802277, + 0.012774950824677944, + 0.009382020682096481, + 0.009982322342693806, + 0.01061254646629095, + 0.012135623954236507, + 0.01567727141082287, + 0.015948472544550896, + 0.0027650725096464157, + 0.004126167856156826, + 0.0022199261002242565, + 0.0025586276315152645, + 0.0023046184796839952, + 0.003465077606961131, + 0.0019529950805008411, + 0.0021674558520317078, + 0.005913870874792337, + 0.008852818049490452, + 0.007009583991020918, + 0.008103182539343834, + 0.009963939897716045, + 0.01235780119895935, + 0.007079813629388809, + 0.007310768589377403, + 0.004911698400974274, + 0.004833178129047155, + 0.013457746244966984, + 0.003324134275317192, + 0.003860167693346739, + 0.016146905720233917, + 0.003897196613252163, + 0.013929505832493305, + 0.018430659547448158, + 0.01572662964463234, + 0.06347905099391937, + 0.01861906796693802, + 0.023665662854909897, + 0.11812812834978104, + 0.02327544055879116, + 0.017679870128631592, + 0.007706151809543371, + 0.007781268563121557, + 0.008216571994125843, + 0.010734078474342823, + 0.005334675312042236, + 0.007607916835695505, + 0.005817900411784649, + 0.0062461672350764275, + 0.0127280717715621, + 0.014519942924380302, + 0.029147017747163773, + 0.02328076772391796, + 0.007526025641709566, + 0.009824469685554504, + 0.024463526904582977, + 0.019487544894218445, + 0.00947099644690752, + 0.006689972709864378, + 0.013208356685936451, + 0.011218872852623463, + 0.006712482776492834, + 0.010110844857990742, + 0.010093794204294682, + 0.01036767102777958, + 0.017011672258377075, + 0.009601233527064323, + 0.01696545071899891, + 0.017366424202919006, + 0.015313277952373028, + 0.019115479663014412, + 0.021555189043283463, + 0.02307860739529133, + 0.004077743273228407, + 0.0029455828480422497, + 0.0030334910843521357, + 0.0031092846766114235, + 0.003232871415093541, + 0.005406876560300589, + 0.005285616964101791, + 0.007851256057620049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.014194712042808533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant hardcover biography book?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the book cover deep desaturated crimson with a matte texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spin slowly and gracefully on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cream-colored pages flutter and separate like a slow-motion waterfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus void illuminated by a single soft diffuse light from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11020360887050629, + 0.08770035952329636, + 0.025672297924757004, + 0.051186297088861465, + 0.1913822591304779, + 0.2231495976448059, + 0.09464388340711594, + 0.0761546865105629, + 0.0584908127784729, + 0.08141621947288513 + ], + "credit_map": [ + 0.00862735603004694, + 0.008520659059286118, + 0.01158651802688837, + 0.009257277473807335, + 0.0085489172488451, + 0.021043527871370316, + 0.012921659275889397, + 0.02969769947230816, + 0.009177767671644688, + 0.009125332348048687, + 0.013889706693589687, + 0.01052154041826725, + 0.009269909001886845, + 0.012710893526673317, + 0.006482611410319805, + 0.016522597521543503, + 0.002028051298111677, + 0.0027937216218560934, + 0.002764526754617691, + 0.003213013056665659, + 0.0036624171771109104, + 0.0038390993140637875, + 0.003665169235318899, + 0.00370630226098001, + 0.004260481335222721, + 0.006869099102914333, + 0.007437354885041714, + 0.0034914687275886536, + 0.00526922894641757, + 0.012509562075138092, + 0.003049232065677643, + 0.008299865759909153, + 0.010985583998262882, + 0.008165317587554455, + 0.03493419289588928, + 0.0070979236625134945, + 0.015374336391687393, + 0.0728258267045021, + 0.010658806189894676, + 0.03134026750922203, + 0.022921660915017128, + 0.024373849853873253, + 0.032955508679151535, + 0.03947993740439415, + 0.014183012768626213, + 0.031193289905786514, + 0.022498026490211487, + 0.035544294863939285, + 0.009003587067127228, + 0.0113452123478055, + 0.015744782984256744, + 0.011748513206839561, + 0.007230574730783701, + 0.014477203600108624, + 0.013281040824949741, + 0.011812970042228699, + 0.011950352229177952, + 0.006376747973263264, + 0.009943627752363682, + 0.009245943278074265, + 0.005837155040353537, + 0.011476090177893639, + 0.009878074750304222, + 0.01144669484347105, + 0.009630711749196053, + 0.0040261331014335155, + 0.005586900282651186, + 0.00492226192727685, + 0.00476448331028223, + 0.0074143847450613976, + 0.008355425670742989, + 0.013790511526167393, + 0.011320607736706734, + 0.007088431157171726, + 0.008207738399505615, + 0.006317534483969212, + 0.008450038731098175, + 0.013543244451284409, + 0.006893340498209, + 0.019595278427004814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0067326370626688 + } + ], + "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/0000052/group_14/index_tensor([2405, 2405])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_14/index_tensor([2405, 2405])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..344b45b63a2bc1d669e20c9061c50264c22a135c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_14/index_tensor([2405, 2405])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from an upward angle, the camera slowly pans right across the murky seafloor, revealing a giant octopus perched atop a patch of ancient volcanic rock covered in deep green algae. Its bulbous head is massive and heavy, its wrinkled and textured skin a shifting iridescence, transitioning from deep indigo to vibrant magenta, dotted with faint, star-like bioluminescent specks. One of its enormous, iridescent tentacle unfolds gracefully and powerfully, like a giant beast awakening. The neatly arranged white suction cups on its inner surface expand and contract in the water as it stretches, as if testing the tranquil waters surrounding it. A few faint rays of Jesus light penetrated through the sea surface, forming swaying beams of light in the deep blue water full of suspended particles, which just illuminated the octopus's outstretched limbs, highlighting the slippery luster and complex color changes of its skin. The occasional silvery schools of fish quickly passed by in the background, further contrasting the calm and composed posture of this giant beast. The whole scene was filled with a tranquil, mysterious and awe-inspiring atmosphere, showing the magnificence and wonder of deep-sea life.", + "seed": 908, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_14/index_2405_seed_908_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": -0.12652352498844266, + "vlm_logprob": -0.12652352498844266, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.12652352498844266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2974456250667572 + ], + "opposite_logprob_tokens": [ + -1.4224456548690796 + ], + "logprob_reward": -0.2974456250667572, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2974456250667572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026333559304475784, + 0.013395789079368114, + 0.025414370000362396, + 0.04380745440721512, + 0.04788820818066597, + 0.05646069720387459, + 0.704546332359314, + 0.04131796956062317, + 0.02813641168177128, + 0.012699230574071407 + ], + "credit_map": [ + 0.002475256798788905, + 0.00302661070600152, + 0.0033293701708316803, + 0.004093831405043602, + 0.0024956855922937393, + 0.004415228497236967, + 0.0034472504630684853, + 0.0030503247398883104, + 0.0016207908047363162, + 0.0015439982526004314, + 0.0016943643568083644, + 0.0018569715321063995, + 0.0010579212103039026, + 0.002013596473261714, + 0.002245950512588024, + 0.0013621965190395713, + 0.004079814534634352, + 0.0033061837311834097, + 0.0030590586829930544, + 0.0033990549854934216, + 0.0030352240428328514, + 0.0034146008547395468, + 0.0021887256298214197, + 0.0029317070730030537, + 0.00730592617765069, + 0.006391663569957018, + 0.003658807836472988, + 0.004789762664586306, + 0.0068700965493917465, + 0.0045156171545386314, + 0.005894268397241831, + 0.004381314851343632, + 0.007596371695399284, + 0.010389182716608047, + 0.004353371448814869, + 0.005912892986088991, + 0.004958863370120525, + 0.0052557471208274364, + 0.003679362591356039, + 0.005742418114095926, + 0.0064741079695522785, + 0.008270408026874065, + 0.005323194898664951, + 0.007143175229430199, + 0.008154936134815216, + 0.005430765450000763, + 0.007721857633441687, + 0.007942252792418003, + 0.3183840811252594, + 0.3027200698852539, + 0.011210043914616108, + 0.015717385336756706, + 0.0195058211684227, + 0.010729282163083553, + 0.014209295623004436, + 0.012070349417626858, + 0.004420817364007235, + 0.005284118931740522, + 0.004725186619907618, + 0.0056076799519360065, + 0.004918780643492937, + 0.005501840263605118, + 0.00554679473862052, + 0.005312748719006777, + 0.004604429006576538, + 0.006878118962049484, + 0.002058468060567975, + 0.0032963186968117952, + 0.00322736706584692, + 0.002370416186749935, + 0.0023390238638967276, + 0.003362267278134823, + 0.0015998798189684749, + 0.0017243644688278437, + 0.0014152094954624772, + 0.000889633665792644, + 0.0030182513874024153, + 0.0016726640751585364, + 0.0009736738866195083, + 0.0014055540086701512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.2974456250667572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813163101673126 + ], + "opposite_logprob_tokens": [ + -1.4063162803649902 + ], + "logprob_reward": -0.2813163101673126, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813163101673126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033153992146253586, + 0.046873144805431366, + 0.07808748632669449, + 0.2230205535888672, + 0.16586196422576904, + 0.037638839334249496, + 0.27362772822380066, + 0.029169000685214996, + 0.0985821783542633, + 0.01398514211177826 + ], + "credit_map": [ + 0.004227413795888424, + 0.004194946959614754, + 0.003956336062401533, + 0.00382025889120996, + 0.003083225805312395, + 0.004709052853286266, + 0.004220941103994846, + 0.004941817373037338, + 0.007501036394387484, + 0.006745365913957357, + 0.00492835370823741, + 0.007123897317796946, + 0.003884501988068223, + 0.005837431643158197, + 0.006260208319872618, + 0.004592349752783775, + 0.01510005071759224, + 0.0161103717982769, + 0.013317740522325039, + 0.008836997672915459, + 0.005010590422898531, + 0.0061835553497076035, + 0.00633772648870945, + 0.007190457079559565, + 0.03189451992511749, + 0.021085835993289948, + 0.02203955501317978, + 0.021620219573378563, + 0.03761410340666771, + 0.03429684787988663, + 0.025930246338248253, + 0.02853921428322792, + 0.02931254357099533, + 0.041381314396858215, + 0.015899525955319405, + 0.023203693330287933, + 0.013405930250883102, + 0.018062327057123184, + 0.009514784440398216, + 0.015081854537129402, + 0.006011149380356073, + 0.007603236939758062, + 0.0026835838798433542, + 0.005086027085781097, + 0.00400685565546155, + 0.0029660798609256744, + 0.005052207503467798, + 0.0042296964675188065, + 0.12330527603626251, + 0.11711468547582626, + 0.0038338839076459408, + 0.005219609010964632, + 0.008555001579225063, + 0.005169084761291742, + 0.005164924077689648, + 0.005265255458652973, + 0.003986756317317486, + 0.005427309777587652, + 0.0027533965185284615, + 0.0032874918542802334, + 0.0027965574990957975, + 0.0036544748581945896, + 0.0035452647134661674, + 0.0037177489139139652, + 0.011887137778103352, + 0.010494341142475605, + 0.010709904134273529, + 0.007812109310179949, + 0.010995187796652317, + 0.015949659049510956, + 0.011586198583245277, + 0.01914764568209648, + 0.0016415166901424527, + 0.0016397510189563036, + 0.0012200067285448313, + 0.0012475671246647835, + 0.002966764383018017, + 0.0019186707213521004, + 0.001171574811451137, + 0.0021792904008179903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.2813163101673126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254771463572979 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.01254771463572979, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254771463572979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034862030297517776, + 0.03310983255505562, + 0.03720194473862648, + 0.02499113604426384, + 0.04718597233295441, + 0.021175403147935867, + 0.7341273427009583, + 0.022849079221487045, + 0.035310737788677216, + 0.00918655190616846 + ], + "credit_map": [ + 0.004573225509375334, + 0.004225066397339106, + 0.004455295391380787, + 0.004416592884808779, + 0.004365807399153709, + 0.005172484554350376, + 0.003528701141476631, + 0.004124860279262066, + 0.003612287575379014, + 0.004241287708282471, + 0.004694242030382156, + 0.004939447622746229, + 0.003206662368029356, + 0.004313684068620205, + 0.004011136945337057, + 0.004091083537787199, + 0.008785455487668514, + 0.009612401016056538, + 0.003436136059463024, + 0.004457399249076843, + 0.0025555819738656282, + 0.0026086538564413786, + 0.002322934800758958, + 0.003423380432650447, + 0.0031900673639029264, + 0.0022402352187782526, + 0.0039233677089214325, + 0.002367109525948763, + 0.003349345177412033, + 0.0035804782528430223, + 0.002264016307890415, + 0.004076519049704075, + 0.00895722396671772, + 0.011943287216126919, + 0.004264729097485542, + 0.006844726391136646, + 0.003952065482735634, + 0.004262621980160475, + 0.0032762931659817696, + 0.003685023169964552, + 0.0028025645297020674, + 0.0026205750182271004, + 0.001501736231148243, + 0.0016274686204269528, + 0.0036793199833482504, + 0.0037255752831697464, + 0.0024317146744579077, + 0.002786447759717703, + 0.32402244210243225, + 0.30807119607925415, + 0.012115942314267159, + 0.014500576071441174, + 0.02665236033499241, + 0.01754307746887207, + 0.01635514199733734, + 0.014866535551846027, + 0.003434930695220828, + 0.004457368515431881, + 0.0022231487091630697, + 0.0026569904293864965, + 0.0024722469970583916, + 0.002371347276493907, + 0.002852628706023097, + 0.002380418125540018, + 0.004759974777698517, + 0.005553327035158873, + 0.0030249962583184242, + 0.004748817533254623, + 0.004268565680831671, + 0.003122845198959112, + 0.004707376006990671, + 0.00512483436614275, + 0.0010744832688942552, + 0.001085194293409586, + 0.0011389884166419506, + 0.001265917089767754, + 0.0012531563406810164, + 0.0009788057068362832, + 0.0010615995852276683, + 0.001328407321125269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.01254771463572979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626701612025499 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626701612025499, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626701612025499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04392636567354202, + 0.034792210906744, + 0.08353902399539948, + 0.07888665050268173, + 0.16810132563114166, + 0.022107833996415138, + 0.24082931876182556, + 0.09722679108381271, + 0.1391792893409729, + 0.09141109883785248 + ], + "credit_map": [ + 0.006512944120913744, + 0.00556170241907239, + 0.005175381898880005, + 0.005916039925068617, + 0.005020961631089449, + 0.0062481798231601715, + 0.005024456884711981, + 0.004466699901968241, + 0.004387480206787586, + 0.004178163595497608, + 0.004371872171759605, + 0.0055927676148712635, + 0.0033710573334246874, + 0.004630597308278084, + 0.004569867625832558, + 0.0036904048174619675, + 0.020805537700653076, + 0.023629773408174515, + 0.008056988008320332, + 0.010938704945147038, + 0.005461661145091057, + 0.005403041373938322, + 0.004723239224404097, + 0.004520086105912924, + 0.010350614786148071, + 0.006713857874274254, + 0.0076571982353925705, + 0.007524928543716669, + 0.012713401578366756, + 0.01276414655148983, + 0.00955837219953537, + 0.011604135856032372, + 0.031315259635448456, + 0.0445726253092289, + 0.016386734321713448, + 0.024588948115706444, + 0.012725528329610825, + 0.015373141504824162, + 0.009387616999447346, + 0.013751489110291004, + 0.0032454621978104115, + 0.0035850657150149345, + 0.0017362205544486642, + 0.003012469271197915, + 0.0027913847006857395, + 0.0025056202430278063, + 0.002625157590955496, + 0.002606455236673355, + 0.09459499269723892, + 0.08973314613103867, + 0.0067523629404604435, + 0.008305787108838558, + 0.01087119709700346, + 0.009005562402307987, + 0.013186515308916569, + 0.008379763923585415, + 0.011799285188317299, + 0.013052590191364288, + 0.011997783556580544, + 0.013731499202549458, + 0.010325401090085506, + 0.010809873230755329, + 0.013741283677518368, + 0.01176909077912569, + 0.021695677191019058, + 0.03314735367894173, + 0.010765875689685345, + 0.011747404932975769, + 0.01580657809972763, + 0.014599105343222618, + 0.014814496040344238, + 0.016602816060185432, + 0.011121437884867191, + 0.01417534239590168, + 0.011236223392188549, + 0.010447264648973942, + 0.010736697353422642, + 0.008601170033216476, + 0.011297470889985561, + 0.013795501552522182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.007626701612025499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033681273460388184 + ], + "opposite_logprob_tokens": [ + -3.4086813926696777 + ], + "logprob_reward": -0.033681273460388184, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033681273460388184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13087649643421173, + 0.062571220099926, + 0.11132057011127472, + 0.08239685744047165, + 0.1129671260714531, + 0.018408242613077164, + 0.39367374777793884, + 0.010221938602626324, + 0.05502643063664436, + 0.02253740467131138 + ], + "credit_map": [ + 0.019841833040118217, + 0.01833789236843586, + 0.015682818368077278, + 0.022441908717155457, + 0.0070876250974833965, + 0.015089963562786579, + 0.014762760140001774, + 0.01763169839978218, + 0.013436298817396164, + 0.00969039835035801, + 0.012185035273432732, + 0.005805757362395525, + 0.0044938381761312485, + 0.006363384425640106, + 0.004055467434227467, + 0.00654103746637702, + 0.02441202849149704, + 0.027554238215088844, + 0.0118857491761446, + 0.01459130272269249, + 0.007416470907628536, + 0.00788839627057314, + 0.008635055273771286, + 0.008937321603298187, + 0.012281252071261406, + 0.008353921584784985, + 0.009250747039914131, + 0.007441711612045765, + 0.013979057781398296, + 0.011546637862920761, + 0.009017969481647015, + 0.010525546036660671, + 0.020879847928881645, + 0.029588527977466583, + 0.010095005854964256, + 0.017535429447889328, + 0.008674299344420433, + 0.010970398783683777, + 0.006054423283785582, + 0.00916918832808733, + 0.0027766262646764517, + 0.002190944505855441, + 0.001741343759931624, + 0.00177426275331527, + 0.0021031643263995647, + 0.0027025288436561823, + 0.002885748166590929, + 0.002233624691143632, + 0.17162184417247772, + 0.1631641834974289, + 0.006721075624227524, + 0.008251186460256577, + 0.013603050261735916, + 0.008068558759987354, + 0.011205197311937809, + 0.01103861816227436, + 0.001164293847978115, + 0.0012497525895014405, + 0.001101102214306593, + 0.0011108465259894729, + 0.001199550461024046, + 0.0013817079598084092, + 0.0012636510655283928, + 0.0017510333564132452, + 0.006784981116652489, + 0.006284442730247974, + 0.005955942906439304, + 0.004495987668633461, + 0.005946921184659004, + 0.008175545372068882, + 0.006944917608052492, + 0.010437688790261745, + 0.0029221500735729933, + 0.004024151712656021, + 0.0023978978861123323, + 0.0023918121587485075, + 0.0029532734770327806, + 0.0018174245487898588, + 0.002941369777545333, + 0.00308932363986969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.033681273460388184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an upward angle, the camera slowly pans right across the murky seafloor, revealing a giant octopus perched atop a patch of ancient volcanic rock covered in deep green algae. Its bulbous head is massive and heavy, its wrinkled and textured skin a shifting iridescence, transitioning from deep indigo to vibrant magenta, dotted with faint, star-like bioluminescent specks. One of its enormous, iridescent tentacle unfolds gracefully and powerfully, like a giant beast awakening. The neatly arranged white suction cups on its inner surface expand and contract in the water as it stretches, as if testing the tranquil waters surrounding it. A few faint rays of Jesus light penetrated through the sea surface, forming swaying beams of light in the deep blue water full of suspended particles, which just illuminated the octopus's outstretched limbs, highlighting the slippery luster and complex color changes of its skin. The occasional silvery schools of fish quickly passed by in the background, further contrasting the calm and composed posture of this giant beast. The whole scene was filled with a tranquil, mysterious and awe-inspiring atmosphere, showing the magnificence and wonder of deep-sea life.", + "seed": 909, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_14/index_2405_seed_909_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": -0.0838346746750176, + "vlm_logprob": -0.0838346746750176, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.0838346746750176, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2977783679962158 + ], + "opposite_logprob_tokens": [ + -1.4227783679962158 + ], + "logprob_reward": -0.2977783679962158, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2977783679962158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10355489701032639, + 0.031178170815110207, + 0.03982072323560715, + 0.03686085343360901, + 0.0192932840436697, + 0.10563217848539352, + 0.15442387759685516, + 0.27447935938835144, + 0.21057596802711487, + 0.024180732667446136 + ], + "credit_map": [ + 0.01409938931465149, + 0.008641407825052738, + 0.011362175457179546, + 0.011641953140497208, + 0.014081375673413277, + 0.013650556094944477, + 0.018437638878822327, + 0.011640398763120174, + 0.003860571887344122, + 0.0037518590688705444, + 0.0038975172210484743, + 0.0033076987601816654, + 0.0020263632759451866, + 0.0037001546006649733, + 0.0033873787615448236, + 0.007246626541018486, + 0.0056650275364518166, + 0.006531591061502695, + 0.0053883325308561325, + 0.003611912950873375, + 0.005138007923960686, + 0.004595847334712744, + 0.004286704584956169, + 0.0046032993122935295, + 0.006899351719766855, + 0.006052027456462383, + 0.005962091963738203, + 0.002912989817559719, + 0.0036444782745093107, + 0.005177910439670086, + 0.003316141664981842, + 0.0028958595357835293, + 0.0020460074301809072, + 0.003110789693892002, + 0.00236591137945652, + 0.0020434593316167593, + 0.002222586190328002, + 0.0025249789468944073, + 0.002628613729029894, + 0.002350935712456703, + 0.01789766550064087, + 0.02030581422150135, + 0.012852820567786694, + 0.011901837773621082, + 0.009091093204915524, + 0.01113878283649683, + 0.01396664697676897, + 0.008477510884404182, + 0.025045622140169144, + 0.029211435467004776, + 0.015701616182923317, + 0.01125610712915659, + 0.01832994818687439, + 0.014000041410326958, + 0.0197638850659132, + 0.02111521176993847, + 0.028679944574832916, + 0.036014311015605927, + 0.02460244484245777, + 0.02348501794040203, + 0.061616457998752594, + 0.036518245935440063, + 0.029354361817240715, + 0.03420858085155487, + 0.02512373775243759, + 0.03206245228648186, + 0.018399853259325027, + 0.024954278022050858, + 0.036641597747802734, + 0.02812621183693409, + 0.021617969498038292, + 0.023649856448173523, + 0.0026123677380383015, + 0.0028781562577933073, + 0.0014038441004231572, + 0.002105357823893428, + 0.0028806154150515795, + 0.0032832154538482428, + 0.003666051896288991, + 0.0053511252626776695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.2977783679962158 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899461686611176 + ], + "opposite_logprob_tokens": [ + -2.5789945125579834 + ], + "logprob_reward": -0.07899461686611176, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899461686611176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11306490749120712, + 0.05597608909010887, + 0.036967676132917404, + 0.04407099634408951, + 0.04500553384423256, + 0.09386511892080307, + 0.13123945891857147, + 0.2645080089569092, + 0.18916702270507812, + 0.026135168969631195 + ], + "credit_map": [ + 0.013842216692864895, + 0.01175833772867918, + 0.010970418341457844, + 0.013110126368701458, + 0.010427839122712612, + 0.012026236392557621, + 0.02062220871448517, + 0.020307516679167747, + 0.00835365615785122, + 0.007445409428328276, + 0.009834378026425838, + 0.005502098705619574, + 0.003653504652902484, + 0.0069036525674164295, + 0.007365558296442032, + 0.006917832884937525, + 0.007544190622866154, + 0.004498733207583427, + 0.004207744263112545, + 0.003672342747449875, + 0.0037487968802452087, + 0.005023755133152008, + 0.003570146393030882, + 0.00470196595415473, + 0.008087452501058578, + 0.004635633900761604, + 0.00633267080411315, + 0.0033140857703983784, + 0.004234259016811848, + 0.0066743637435138226, + 0.002952874405309558, + 0.007839656434953213, + 0.006860711611807346, + 0.006108737085014582, + 0.006185307167470455, + 0.005368392448872328, + 0.004603099077939987, + 0.0044036502949893475, + 0.00587492948397994, + 0.0056007071398198605, + 0.012877569533884525, + 0.010198467411100864, + 0.01707850955426693, + 0.009151075035333633, + 0.011716796085238457, + 0.008741507306694984, + 0.011724729090929031, + 0.012376470491290092, + 0.019553374499082565, + 0.02118276245892048, + 0.012422880157828331, + 0.009702633135020733, + 0.022431615740060806, + 0.01599767431616783, + 0.01308384072035551, + 0.01686467044055462, + 0.024705251678824425, + 0.02768213115632534, + 0.015633774921298027, + 0.02116839960217476, + 0.0652960017323494, + 0.04458817094564438, + 0.023305751383304596, + 0.042128536850214005, + 0.020694328472018242, + 0.023065315559506416, + 0.017622513696551323, + 0.023254388943314552, + 0.02489602565765381, + 0.016602186486124992, + 0.026629840955138206, + 0.03640242666006088, + 0.0025762261357158422, + 0.005120707675814629, + 0.0016033472493290901, + 0.002383718267083168, + 0.0052823638543486595, + 0.003017720766365528, + 0.002649100963026285, + 0.0035019838251173496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.07899461686611176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635692600160837 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635692600160837, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635692600160837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07816296070814133, + 0.036567311733961105, + 0.02022533118724823, + 0.033430065959692, + 0.05651557072997093, + 0.07953808456659317, + 0.08456018567085266, + 0.30442482233047485, + 0.2620365619659424, + 0.04453910142183304 + ], + "credit_map": [ + 0.016011854633688927, + 0.008013153448700905, + 0.008777568116784096, + 0.008311480283737183, + 0.009289232082664967, + 0.010135937482118607, + 0.007680066395550966, + 0.009943673387169838, + 0.00452537601813674, + 0.005443400703370571, + 0.004894256591796875, + 0.0036718000192195177, + 0.003441460197791457, + 0.00463431142270565, + 0.005269522313028574, + 0.004687184002250433, + 0.0035476218909025192, + 0.0024594408459961414, + 0.0030430739279836416, + 0.0016558326315134764, + 0.002089096698909998, + 0.0029552748892456293, + 0.0026312307454645634, + 0.0018437601393088698, + 0.005576568655669689, + 0.004140864126384258, + 0.0059450408443808556, + 0.0025556664913892746, + 0.0033674179576337337, + 0.004002085886895657, + 0.0034993684384971857, + 0.004343056119978428, + 0.007259690202772617, + 0.00894074235111475, + 0.00638913456350565, + 0.005542726255953312, + 0.008638911880552769, + 0.009322737343609333, + 0.005201052408665419, + 0.005220574326813221, + 0.010565766133368015, + 0.011420427821576595, + 0.009678486734628677, + 0.007422957569360733, + 0.010474439710378647, + 0.014037509448826313, + 0.009816269390285015, + 0.006122229620814323, + 0.012711635790765285, + 0.016480131074786186, + 0.009071013890206814, + 0.006319599691778421, + 0.014325713738799095, + 0.009806964546442032, + 0.008989034220576286, + 0.006856090389192104, + 0.03380313143134117, + 0.04129280894994736, + 0.029118523001670837, + 0.026342637836933136, + 0.06540530920028687, + 0.03985783830285072, + 0.03515361621975899, + 0.03345094621181488, + 0.032180432230234146, + 0.03915568068623543, + 0.021374698728322983, + 0.03178485482931137, + 0.046460360288619995, + 0.03442918881773949, + 0.026177838444709778, + 0.0304734967648983, + 0.006124080624431372, + 0.005183886736631393, + 0.0024845630396157503, + 0.0036276355385780334, + 0.009378678165376186, + 0.005970694124698639, + 0.00565976882353425, + 0.0061097946017980576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.007635692600160837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579621195793152 + ], + "opposite_logprob_tokens": [ + -3.8955795764923096 + ], + "logprob_reward": -0.020579621195793152, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579621195793152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06662170588970184, + 0.02722768671810627, + 0.025311054661870003, + 0.03132534772157669, + 0.031514883041381836, + 0.08979517966508865, + 0.0682789534330368, + 0.35681238770484924, + 0.25553569197654724, + 0.04757711663842201 + ], + "credit_map": [ + 0.010545143857598305, + 0.006869559176266193, + 0.009922835044562817, + 0.007818183861672878, + 0.006143774371594191, + 0.007712610997259617, + 0.009710620157420635, + 0.00789897982031107, + 0.004705170169472694, + 0.004337899852544069, + 0.003596646012738347, + 0.0027814654167741537, + 0.0022671727929264307, + 0.003400300396606326, + 0.0035856696777045727, + 0.002553358906880021, + 0.003635171102359891, + 0.0044991192407906055, + 0.0030617841985076666, + 0.002035875106230378, + 0.002568095689639449, + 0.0033201989717781544, + 0.0027249236591160297, + 0.0034658832009881735, + 0.005415484309196472, + 0.005078404676169157, + 0.004653737414628267, + 0.0024523744359612465, + 0.002994758542627096, + 0.004073863849043846, + 0.003448198549449444, + 0.0032085194252431393, + 0.0037974119186401367, + 0.0036821323446929455, + 0.004029125906527042, + 0.003320354036986828, + 0.005582377314567566, + 0.004471568390727043, + 0.003722989233210683, + 0.002908922266215086, + 0.018745096400380135, + 0.013517534360289574, + 0.010205818340182304, + 0.0093527315184474, + 0.008491030894219875, + 0.011861898936331272, + 0.010665914043784142, + 0.0069551412016153336, + 0.009837329387664795, + 0.012278968468308449, + 0.008191067725419998, + 0.005774972029030323, + 0.008912955410778522, + 0.005958203691989183, + 0.010233554989099503, + 0.0070918952114880085, + 0.03726407140493393, + 0.04572250321507454, + 0.032159868627786636, + 0.029209675267338753, + 0.08088921755552292, + 0.04847649484872818, + 0.040039390325546265, + 0.04305114224553108, + 0.02983636036515236, + 0.040433138608932495, + 0.02253570221364498, + 0.030821984633803368, + 0.04339015856385231, + 0.03312268480658531, + 0.026024632155895233, + 0.029370998963713646, + 0.0061541227623820305, + 0.006483916658908129, + 0.002771493047475815, + 0.0038152113556861877, + 0.011120336130261421, + 0.007007535081356764, + 0.0036146556958556175, + 0.0066098403185606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.020579621195793152 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014185074716806412 + ], + "opposite_logprob_tokens": [ + -4.264184951782227 + ], + "logprob_reward": -0.014185074716806412, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014185074716806412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07733528316020966, + 0.0662870854139328, + 0.017554374411702156, + 0.02999104931950569, + 0.02131364680826664, + 0.07323525846004486, + 0.07205592095851898, + 0.31740692257881165, + 0.274722158908844, + 0.050098318606615067 + ], + "credit_map": [ + 0.013073619455099106, + 0.005946641322225332, + 0.01204307097941637, + 0.010251548141241074, + 0.007201489992439747, + 0.008757775649428368, + 0.005035011097788811, + 0.015026120468974113, + 0.013870268128812313, + 0.006773665081709623, + 0.009921827353537083, + 0.007608084939420223, + 0.006268724333494902, + 0.009869528003036976, + 0.005854452028870583, + 0.006120531819760799, + 0.0038051533047109842, + 0.0020982620771974325, + 0.0024077147245407104, + 0.0014300572220236063, + 0.0013518893392756581, + 0.002556969877332449, + 0.001628164085559547, + 0.0022761644795536995, + 0.005165649112313986, + 0.003238426987081766, + 0.0054186019115149975, + 0.0019884707871824503, + 0.002732464810833335, + 0.004139214754104614, + 0.0029467956628650427, + 0.0043614269234240055, + 0.002666447777301073, + 0.003258882323279977, + 0.0029237717390060425, + 0.002545783529058099, + 0.0030802306719124317, + 0.0023638156708329916, + 0.0022084354422986507, + 0.002266280585899949, + 0.012728913687169552, + 0.012239119969308376, + 0.009128670208156109, + 0.008616731502115726, + 0.006266123615205288, + 0.010229735635221004, + 0.00693257013335824, + 0.007093401160091162, + 0.01033718977123499, + 0.01380668580532074, + 0.008428973145782948, + 0.006158370058983564, + 0.009357083588838577, + 0.007092827465385199, + 0.009225964546203613, + 0.007648831699043512, + 0.0343531109392643, + 0.04580741003155708, + 0.02750147134065628, + 0.030402788892388344, + 0.06182963773608208, + 0.0391533225774765, + 0.035192351788282394, + 0.04316681623458862, + 0.032676372677087784, + 0.04338556155562401, + 0.02255067601799965, + 0.03294931352138519, + 0.04767204448580742, + 0.03530395030975342, + 0.02737872116267681, + 0.032805535942316055, + 0.00688336743041873, + 0.006095529068261385, + 0.0025733783841133118, + 0.004373845178633928, + 0.010862048715353012, + 0.00687568262219429, + 0.0049145519733428955, + 0.0075199115090072155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.014185074716806412 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an upward angle, the camera slowly pans right across the murky seafloor, revealing a giant octopus perched atop a patch of ancient volcanic rock covered in deep green algae. Its bulbous head is massive and heavy, its wrinkled and textured skin a shifting iridescence, transitioning from deep indigo to vibrant magenta, dotted with faint, star-like bioluminescent specks. One of its enormous, iridescent tentacle unfolds gracefully and powerfully, like a giant beast awakening. The neatly arranged white suction cups on its inner surface expand and contract in the water as it stretches, as if testing the tranquil waters surrounding it. A few faint rays of Jesus light penetrated through the sea surface, forming swaying beams of light in the deep blue water full of suspended particles, which just illuminated the octopus's outstretched limbs, highlighting the slippery luster and complex color changes of its skin. The occasional silvery schools of fish quickly passed by in the background, further contrasting the calm and composed posture of this giant beast. The whole scene was filled with a tranquil, mysterious and awe-inspiring atmosphere, showing the magnificence and wonder of deep-sea life.", + "seed": 910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_14/index_2405_seed_910_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": -0.0561766155064106, + "vlm_logprob": -0.0561766155064106, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.0561766155064106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21078000962734222 + ], + "opposite_logprob_tokens": [ + -1.7107800245285034 + ], + "logprob_reward": -0.21078000962734222, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21078000962734222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20980887115001678, + 0.06725236773490906, + 0.07377218455076218, + 0.1392294466495514, + 0.0832623541355133, + 0.05184582620859146, + 0.2145649492740631, + 0.03387199714779854, + 0.054187554866075516, + 0.07220444828271866 + ], + "credit_map": [ + 0.03196004033088684, + 0.028187304735183716, + 0.02876965142786503, + 0.019396193325519562, + 0.014006609097123146, + 0.04808536544442177, + 0.01648690178990364, + 0.02291679196059704, + 0.008195693604648113, + 0.012317393906414509, + 0.008069990202784538, + 0.007157153449952602, + 0.006192865781486034, + 0.009692528285086155, + 0.008540368638932705, + 0.007086370140314102, + 0.011230814270675182, + 0.006332987919449806, + 0.011695103719830513, + 0.007244572974741459, + 0.0073428587056696415, + 0.012514987029135227, + 0.00619127694517374, + 0.011219583451747894, + 0.03177858516573906, + 0.017657825723290443, + 0.013147256337106228, + 0.015345166437327862, + 0.008454366587102413, + 0.023669635877013206, + 0.013709656894207, + 0.015466959215700626, + 0.008697450160980225, + 0.016305679455399513, + 0.00797741487622261, + 0.009920640848577023, + 0.008894325233995914, + 0.011278566904366016, + 0.011697335168719292, + 0.008490946143865585, + 0.007501632906496525, + 0.0070379795506596565, + 0.008085310459136963, + 0.005528033711016178, + 0.004818144254386425, + 0.005196467973291874, + 0.006804629694670439, + 0.006873628590255976, + 0.023393729701638222, + 0.027523178607225418, + 0.018087005242705345, + 0.01943088136613369, + 0.019372057169675827, + 0.022967873141169548, + 0.03191031888127327, + 0.051879897713661194, + 0.0032631722278892994, + 0.004518562462180853, + 0.004577278159558773, + 0.003617445472627878, + 0.0034413011744618416, + 0.004137909971177578, + 0.0038454080931842327, + 0.006470921449363232, + 0.008653203025460243, + 0.008812935091555119, + 0.0042960685677826405, + 0.0065630157478153706, + 0.005634050816297531, + 0.006867816671729088, + 0.006674314849078655, + 0.006686150096356869, + 0.0067678699269890785, + 0.008733756840229034, + 0.00634722551330924, + 0.00672258622944355, + 0.008454589173197746, + 0.010029644705355167, + 0.006618218496441841, + 0.018530553206801414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.21078000962734222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029818108305335045 + ], + "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": [ + 0.12492125481367111, + 0.18683861196041107, + 0.11360685527324677, + 0.16097202897071838, + 0.031786322593688965, + 0.032857298851013184, + 0.06323378533124924, + 0.2319454550743103, + 0.031048065051436424, + 0.022790275514125824 + ], + "credit_map": [ + 0.02277626283466816, + 0.015578985214233398, + 0.017987845465540886, + 0.01497667282819748, + 0.013311920687556267, + 0.014229204505681992, + 0.011087797582149506, + 0.014972553588449955, + 0.030545687302947044, + 0.02221851609647274, + 0.02548958919942379, + 0.018890580162405968, + 0.02006317302584648, + 0.03330559283494949, + 0.017531560733914375, + 0.018793923780322075, + 0.01837713085114956, + 0.012724519707262516, + 0.017671318724751472, + 0.011139259673655033, + 0.011550791561603546, + 0.02022707089781761, + 0.008518243208527565, + 0.013398525305092335, + 0.037232961505651474, + 0.016426919028162956, + 0.019770003855228424, + 0.010464312508702278, + 0.011804626323282719, + 0.030840277671813965, + 0.010854026302695274, + 0.02357889711856842, + 0.0033261850476264954, + 0.007250846829265356, + 0.004002298694103956, + 0.0033170259557664394, + 0.0028719380497932434, + 0.0035523551050573587, + 0.003839602693915367, + 0.0036260723136365414, + 0.003730020485818386, + 0.004261413123458624, + 0.005736073013395071, + 0.0033186201471835375, + 0.003429221687838435, + 0.003186057787388563, + 0.004983973223716021, + 0.004211917985230684, + 0.005696502979844809, + 0.00781504437327385, + 0.006673090159893036, + 0.007141899783164263, + 0.006493141409009695, + 0.006398150697350502, + 0.011839126236736774, + 0.011176827363669872, + 0.03392188996076584, + 0.026312632486224174, + 0.02671733871102333, + 0.023940829560160637, + 0.018960097804665565, + 0.027164805680513382, + 0.01912413351237774, + 0.05580373480916023, + 0.004156260285526514, + 0.0030210628174245358, + 0.002405096311122179, + 0.0027409903705120087, + 0.0030815734062343836, + 0.003982482478022575, + 0.0031889069359749556, + 0.00847169291228056, + 0.0022977255284786224, + 0.002961821388453245, + 0.0022025122307240963, + 0.002149702748283744, + 0.0038140390533953905, + 0.002843995112925768, + 0.0023295260034501553, + 0.004190951120108366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.029818108305335045 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081226170063019 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081226170063019, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081226170063019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14121221005916595, + 0.14382098615169525, + 0.09909993410110474, + 0.09133534878492355, + 0.05416785925626755, + 0.05215281620621681, + 0.16268812119960785, + 0.0770668238401413, + 0.1529175192117691, + 0.025538405403494835 + ], + "credit_map": [ + 0.023282209411263466, + 0.01810244843363762, + 0.02302425168454647, + 0.017719391733407974, + 0.009961836971342564, + 0.019231446087360382, + 0.012728807516396046, + 0.017161808907985687, + 0.02409965544939041, + 0.01814574934542179, + 0.018499920144677162, + 0.009737296961247921, + 0.016142917796969414, + 0.02666025608778, + 0.013588478788733482, + 0.016946716234087944, + 0.016402294859290123, + 0.00977031234651804, + 0.013172130100429058, + 0.011421977542340755, + 0.008207576349377632, + 0.01730825938284397, + 0.010489975102245808, + 0.012327409349381924, + 0.01735210232436657, + 0.01586093194782734, + 0.008172974921762943, + 0.011045696213841438, + 0.006813645362854004, + 0.014152026735246181, + 0.008374900557100773, + 0.009563074447214603, + 0.0067733582109212875, + 0.009382643736898899, + 0.0064873346127569675, + 0.005723349284380674, + 0.006945488974452019, + 0.006582224275916815, + 0.006307726725935936, + 0.005965736228972673, + 0.008394421078264713, + 0.007957539521157742, + 0.0052705141715705395, + 0.00601160665974021, + 0.0050119273364543915, + 0.0055304900743067265, + 0.007037297356873751, + 0.006939018610864878, + 0.021887565031647682, + 0.017179911956191063, + 0.013972149230539799, + 0.016053389757871628, + 0.013886676169931889, + 0.016427159309387207, + 0.02535799704492092, + 0.037923265248537064, + 0.010914073325693607, + 0.011875691823661327, + 0.009856823831796646, + 0.009176371619105339, + 0.008009366691112518, + 0.0082785002887249, + 0.008492988534271717, + 0.010463004931807518, + 0.017597563564777374, + 0.023097284138202667, + 0.01147677842527628, + 0.017460573464632034, + 0.013802836649119854, + 0.02004488743841648, + 0.023297283798456192, + 0.02614031545817852, + 0.0043957713060081005, + 0.00353018706664443, + 0.001570038148202002, + 0.001669440302066505, + 0.0026339914184063673, + 0.0032188890036195517, + 0.003390243276953697, + 0.005129844415932894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.011081226170063019 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635440841317177 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635440841317177, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635440841317177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10927345603704453, + 0.09993059933185577, + 0.06774702668190002, + 0.05423226207494736, + 0.07882577925920486, + 0.08983428031206131, + 0.17787180840969086, + 0.07993977516889572, + 0.13705193996429443, + 0.10529309511184692 + ], + "credit_map": [ + 0.017782656475901604, + 0.015160158276557922, + 0.013072875328361988, + 0.012407701462507248, + 0.011516101658344269, + 0.01550513319671154, + 0.012686985544860363, + 0.01114184595644474, + 0.012483003549277782, + 0.016432175412774086, + 0.010434931144118309, + 0.011824309825897217, + 0.008493529632687569, + 0.014207471162080765, + 0.014981438405811787, + 0.011073735542595387, + 0.009084273129701614, + 0.007460324093699455, + 0.011160996742546558, + 0.008168073371052742, + 0.007557484321296215, + 0.007285884581506252, + 0.008063155226409435, + 0.008966836147010326, + 0.007826372049748898, + 0.007816906087100506, + 0.005423300433903933, + 0.008222861215472221, + 0.007169738411903381, + 0.007098844274878502, + 0.0062684668228030205, + 0.00440577557310462, + 0.009755212813615799, + 0.012098303996026516, + 0.009970061480998993, + 0.008632062934339046, + 0.007976286113262177, + 0.009062889032065868, + 0.009455984458327293, + 0.011874980293214321, + 0.009264358319342136, + 0.013737669214606285, + 0.011797355487942696, + 0.009372628293931484, + 0.0120540177449584, + 0.009071976877748966, + 0.01505235955119133, + 0.009483910165727139, + 0.02053992822766304, + 0.017518358305096626, + 0.01458071544766426, + 0.014191211201250553, + 0.013248014263808727, + 0.022857317700982094, + 0.01745743490755558, + 0.05747883766889572, + 0.013167696073651314, + 0.013402527198195457, + 0.008997738361358643, + 0.008445139974355698, + 0.008394859731197357, + 0.00911705195903778, + 0.006982398685067892, + 0.011432363651692867, + 0.02676534093916416, + 0.020871451124548912, + 0.006999734323471785, + 0.01207983773201704, + 0.012577913701534271, + 0.017865460366010666, + 0.017972709611058235, + 0.021919498220086098, + 0.014486747793853283, + 0.0251137875020504, + 0.0077611650340259075, + 0.008111619390547276, + 0.010529185645282269, + 0.01209032442420721, + 0.0160602405667305, + 0.011140020564198494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.008635440841317177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056829258799553 + ], + "opposite_logprob_tokens": [ + -3.895568370819092 + ], + "logprob_reward": -0.02056829258799553, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056829258799553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13152673840522766, + 0.3180284798145294, + 0.09582215547561646, + 0.049038954079151154, + 0.05136727914214134, + 0.05015572905540466, + 0.08635500073432922, + 0.12045743316411972, + 0.060691941529512405, + 0.036556344479322433 + ], + "credit_map": [ + 0.018648257479071617, + 0.016587253659963608, + 0.023750796914100647, + 0.012606288306415081, + 0.01213826984167099, + 0.022999759763479233, + 0.009501896798610687, + 0.015294210985302925, + 0.05572016164660454, + 0.0414014533162117, + 0.038599759340286255, + 0.025346938520669937, + 0.036091845482587814, + 0.06289215385913849, + 0.028812240809202194, + 0.029163895174860954, + 0.014693379402160645, + 0.009203469380736351, + 0.017056725919246674, + 0.011045833118259907, + 0.008721521124243736, + 0.013512472622096539, + 0.008872641250491142, + 0.01271610613912344, + 0.008953508920967579, + 0.005900308024138212, + 0.007143137510865927, + 0.0040915184654295444, + 0.004588537849485874, + 0.010277282446622849, + 0.0033687474206089973, + 0.004715912509709597, + 0.007303133606910706, + 0.008884498849511147, + 0.007244724314659834, + 0.005772117525339127, + 0.005172529257833958, + 0.00525832362473011, + 0.005982034839689732, + 0.005749916657805443, + 0.005786375608295202, + 0.005125307012349367, + 0.0065636844374239445, + 0.006384761072695255, + 0.004555811639875174, + 0.0053758095018565655, + 0.007712946273386478, + 0.008651030249893665, + 0.011870172806084156, + 0.007864682003855705, + 0.007195499259978533, + 0.006701917387545109, + 0.006598187610507011, + 0.011131583712995052, + 0.010452345944941044, + 0.024540605023503304, + 0.015998907387256622, + 0.01659029722213745, + 0.013436277396976948, + 0.010640980675816536, + 0.012297464534640312, + 0.014740011654794216, + 0.01031961478292942, + 0.026433879509568214, + 0.011105832643806934, + 0.0075287348590791225, + 0.003883176017552614, + 0.005226967390626669, + 0.0066529689356684685, + 0.007609642576426268, + 0.005295615177601576, + 0.01338900811970234, + 0.004972437862306833, + 0.005835792049765587, + 0.003003492020070553, + 0.003147454233840108, + 0.004218923859298229, + 0.004798804875463247, + 0.004592621233314276, + 0.005986820440739393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.02056829258799553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an upward angle, the camera slowly pans right across the murky seafloor, revealing a giant octopus perched atop a patch of ancient volcanic rock covered in deep green algae. Its bulbous head is massive and heavy, its wrinkled and textured skin a shifting iridescence, transitioning from deep indigo to vibrant magenta, dotted with faint, star-like bioluminescent specks. One of its enormous, iridescent tentacle unfolds gracefully and powerfully, like a giant beast awakening. The neatly arranged white suction cups on its inner surface expand and contract in the water as it stretches, as if testing the tranquil waters surrounding it. A few faint rays of Jesus light penetrated through the sea surface, forming swaying beams of light in the deep blue water full of suspended particles, which just illuminated the octopus's outstretched limbs, highlighting the slippery luster and complex color changes of its skin. The occasional silvery schools of fish quickly passed by in the background, further contrasting the calm and composed posture of this giant beast. The whole scene was filled with a tranquil, mysterious and awe-inspiring atmosphere, showing the magnificence and wonder of deep-sea life.", + "seed": 911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_14/index_2405_seed_911_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": -0.08163059530779719, + "vlm_logprob": -0.08163059530779719, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.08163059530779719, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24670973420143127 + ], + "opposite_logprob_tokens": [ + -1.6217097043991089 + ], + "logprob_reward": -0.24670973420143127, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24670973420143127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11323867738246918, + 0.058586545288562775, + 0.030470075085759163, + 0.042445845901966095, + 0.06607083231210709, + 0.15438833832740784, + 0.09501567482948303, + 0.14084763824939728, + 0.24353858828544617, + 0.05539786070585251 + ], + "credit_map": [ + 0.013665464706718922, + 0.01387545745819807, + 0.019337691366672516, + 0.009457476437091827, + 0.012817436829209328, + 0.013435606844723225, + 0.01038598082959652, + 0.0202635508030653, + 0.006411676295101643, + 0.005830659065395594, + 0.00741173792630434, + 0.006727706640958786, + 0.005907607264816761, + 0.005896897055208683, + 0.007579676806926727, + 0.012820572592318058, + 0.003927531186491251, + 0.004028880503028631, + 0.0030950389336794615, + 0.0031546384561806917, + 0.004793111700564623, + 0.003711530938744545, + 0.003577233524993062, + 0.00418210681527853, + 0.004786169622093439, + 0.0055920640006661415, + 0.0027367426082491875, + 0.0039183408953249454, + 0.007241461891680956, + 0.005078856833279133, + 0.005193961318582296, + 0.007898244075477123, + 0.007973510771989822, + 0.007185994181782007, + 0.007152067497372627, + 0.006069404538720846, + 0.009789089672267437, + 0.008078609593212605, + 0.007748024072498083, + 0.0120741231366992, + 0.012746124528348446, + 0.01998876966536045, + 0.015905836597085, + 0.017788857221603394, + 0.02421271614730358, + 0.018781229853630066, + 0.026699062436819077, + 0.018265720456838608, + 0.007810000330209732, + 0.009839673526585102, + 0.008993953466415405, + 0.009014913812279701, + 0.01690421812236309, + 0.011511094868183136, + 0.018164649605751038, + 0.012777159921824932, + 0.019873255863785744, + 0.021183790639042854, + 0.015345470979809761, + 0.016197528690099716, + 0.014547555707395077, + 0.017007675021886826, + 0.016844645142555237, + 0.019847707822918892, + 0.023918015882372856, + 0.03470959514379501, + 0.03424476832151413, + 0.027907205745577812, + 0.03540675714612007, + 0.032416168600320816, + 0.027203768491744995, + 0.027732286602258682, + 0.005064298864454031, + 0.007520024199038744, + 0.006951457355171442, + 0.005811250302940607, + 0.0066251675598323345, + 0.006711656227707863, + 0.010555394925177097, + 0.006158605217933655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.24670973420143127 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129465103149414 + ], + "opposite_logprob_tokens": [ + -2.2379465103149414 + ], + "logprob_reward": -0.1129465103149414, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129465103149414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1493646651506424, + 0.09150952845811844, + 0.21940338611602783, + 0.17247098684310913, + 0.034753505140542984, + 0.06903521716594696, + 0.05529163032770157, + 0.09744763374328613, + 0.02592877671122551, + 0.08479469269514084 + ], + "credit_map": [ + 0.023505156859755516, + 0.015185557305812836, + 0.022797966375947, + 0.014030653983354568, + 0.02039027027785778, + 0.020715638995170593, + 0.014633840881288052, + 0.01810557395219803, + 0.010550067760050297, + 0.012035690248012543, + 0.00959213450551033, + 0.012628738768398762, + 0.015624228864908218, + 0.010251101106405258, + 0.010927827097475529, + 0.009899740107357502, + 0.03551663085818291, + 0.022970467805862427, + 0.015117247588932514, + 0.020907238125801086, + 0.03655247390270233, + 0.026813676580786705, + 0.022633155807852745, + 0.03889249637722969, + 0.018043305724859238, + 0.02097446657717228, + 0.016575146466493607, + 0.013149015605449677, + 0.02421349659562111, + 0.019193163141608238, + 0.025132998824119568, + 0.03518940880894661, + 0.004988927859812975, + 0.00471781101077795, + 0.0043663871474564075, + 0.002309845294803381, + 0.003656104439869523, + 0.004060938023030758, + 0.0048875086940824986, + 0.0057659815065562725, + 0.008022145368158817, + 0.010357746854424477, + 0.010851495899260044, + 0.007325306069105864, + 0.0069140722043812275, + 0.008362852968275547, + 0.009463232941925526, + 0.007738363929092884, + 0.0056326789781451225, + 0.006978973280638456, + 0.009811211377382278, + 0.006186637096107006, + 0.006099266465753317, + 0.007808513008058071, + 0.0061968606896698475, + 0.006577490828931332, + 0.013504869304597378, + 0.014342926442623138, + 0.009441085159778595, + 0.013188706710934639, + 0.011650511994957924, + 0.011973110027611256, + 0.010171368718147278, + 0.013175052590668201, + 0.003267919411882758, + 0.004490661900490522, + 0.0034136450849473476, + 0.002536922227591276, + 0.003029456827789545, + 0.00397027051076293, + 0.002326427260413766, + 0.00289347255602479, + 0.012456758879125118, + 0.014607089571654797, + 0.008454244583845139, + 0.008097177371382713, + 0.0107661047950387, + 0.010569727048277855, + 0.011352187022566795, + 0.008491403423249722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.1129465103149414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085116304457188 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011085116304457188, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085116304457188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1312946379184723, + 0.036153294146060944, + 0.04816516488790512, + 0.058790840208530426, + 0.03653028979897499, + 0.12514738738536835, + 0.036301206797361374, + 0.3114360272884369, + 0.11661908030509949, + 0.09956206381320953 + ], + "credit_map": [ + 0.016757238656282425, + 0.016826948150992393, + 0.02075522392988205, + 0.008668586611747742, + 0.016120191663503647, + 0.015728922560811043, + 0.015065168961882591, + 0.021372361108660698, + 0.004809498321264982, + 0.004058989230543375, + 0.004196925554424524, + 0.003528960281983018, + 0.0060251634567976, + 0.004364427179098129, + 0.004200279712677002, + 0.004969052970409393, + 0.006075319834053516, + 0.005986378062516451, + 0.005085946060717106, + 0.004601713735610247, + 0.007499253377318382, + 0.005524737294763327, + 0.0074522774666547775, + 0.005939539521932602, + 0.010129346512258053, + 0.006929120980203152, + 0.004162458702921867, + 0.0046595134772360325, + 0.00765041820704937, + 0.007154261227697134, + 0.0067641641944646835, + 0.01134155597537756, + 0.005124977324157953, + 0.005026878789067268, + 0.0036804918199777603, + 0.0034960107877850533, + 0.004936631303280592, + 0.004215034190565348, + 0.004068284295499325, + 0.005981979891657829, + 0.009674587287008762, + 0.015665821731090546, + 0.013210910372436047, + 0.014873874373733997, + 0.01948525197803974, + 0.014238722622394562, + 0.023330481722950935, + 0.014667736366391182, + 0.004594390280544758, + 0.005750835407525301, + 0.004863719921559095, + 0.002679330063983798, + 0.005068608094006777, + 0.004633622709661722, + 0.0037300358526408672, + 0.004980665165930986, + 0.05780233070254326, + 0.061652444303035736, + 0.028244808316230774, + 0.03648751229047775, + 0.03099006786942482, + 0.0415445975959301, + 0.02198183536529541, + 0.03273243457078934, + 0.013443474657833576, + 0.01732700876891613, + 0.017333533614873886, + 0.012859315611422062, + 0.0144188879057765, + 0.015971714630723, + 0.01150432601571083, + 0.0137608228251338, + 0.01131658535450697, + 0.017953479662537575, + 0.012121760286390781, + 0.009936819784343243, + 0.01162903755903244, + 0.01264319196343422, + 0.01361118070781231, + 0.010350005701184273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.011085116304457188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627411279827356 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627411279827356, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627411279827356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0465753935277462, + 0.050799839198589325, + 0.02063853107392788, + 0.038466740399599075, + 0.054992031306028366, + 0.09116671234369278, + 0.0738026425242424, + 0.10154102742671967, + 0.20503073930740356, + 0.3169863522052765 + ], + "credit_map": [ + 0.006359599996358156, + 0.006817281246185303, + 0.005422019399702549, + 0.004882297944277525, + 0.007128213532269001, + 0.004981053527444601, + 0.005978924222290516, + 0.005006003193557262, + 0.005745766218751669, + 0.0064636459574103355, + 0.006437534932047129, + 0.005859681870788336, + 0.00837612058967352, + 0.005366202909499407, + 0.006508184131234884, + 0.0060427007265388966, + 0.004192981403321028, + 0.0026799612678587437, + 0.002243029186502099, + 0.0017880917293950915, + 0.002613471355289221, + 0.003230791073292494, + 0.001800556550733745, + 0.002089648274704814, + 0.005522415041923523, + 0.004258357919752598, + 0.0037743605207651854, + 0.004060007631778717, + 0.00477222865447402, + 0.004102669190615416, + 0.006319487001746893, + 0.005657216068357229, + 0.007061800453811884, + 0.011205489747226238, + 0.006983538623899221, + 0.004643143620342016, + 0.006602524779736996, + 0.0061682225205004215, + 0.005967378616333008, + 0.0063599348068237305, + 0.008085113018751144, + 0.010294154286384583, + 0.013800570741295815, + 0.014209363609552383, + 0.012738188728690147, + 0.010154737159609795, + 0.011546898633241653, + 0.010337681509554386, + 0.006798881571739912, + 0.007620643358677626, + 0.00906146690249443, + 0.00805033277720213, + 0.012095167301595211, + 0.009229692630469799, + 0.012178146280348301, + 0.008768309839069843, + 0.02046739123761654, + 0.019099118188023567, + 0.00918375700712204, + 0.012061295099556446, + 0.010354915633797646, + 0.014170201495289803, + 0.005955501459538937, + 0.010248834267258644, + 0.025603819638490677, + 0.027242567390203476, + 0.024826601147651672, + 0.020968768745660782, + 0.03393762186169624, + 0.0323360376060009, + 0.019771184772253036, + 0.020344143733382225, + 0.03659190237522125, + 0.05885356664657593, + 0.03957739099860191, + 0.03291122615337372, + 0.03447245806455612, + 0.04098126292228699, + 0.04102610796689987, + 0.0325724296271801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.007627411279827356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an upward angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant octopus visible perched on ancient volcanic rock covered in deep green algae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus's skin display a shifting iridescence transitioning from deep indigo to vibrant magenta with faint bioluminescent specks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the octopus's enormous tentacles unfolding gracefully while its white suction cups expand and contract?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are swaying beams of light penetrating the deep blue water full of suspended particles to illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029784204438328743 + ], + "opposite_logprob_tokens": [ + -3.5297842025756836 + ], + "logprob_reward": -0.029784204438328743, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029784204438328743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20089957118034363, + 0.20544947683811188, + 0.08594886213541031, + 0.05167999118566513, + 0.054967258125543594, + 0.1309623271226883, + 0.056941743940114975, + 0.08519236743450165, + 0.04831339046359062, + 0.07964510470628738 + ], + "credit_map": [ + 0.03199179843068123, + 0.023317089304327965, + 0.03208835795521736, + 0.01927649788558483, + 0.026263179257512093, + 0.031097227707505226, + 0.018889926373958588, + 0.017975473776459694, + 0.02915292978286743, + 0.0184030681848526, + 0.02809702791273594, + 0.01873297244310379, + 0.016216564923524857, + 0.025740832090377808, + 0.021205080673098564, + 0.047900982201099396, + 0.015917489305138588, + 0.009059488773345947, + 0.01060398854315281, + 0.007410020567476749, + 0.007782530039548874, + 0.010953014716506004, + 0.00985279493033886, + 0.014369522221386433, + 0.008612826466560364, + 0.007131272926926613, + 0.004317810758948326, + 0.0038170607294887304, + 0.007101801689714193, + 0.0068521504290401936, + 0.005917140282690525, + 0.007929923012852669, + 0.009096840396523476, + 0.00691200839355588, + 0.005384750664234161, + 0.004603294655680656, + 0.006547935307025909, + 0.006715868134051561, + 0.007127269171178341, + 0.008579284884035587, + 0.01292130071669817, + 0.013011442497372627, + 0.022883210331201553, + 0.018350854516029358, + 0.01597018353641033, + 0.01526778843253851, + 0.017477519810199738, + 0.015080024488270283, + 0.007236705161631107, + 0.006550906226038933, + 0.0071814763359725475, + 0.003724733367562294, + 0.01005953922867775, + 0.008039805106818676, + 0.00801566056907177, + 0.0061329081654548645, + 0.009533329866826534, + 0.01609514281153679, + 0.01004964578896761, + 0.00885189138352871, + 0.009353908710181713, + 0.010407832451164722, + 0.007641386706382036, + 0.0132592199370265, + 0.005797787569463253, + 0.008383270353078842, + 0.006504326593130827, + 0.0047504534013569355, + 0.005595844238996506, + 0.006252510938793421, + 0.005491957068443298, + 0.005537231918424368, + 0.011966366320848465, + 0.013255455531179905, + 0.010145521722733974, + 0.005896065849810839, + 0.011961225420236588, + 0.012907767668366432, + 0.006102637387812138, + 0.00741004990413785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.029784204438328743 + } + ], + "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/0000052/group_15/index_tensor([7043, 7043])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_15/index_tensor([7043, 7043])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a574513503e2ff34e441737fdc55950564591c65 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_15/index_tensor([7043, 7043])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely close, still photograph is tightly focused on a car's interior rearview mirror, encased in a matte black plastic casing. The faint ambient glow from the unseen dashboard barely captures the edges of the mirror's frame and the black fabric of the car's roof above. The entire drama unfolds within the mirror's glass surface, which reflects the nighttime multi-lane highway, rapidly receding into a pitch-black void. The bright white headlights of the following car in the distance are initially captured as distinct balls of light, but they quickly expand into thin, distorted streaks, their glow smeared on the glass by the car's immense speed. Then, these streaks of light rapidly shrink and dim, from vibrant lines to faint, isolated points of light that flicker unsteadily against the overwhelming darkness before vanishing entirely. The painted lane dividers on the asphalt appear to converge at an accelerating pace, swallowed by vanishing points as the car surges forward, creating a powerful and melancholic vision of departure and irreversible distance, immersed in the cool colors of deep blue and absolute black.", + "seed": 8517, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_15/index_7043_seed_8517_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": -0.06953472709283233, + "vlm_logprob": -0.06953472709283233, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.06953472709283233, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012734993360936642 + ], + "opposite_logprob_tokens": [ + -5.512734889984131 + ], + "logprob_reward": -0.012734993360936642, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.012734993360936642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17364446818828583, + 0.21514466404914856, + 0.11755289882421494, + 0.0373968705534935, + 0.05238097161054611, + 0.04923485592007637, + 0.06976888328790665, + 0.06690903007984161, + 0.10732349753379822, + 0.11064393073320389 + ], + "credit_map": [ + 0.015792999416589737, + 0.011952299624681473, + 0.02244907058775425, + 0.060732364654541016, + 0.016713788732886314, + 0.013026827946305275, + 0.012855196371674538, + 0.02012191154062748, + 0.04070629924535751, + 0.01416079606860876, + 0.03801817074418068, + 0.01510376576334238, + 0.04489217326045036, + 0.020899318158626556, + 0.022266048938035965, + 0.019098082557320595, + 0.014564537443220615, + 0.009421329014003277, + 0.00880969688296318, + 0.008837290108203888, + 0.015400486998260021, + 0.04028083384037018, + 0.012225390411913395, + 0.008013329468667507, + 0.003841777564957738, + 0.0037755845114588737, + 0.004510240629315376, + 0.005802517756819725, + 0.0043575051240623, + 0.003502716077491641, + 0.009399639442563057, + 0.002206890843808651, + 0.007335868198424578, + 0.006888806354254484, + 0.0060607027262449265, + 0.006300316192209721, + 0.007494964171200991, + 0.00592075614258647, + 0.008462187834084034, + 0.003917370922863483, + 0.005474914330989122, + 0.0066541824489831924, + 0.0051786391995847225, + 0.004456585738807917, + 0.006015442311763763, + 0.009944463148713112, + 0.00588248809799552, + 0.005628138314932585, + 0.00544397160410881, + 0.009026706218719482, + 0.0056657493114471436, + 0.010576633736491203, + 0.014337118715047836, + 0.01126512698829174, + 0.0079566715285182, + 0.005496907513588667, + 0.006507773883640766, + 0.008652672171592712, + 0.004519771784543991, + 0.007549329195171595, + 0.017466669902205467, + 0.011839262209832668, + 0.006315429694950581, + 0.004058120772242546, + 0.009299089200794697, + 0.011785322800278664, + 0.01152693573385477, + 0.013143240474164486, + 0.02367127314209938, + 0.02089536562561989, + 0.006886770483106375, + 0.010115492157638073, + 0.007739323657006025, + 0.011110793799161911, + 0.009563232772052288, + 0.00668503250926733, + 0.029816102236509323, + 0.02532283402979374, + 0.00834103673696518, + 0.012065567076206207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.012734993360936642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349563151597977 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349563151597977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349563151597977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.246049702167511, + 0.18605272471904755, + 0.10122400522232056, + 0.022322891280055046, + 0.05272310972213745, + 0.08419669419527054, + 0.045224882662296295, + 0.05939473211765289, + 0.1005813479423523, + 0.10222990810871124 + ], + "credit_map": [ + 0.022942405194044113, + 0.023638704791665077, + 0.029672542586922646, + 0.058554958552122116, + 0.028047146275639534, + 0.028379565104842186, + 0.016872573643922806, + 0.037941817194223404, + 0.026837309822440147, + 0.008438934572041035, + 0.07006391137838364, + 0.016520807519555092, + 0.017025325447320938, + 0.012613808736205101, + 0.01640799082815647, + 0.0181446373462677, + 0.008887981995940208, + 0.010579396039247513, + 0.006891786120831966, + 0.00551879545673728, + 0.012480499222874641, + 0.03451186791062355, + 0.014167393557727337, + 0.008186286315321922, + 0.0028721848502755165, + 0.002974000060930848, + 0.0021889929194003344, + 0.0036104347091168165, + 0.002667137188836932, + 0.002275787992402911, + 0.004020887892693281, + 0.0017134675290435553, + 0.008336139842867851, + 0.005438059568405151, + 0.006916442420333624, + 0.009471604600548744, + 0.004471002612262964, + 0.007075963541865349, + 0.005689991172403097, + 0.005323909223079681, + 0.006538399029523134, + 0.00984297413378954, + 0.0085895461961627, + 0.008729503490030766, + 0.009147491306066513, + 0.02519872412085533, + 0.009413991123437881, + 0.006736069917678833, + 0.0040577673353254795, + 0.007072107866406441, + 0.00563975190743804, + 0.006435244344174862, + 0.00785952340811491, + 0.0055739013478159904, + 0.004340140148997307, + 0.004246450494974852, + 0.005346750840544701, + 0.0072148689068853855, + 0.004728290252387524, + 0.00653348071500659, + 0.015506108291447163, + 0.01062284130603075, + 0.005855991039425135, + 0.0035864117089658976, + 0.00675147445872426, + 0.009187440387904644, + 0.009751771576702595, + 0.014905326999723911, + 0.018246112391352654, + 0.017641153186559677, + 0.009295904077589512, + 0.014802176505327225, + 0.0067619262263178825, + 0.011575351469218731, + 0.011876733973622322, + 0.009308721870183945, + 0.019882163032889366, + 0.021545300260186195, + 0.010414314456284046, + 0.010865404270589352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.026349563151597977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156273245811462 + ], + "opposite_logprob_tokens": [ + -1.701562762260437 + ], + "logprob_reward": -0.20156273245811462, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156273245811462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14725323021411896, + 0.11981139332056046, + 0.07018430531024933, + 0.026874709874391556, + 0.08797203749418259, + 0.04352327808737755, + 0.050036992877721786, + 0.03274274244904518, + 0.22388218343257904, + 0.19771914184093475 + ], + "credit_map": [ + 0.013302787207067013, + 0.009915932081639767, + 0.02035789005458355, + 0.0499148815870285, + 0.011824289336800575, + 0.009503175504505634, + 0.013331410475075245, + 0.019102856516838074, + 0.011651340872049332, + 0.0073716603219509125, + 0.026522483676671982, + 0.013715931214392185, + 0.02948642522096634, + 0.006975431926548481, + 0.014494935050606728, + 0.009593183174729347, + 0.010307970456779003, + 0.008290483616292477, + 0.005693218670785427, + 0.0040787216275930405, + 0.009699488990008831, + 0.020635021850466728, + 0.007014983333647251, + 0.004464418161660433, + 0.0033248947001993656, + 0.0030946305487304926, + 0.002803280483931303, + 0.004792686551809311, + 0.0026056221686303616, + 0.002214300911873579, + 0.005517757963389158, + 0.002521534450352192, + 0.009410514496266842, + 0.009489728137850761, + 0.00864648912101984, + 0.015460354276001453, + 0.011889847926795483, + 0.010681361891329288, + 0.013541625812649727, + 0.008852112106978893, + 0.0040230401791632175, + 0.005253423936665058, + 0.004849471617490053, + 0.006424375809729099, + 0.00411727512255311, + 0.008979341946542263, + 0.005550895817577839, + 0.004325455520302057, + 0.0049660238437354565, + 0.007472600322216749, + 0.0061207362450659275, + 0.00631908280774951, + 0.007840381003916264, + 0.006922489497810602, + 0.006102320738136768, + 0.004293358884751797, + 0.0037645860575139523, + 0.003335552755743265, + 0.0040615941397845745, + 0.0059167128056287766, + 0.0047131492756307125, + 0.005209657363593578, + 0.0030069039203226566, + 0.002734588459134102, + 0.01485452614724636, + 0.020451493561267853, + 0.02184060774743557, + 0.030051978304982185, + 0.05056284740567207, + 0.045733969658613205, + 0.0143039645627141, + 0.026082808151841164, + 0.015704065561294556, + 0.019667547196149826, + 0.025683270767331123, + 0.022641276940703392, + 0.035389501601457596, + 0.03588376194238663, + 0.02029012143611908, + 0.022459613159298897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.20156273245811462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029134154319763 + ], + "opposite_logprob_tokens": [ + -2.3502912521362305 + ], + "logprob_reward": -0.10029134154319763, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029134154319763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044747915118932724, + 0.25897490978240967, + 0.05184901878237724, + 0.02926522307097912, + 0.025402676314115524, + 0.1516772210597992, + 0.045464858412742615, + 0.03298602253198624, + 0.21706414222717285, + 0.14256802201271057 + ], + "credit_map": [ + 0.005087301600724459, + 0.0036700069904327393, + 0.004930485505610704, + 0.012872478924691677, + 0.0036831728648394346, + 0.003992603626102209, + 0.005506428424268961, + 0.005005436483770609, + 0.029159752652049065, + 0.00853643286973238, + 0.10282012075185776, + 0.015060756355524063, + 0.05121730640530586, + 0.012618429027497768, + 0.01329785492271185, + 0.026264240965247154, + 0.006176861468702555, + 0.008673157542943954, + 0.003987706732004881, + 0.005653479602187872, + 0.0074090538546442986, + 0.008207838051021099, + 0.009070727974176407, + 0.0026701930910348892, + 0.003986226860433817, + 0.003412865800783038, + 0.0028451087418943644, + 0.004503266885876656, + 0.003922838252037764, + 0.0032063296530395746, + 0.004771851468831301, + 0.0026167347095906734, + 0.0036550869699567556, + 0.003252476453781128, + 0.00288914586417377, + 0.004049771931022406, + 0.0034354908857494593, + 0.002853742800652981, + 0.0033537650015205145, + 0.001913192798383534, + 0.011533604003489017, + 0.01644613780081272, + 0.014391585253179073, + 0.01659265160560608, + 0.015184416435658932, + 0.044200971722602844, + 0.02025819756090641, + 0.013069656677544117, + 0.004062661435455084, + 0.006500688847154379, + 0.004572708159685135, + 0.005701659247279167, + 0.008939573541283607, + 0.006597318220883608, + 0.0047034225426614285, + 0.004386823624372482, + 0.0036449593026190996, + 0.005872438661754131, + 0.0034160155337303877, + 0.0056952983140945435, + 0.004614564590156078, + 0.004493005573749542, + 0.0027135631535202265, + 0.002536176471039653, + 0.012721161358058453, + 0.01827818714082241, + 0.020922251045703888, + 0.022746894508600235, + 0.05664033815264702, + 0.05080220103263855, + 0.013307776302099228, + 0.021645335480570793, + 0.009042494930326939, + 0.010553069412708282, + 0.0136043606325984, + 0.030901892110705376, + 0.022045617923140526, + 0.020292185246944427, + 0.01642131805419922, + 0.019707074388861656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.10029134154319763 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673500495031476 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.00673500495031476, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673500495031476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18974387645721436, + 0.17299175262451172, + 0.11285696923732758, + 0.03525848314166069, + 0.038839563727378845, + 0.07924888283014297, + 0.14245139062404633, + 0.0362350158393383, + 0.06698225438594818, + 0.12539182603359222 + ], + "credit_map": [ + 0.01578160934150219, + 0.01744071952998638, + 0.02933770790696144, + 0.056893061846494675, + 0.02101367525756359, + 0.013709642924368382, + 0.01656760834157467, + 0.018999852240085602, + 0.0321987047791481, + 0.007980283349752426, + 0.06931259483098984, + 0.009472839534282684, + 0.016819357872009277, + 0.01180704403668642, + 0.011043627746403217, + 0.01435729768127203, + 0.011888407170772552, + 0.00806023832410574, + 0.007254841737449169, + 0.005320678930729628, + 0.016801664605736732, + 0.043804753571748734, + 0.01233762875199318, + 0.007388752419501543, + 0.004337224643677473, + 0.0038962732069194317, + 0.003887936705723405, + 0.004029996693134308, + 0.00604945607483387, + 0.00575379328802228, + 0.0048782904632389545, + 0.0024255074094980955, + 0.004003662616014481, + 0.003973617218434811, + 0.004122364334762096, + 0.005749614909291267, + 0.00621144101023674, + 0.004866440314799547, + 0.006547120399773121, + 0.003365302924066782, + 0.009152572602033615, + 0.008127973414957523, + 0.008752932772040367, + 0.009984100237488747, + 0.010443413630127907, + 0.010888967663049698, + 0.014184230007231236, + 0.007714695297181606, + 0.009348351508378983, + 0.01845432072877884, + 0.011937378905713558, + 0.023096414282917976, + 0.03444337844848633, + 0.02250565215945244, + 0.014221766963601112, + 0.008444132283329964, + 0.0024430931080132723, + 0.004778634756803513, + 0.003486420027911663, + 0.0057858871296048164, + 0.00618433253839612, + 0.006360978353768587, + 0.003919682465493679, + 0.003275986760854721, + 0.003026201855391264, + 0.005141971167176962, + 0.005291275214403868, + 0.007892518304288387, + 0.017000673338770866, + 0.01615455560386181, + 0.005233291536569595, + 0.007241772022098303, + 0.009303241036832333, + 0.012072226963937283, + 0.011364557780325413, + 0.01732911914587021, + 0.029634203761816025, + 0.024054918438196182, + 0.011812780983746052, + 0.009820781648159027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.00673500495031476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, still photograph is tightly focused on a car's interior rearview mirror, encased in a matte black plastic casing. The faint ambient glow from the unseen dashboard barely captures the edges of the mirror's frame and the black fabric of the car's roof above. The entire drama unfolds within the mirror's glass surface, which reflects the nighttime multi-lane highway, rapidly receding into a pitch-black void. The bright white headlights of the following car in the distance are initially captured as distinct balls of light, but they quickly expand into thin, distorted streaks, their glow smeared on the glass by the car's immense speed. Then, these streaks of light rapidly shrink and dim, from vibrant lines to faint, isolated points of light that flicker unsteadily against the overwhelming darkness before vanishing entirely. The painted lane dividers on the asphalt appear to converge at an accelerating pace, swallowed by vanishing points as the car surges forward, creating a powerful and melancholic vision of departure and irreversible distance, immersed in the cool colors of deep blue and absolute black.", + "seed": 8518, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_15/index_7043_seed_8518_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": -0.2442143182270229, + "vlm_logprob": -0.2442143182270229, + "vlm_yes_no_margin": 2.675000011920929, + "vlm_token_credit": -0.2442143182270229, + "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 a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3517691493034363 + ], + "opposite_logprob_tokens": [ + -1.226769208908081 + ], + "logprob_reward": -0.3517691493034363, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3517691493034363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1495155394077301, + 0.1066623106598854, + 0.05954154580831528, + 0.08269219845533371, + 0.03419821336865425, + 0.07732059806585312, + 0.2477351427078247, + 0.1335868537425995, + 0.05508441850543022, + 0.05366325378417969 + ], + "credit_map": [ + 0.01365826465189457, + 0.014035580679774284, + 0.025332136079669, + 0.013838921673595905, + 0.015831349417567253, + 0.026533178985118866, + 0.013999806717038155, + 0.026286281645298004, + 0.011789650656282902, + 0.009517972357571125, + 0.02404852770268917, + 0.017268775030970573, + 0.007281958125531673, + 0.0135817751288414, + 0.008862973190844059, + 0.014310669153928757, + 0.011966359801590443, + 0.004194955807179213, + 0.012003857642412186, + 0.007433039601892233, + 0.005639643874019384, + 0.0066556851379573345, + 0.004832686856389046, + 0.006815311033278704, + 0.005563818383961916, + 0.0034052638802677393, + 0.00881995353847742, + 0.005865040235221386, + 0.005461752414703369, + 0.020056329667568207, + 0.010880030691623688, + 0.02263999730348587, + 0.0036239821929484606, + 0.0034814621321856976, + 0.006866014562547207, + 0.004232012666761875, + 0.0027690399438142776, + 0.004523367155343294, + 0.0039635938592255116, + 0.004738739226013422, + 0.00992768444120884, + 0.008637571707367897, + 0.008916287682950497, + 0.007878782227635384, + 0.011256025172770023, + 0.014881160110235214, + 0.006309063173830509, + 0.009514016099274158, + 0.026954855769872665, + 0.04333813115954399, + 0.021263305097818375, + 0.02414017543196678, + 0.042867641896009445, + 0.03227990120649338, + 0.026755020022392273, + 0.030136100947856903, + 0.018937474116683006, + 0.014470808207988739, + 0.010651873424649239, + 0.01436394453048706, + 0.02451619692146778, + 0.026601357385516167, + 0.009201422333717346, + 0.014843755401670933, + 0.005723654292523861, + 0.0057444521225988865, + 0.004380553029477596, + 0.005727166775614023, + 0.009912756271660328, + 0.01255626417696476, + 0.0033994445111602545, + 0.007640128489583731, + 0.006539151072502136, + 0.004845907911658287, + 0.004282244946807623, + 0.00630659656599164, + 0.009049680083990097, + 0.012798961251974106, + 0.0030958715360611677, + 0.006744834594428539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.3517691493034363 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961612798273563 + ], + "opposite_logprob_tokens": [ + -5.1309614181518555 + ], + "logprob_reward": -0.005961612798273563, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961612798273563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11252182722091675, + 0.10894251614809036, + 0.06895805895328522, + 0.07546265423297882, + 0.043304502964019775, + 0.06888376921415329, + 0.12309636920690536, + 0.0870753824710846, + 0.11784543842077255, + 0.19390949606895447 + ], + "credit_map": [ + 0.014091679826378822, + 0.00863770954310894, + 0.01938585937023163, + 0.013925283215939999, + 0.010572954080998898, + 0.018648415803909302, + 0.010939653031527996, + 0.01632027141749859, + 0.011642406694591045, + 0.012270231731235981, + 0.020624052733182907, + 0.01734774187207222, + 0.01077495887875557, + 0.012291924096643925, + 0.009928250685334206, + 0.014062943868339062, + 0.0078296959400177, + 0.004884590394794941, + 0.01252291165292263, + 0.01095493882894516, + 0.007211065385490656, + 0.008745822124183178, + 0.006353939883410931, + 0.010455097071826458, + 0.007618866395205259, + 0.005433501675724983, + 0.007963969372212887, + 0.009768049232661724, + 0.004922050517052412, + 0.015773914754390717, + 0.006606787443161011, + 0.017375510185956955, + 0.005118272267282009, + 0.0025997029151767492, + 0.011567219160497189, + 0.0038038084749132395, + 0.003061659401282668, + 0.005199026316404343, + 0.0053075565956532955, + 0.0066472552716732025, + 0.008490392006933689, + 0.0048630572855472565, + 0.012496978975832462, + 0.0073249503038823605, + 0.005298123229295015, + 0.009271686896681786, + 0.010393528267741203, + 0.010745051316916943, + 0.01529094111174345, + 0.023432934656739235, + 0.009651371277868748, + 0.009215355850756168, + 0.02601015567779541, + 0.017910363152623177, + 0.008328043855726719, + 0.013257204554975033, + 0.012960046529769897, + 0.00993250124156475, + 0.008219065144658089, + 0.009592062793672085, + 0.016994161531329155, + 0.01434650644659996, + 0.006208688486367464, + 0.008822347037494183, + 0.015513300895690918, + 0.012413824908435345, + 0.012456362135708332, + 0.013291248120367527, + 0.016971200704574585, + 0.022358031943440437, + 0.010217098519206047, + 0.014624373987317085, + 0.020492924377322197, + 0.018595604225993156, + 0.014849754050374031, + 0.021005019545555115, + 0.0400354377925396, + 0.03567460924386978, + 0.02000364102423191, + 0.02325250208377838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.005961612798273563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7577891945838928 + ], + "opposite_logprob_tokens": [ + -0.6327891945838928 + ], + "logprob_reward": -0.7577891945838928, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577891945838928, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09282317757606506, + 0.08822596073150635, + 0.06307019293308258, + 0.16855140030384064, + 0.0697849839925766, + 0.06836209446191788, + 0.19902943074703217, + 0.1538778841495514, + 0.02316906861960888, + 0.07310578972101212 + ], + "credit_map": [ + 0.009677588008344173, + 0.010947734117507935, + 0.013678368180990219, + 0.012536166235804558, + 0.012648901902139187, + 0.014590105041861534, + 0.008694597519934177, + 0.010049713775515556, + 0.009272039867937565, + 0.005689126439392567, + 0.021247845143079758, + 0.012114248238503933, + 0.005327790509909391, + 0.00964664388448, + 0.006583083420991898, + 0.01834517903625965, + 0.010945185087621212, + 0.004820585250854492, + 0.014614902436733246, + 0.007080335170030594, + 0.004158914089202881, + 0.007279570214450359, + 0.0059190476313233376, + 0.008251652121543884, + 0.019636766985058784, + 0.015364576131105423, + 0.02169705554842949, + 0.019240180030465126, + 0.013004191219806671, + 0.03182356432080269, + 0.013239743188023567, + 0.03454531356692314, + 0.009066895581781864, + 0.008268771693110466, + 0.013742372393608093, + 0.008776530623435974, + 0.0063154143281280994, + 0.00713349599391222, + 0.005304958671331406, + 0.011176543310284615, + 0.007280366960912943, + 0.004661806393414736, + 0.009558022953569889, + 0.010472404770553112, + 0.011004535481333733, + 0.010443124920129776, + 0.0068033006973564625, + 0.008138532750308514, + 0.024080080911517143, + 0.03435366973280907, + 0.020746659487485886, + 0.014839750714600086, + 0.03170361742377281, + 0.029351668432354927, + 0.01939125545322895, + 0.024562731385231018, + 0.025668205693364143, + 0.017022164538502693, + 0.013673138804733753, + 0.016047073528170586, + 0.02482808753848076, + 0.026915909722447395, + 0.011273938231170177, + 0.018449366092681885, + 0.0024750102311372757, + 0.00237526698037982, + 0.0017166257603093982, + 0.00407703872770071, + 0.002933772513642907, + 0.004250228870660067, + 0.0020594496745616198, + 0.0032816773746162653, + 0.007404531352221966, + 0.0062479074113070965, + 0.006659609265625477, + 0.008859631605446339, + 0.01620309241116047, + 0.013291658833622932, + 0.006466223392635584, + 0.007973136380314827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.7577891945838928 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030234605073929 + ], + "opposite_logprob_tokens": [ + -2.3503024578094482 + ], + "logprob_reward": -0.10030234605073929, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030234605073929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1325756013393402, + 0.04545379430055618, + 0.024290766566991806, + 0.05936581268906593, + 0.03208082541823387, + 0.05677378177642822, + 0.04835522174835205, + 0.0989862009882927, + 0.18046678602695465, + 0.32165125012397766 + ], + "credit_map": [ + 0.014081919565796852, + 0.010142180137336254, + 0.026217268779873848, + 0.022267865017056465, + 0.01340467669069767, + 0.01647929847240448, + 0.012920293025672436, + 0.017062094062566757, + 0.0042176865972578526, + 0.004153973422944546, + 0.00809850636869669, + 0.008175365626811981, + 0.005179458297789097, + 0.005591708235442638, + 0.0039697508327662945, + 0.006067344453185797, + 0.0032555737998336554, + 0.0017382140504196286, + 0.003583187237381935, + 0.002339651808142662, + 0.002674625487998128, + 0.0028457045555114746, + 0.004092914052307606, + 0.0037608949933201075, + 0.003199150785803795, + 0.003217971185222268, + 0.006364486180245876, + 0.006559318397194147, + 0.003172472817823291, + 0.013849235139787197, + 0.004756550770252943, + 0.018246620893478394, + 0.0036280762869864702, + 0.0026895105838775635, + 0.0061083342880010605, + 0.004282826092094183, + 0.002294846111908555, + 0.003485525958240032, + 0.005170052405446768, + 0.00442165182903409, + 0.0064214421436190605, + 0.00635968241840601, + 0.009053351357579231, + 0.006314736325293779, + 0.006982147693634033, + 0.00855282787233591, + 0.005796211306005716, + 0.00729338126257062, + 0.00530601292848587, + 0.006976451724767685, + 0.00467555969953537, + 0.004695226438343525, + 0.005611120723187923, + 0.006998527329415083, + 0.0072603728622198105, + 0.006831946317106485, + 0.012014445848762989, + 0.010597794316709042, + 0.009699113667011261, + 0.011103514581918716, + 0.015244550071656704, + 0.01721690036356449, + 0.011339555494487286, + 0.011770318262279034, + 0.02316744439303875, + 0.016898250207304955, + 0.02504889667034149, + 0.023097986355423927, + 0.020341906696558, + 0.030201077461242676, + 0.020270461216568947, + 0.02144075743854046, + 0.0353454053401947, + 0.03136434778571129, + 0.025136016309261322, + 0.03529087081551552, + 0.06310956925153732, + 0.058826301246881485, + 0.03415991738438606, + 0.03841878101229668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.10030234605073929 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249288398772478 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249288398772478, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249288398772478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22403591871261597, + 0.11906417459249496, + 0.0873822346329689, + 0.09060449153184891, + 0.07551827281713486, + 0.05046999827027321, + 0.05077596381306648, + 0.08794870972633362, + 0.0670509934425354, + 0.14714933931827545 + ], + "credit_map": [ + 0.02401510253548622, + 0.020655957981944084, + 0.0350547730922699, + 0.04326489195227623, + 0.02158525213599205, + 0.03266630321741104, + 0.018817724660038948, + 0.027975870296359062, + 0.012230763211846352, + 0.011174776591360569, + 0.026281587779521942, + 0.014178366400301456, + 0.011999115347862244, + 0.013348899781703949, + 0.012983903288841248, + 0.016866741701960564, + 0.018819157034158707, + 0.006514695007354021, + 0.017961939796805382, + 0.007682856172323227, + 0.009118667803704739, + 0.010079637169837952, + 0.006381615996360779, + 0.010823644697666168, + 0.006045098882168531, + 0.0051539745181798935, + 0.00849453266710043, + 0.005885760765522718, + 0.006664451211690903, + 0.022936448454856873, + 0.009619373828172684, + 0.025804830715060234, + 0.009409699589014053, + 0.008123327046632767, + 0.014061430469155312, + 0.007666264660656452, + 0.007073909975588322, + 0.008527105674147606, + 0.010221313685178757, + 0.010435209609568119, + 0.004696223884820938, + 0.005268564913421869, + 0.007043336518108845, + 0.007307537831366062, + 0.005295554641634226, + 0.009804180823266506, + 0.004826180636882782, + 0.006228409707546234, + 0.0059006293304264545, + 0.007750518620014191, + 0.006202398799359798, + 0.004595193546265364, + 0.005764719098806381, + 0.007152286823838949, + 0.006027101539075375, + 0.0073831044137477875, + 0.0082777738571167, + 0.008608306758105755, + 0.009756037965416908, + 0.015489153563976288, + 0.010276298969984055, + 0.013943727128207684, + 0.010306130163371563, + 0.011291262693703175, + 0.006831285543739796, + 0.00848853588104248, + 0.0051702531054615974, + 0.007134553045034409, + 0.011121806688606739, + 0.014582420699298382, + 0.005102045834064484, + 0.008620080538094044, + 0.014319400303065777, + 0.015363314189016819, + 0.01031370647251606, + 0.015363523736596107, + 0.02778211049735546, + 0.035659059882164, + 0.01009663101285696, + 0.018251555040478706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.005249288398772478 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, still photograph is tightly focused on a car's interior rearview mirror, encased in a matte black plastic casing. The faint ambient glow from the unseen dashboard barely captures the edges of the mirror's frame and the black fabric of the car's roof above. The entire drama unfolds within the mirror's glass surface, which reflects the nighttime multi-lane highway, rapidly receding into a pitch-black void. The bright white headlights of the following car in the distance are initially captured as distinct balls of light, but they quickly expand into thin, distorted streaks, their glow smeared on the glass by the car's immense speed. Then, these streaks of light rapidly shrink and dim, from vibrant lines to faint, isolated points of light that flicker unsteadily against the overwhelming darkness before vanishing entirely. The painted lane dividers on the asphalt appear to converge at an accelerating pace, swallowed by vanishing points as the car surges forward, creating a powerful and melancholic vision of departure and irreversible distance, immersed in the cool colors of deep blue and absolute black.", + "seed": 8519, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_15/index_7043_seed_8519_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": -0.08063878836110235, + "vlm_logprob": -0.08063878836110235, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.08063878836110235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010448398068547249 + ], + "opposite_logprob_tokens": [ + -4.885448455810547 + ], + "logprob_reward": -0.010448398068547249, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010448398068547249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24359163641929626, + 0.1503899246454239, + 0.04182688146829605, + 0.03776095435023308, + 0.04374636709690094, + 0.020046837627887726, + 0.02072085067629814, + 0.027571430429816246, + 0.33321404457092285, + 0.08113108575344086 + ], + "credit_map": [ + 0.024355821311473846, + 0.024663647636771202, + 0.021577779203653336, + 0.04209456220269203, + 0.04002080112695694, + 0.038853198289871216, + 0.026439731940627098, + 0.025586087256669998, + 0.01635480858385563, + 0.014196543954312801, + 0.026477301493287086, + 0.012105110101401806, + 0.013534462079405785, + 0.036719292402267456, + 0.011575011536478996, + 0.019427383318543434, + 0.004651200491935015, + 0.0033431989140808582, + 0.004948435816913843, + 0.0050564114935696125, + 0.004536130931228399, + 0.005387140437960625, + 0.0074092126451432705, + 0.006495146546512842, + 0.0032016581390053034, + 0.006405620835721493, + 0.003648796584457159, + 0.005867069121450186, + 0.0037514420691877604, + 0.00525288563221693, + 0.004995107185095549, + 0.004638370126485825, + 0.004970110487192869, + 0.005795835051685572, + 0.005425127223134041, + 0.006481361109763384, + 0.003913181833922863, + 0.007606911472976208, + 0.004550829995423555, + 0.005003004334867001, + 0.0024106225464493036, + 0.001657360466197133, + 0.0020557637326419353, + 0.002517620800063014, + 0.002775751054286957, + 0.003117004409432411, + 0.002657925011590123, + 0.002854787278920412, + 0.0019449995597824454, + 0.0017195517430081964, + 0.0022443286143243313, + 0.002358016325160861, + 0.0020343069918453693, + 0.003058865200728178, + 0.0036554874386638403, + 0.0037052910774946213, + 0.003334322012960911, + 0.0016438969178125262, + 0.002532560843974352, + 0.003204180160537362, + 0.0032361990306526423, + 0.004452708177268505, + 0.006670364644378424, + 0.0024971964303404093, + 0.018590040504932404, + 0.013769356533885002, + 0.03897963836789131, + 0.02091033384203911, + 0.03002287447452545, + 0.12433002144098282, + 0.0326685830950737, + 0.053943149745464325, + 0.005419113207608461, + 0.0044561829417943954, + 0.010042094625532627, + 0.010973564349114895, + 0.007590580731630325, + 0.015494105406105518, + 0.019357234239578247, + 0.007798201870173216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.010448398068547249 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636639289557934 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.007636639289557934, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636639289557934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13491149246692657, + 0.06186552718281746, + 0.06591030955314636, + 0.14919693768024445, + 0.07616432011127472, + 0.06840422004461288, + 0.03964253142476082, + 0.03339703381061554, + 0.16394813358783722, + 0.2065594494342804 + ], + "credit_map": [ + 0.015235957689583302, + 0.015271692536771297, + 0.01813393644988537, + 0.020405344665050507, + 0.018161892890930176, + 0.02081606350839138, + 0.012468061409890652, + 0.014418544247746468, + 0.006224522367119789, + 0.006493045948445797, + 0.010321561247110367, + 0.005835247226059437, + 0.006417462136596441, + 0.015305312350392342, + 0.003937606234103441, + 0.007330768741667271, + 0.003966177348047495, + 0.0046805026941001415, + 0.016639534384012222, + 0.003603384830057621, + 0.004789914470165968, + 0.02267005294561386, + 0.004125979728996754, + 0.005434767343103886, + 0.012304923497140408, + 0.01630307547748089, + 0.019066831097006798, + 0.019009767100214958, + 0.01483914814889431, + 0.030368663370609283, + 0.021063411608338356, + 0.01624111644923687, + 0.007274857256561518, + 0.007494622375816107, + 0.007779160048812628, + 0.009326817467808723, + 0.008943373337388039, + 0.014494928531348705, + 0.010248804464936256, + 0.01060175709426403, + 0.0056357793509960175, + 0.006260133348405361, + 0.006701699923723936, + 0.007532571442425251, + 0.008024216629564762, + 0.01315133087337017, + 0.008148168213665485, + 0.012950319796800613, + 0.0043541681952774525, + 0.0033962202724069357, + 0.003977889195084572, + 0.0048062908463180065, + 0.004307862371206284, + 0.00704809557646513, + 0.005533718969672918, + 0.006218285299837589, + 0.0027355493512004614, + 0.0017606542678549886, + 0.0039088912308216095, + 0.0036556192208081484, + 0.0037145973183214664, + 0.006475628353655338, + 0.007805758621543646, + 0.0033403353299945593, + 0.022842731326818466, + 0.012600671499967575, + 0.014330576173961163, + 0.01864062249660492, + 0.01890273578464985, + 0.028942931443452835, + 0.03171252831816673, + 0.015975328162312508, + 0.01820814236998558, + 0.012573841027915478, + 0.026504263281822205, + 0.031203899532556534, + 0.018598321825265884, + 0.0365384966135025, + 0.04216274619102478, + 0.02076975256204605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.007636639289557934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486050069332123 + ], + "opposite_logprob_tokens": [ + -1.2236050367355347 + ], + "logprob_reward": -0.3486050069332123, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486050069332123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0606740340590477, + 0.07636664807796478, + 0.08643809705972672, + 0.07654784619808197, + 0.030912868678569794, + 0.054167039692401886, + 0.07709535211324692, + 0.04653673246502876, + 0.34425854682922363, + 0.1470029056072235 + ], + "credit_map": [ + 0.006566592026501894, + 0.006613410077989101, + 0.0055908202193677425, + 0.009231732226908207, + 0.010168733075261116, + 0.00903945043683052, + 0.007244287524372339, + 0.006219005212187767, + 0.008154616691172123, + 0.00910621602088213, + 0.011700358241796494, + 0.008841061033308506, + 0.0077445670031011105, + 0.016300417482852936, + 0.004800035618245602, + 0.009719369001686573, + 0.006813838612288237, + 0.005910268984735012, + 0.014244095422327518, + 0.009163789451122284, + 0.009896034374833107, + 0.01865234225988388, + 0.011528782546520233, + 0.010228930972516537, + 0.007229559123516083, + 0.009101208299398422, + 0.007649314124137163, + 0.012977327220141888, + 0.006266188807785511, + 0.013086267746984959, + 0.010179632343351841, + 0.010058341547846794, + 0.002980574034154415, + 0.00361452903598547, + 0.003715386614203453, + 0.003793236566707492, + 0.002601180225610733, + 0.004986490122973919, + 0.003980477340519428, + 0.00524098938331008, + 0.008330963551998138, + 0.006292604375630617, + 0.006532123778015375, + 0.007062732242047787, + 0.0056257969699800014, + 0.00797850638628006, + 0.003916450310498476, + 0.008427857421338558, + 0.00736567098647356, + 0.007052886765450239, + 0.009463551454246044, + 0.00799541175365448, + 0.007928554899990559, + 0.01915186084806919, + 0.008217106573283672, + 0.00992029532790184, + 0.004423843696713448, + 0.005230220966041088, + 0.006126602180302143, + 0.0074196127243340015, + 0.00499297259375453, + 0.009406357072293758, + 0.005541329737752676, + 0.0033957932610064745, + 0.044096220284700394, + 0.030125204473733902, + 0.030049851164221764, + 0.045678943395614624, + 0.04072728008031845, + 0.046952392905950546, + 0.07483486831188202, + 0.03179376572370529, + 0.014386726543307304, + 0.010021224617958069, + 0.01981395296752453, + 0.018497835844755173, + 0.014258026145398617, + 0.03669450804591179, + 0.019534798339009285, + 0.01379582192748785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.3486050069332123 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033687494695186615 + ], + "opposite_logprob_tokens": [ + -3.4086875915527344 + ], + "logprob_reward": -0.033687494695186615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033687494695186615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13808514177799225, + 0.02877250500023365, + 0.02266346663236618, + 0.10109606385231018, + 0.13448397815227509, + 0.17522306740283966, + 0.03284361958503723, + 0.07066017389297485, + 0.12403324246406555, + 0.17213879525661469 + ], + "credit_map": [ + 0.013711079955101013, + 0.013586455956101418, + 0.014350959099829197, + 0.021888289600610733, + 0.017468122765421867, + 0.024188537150621414, + 0.017002008855342865, + 0.015889683738350868, + 0.002518667606636882, + 0.0023014196194708347, + 0.0032062570098787546, + 0.0036571987438946962, + 0.004843253176659346, + 0.004485501442104578, + 0.003448081435635686, + 0.004312122706323862, + 0.0020810312125831842, + 0.0020344171207398176, + 0.002794586820527911, + 0.0023853203747421503, + 0.0025008912198245525, + 0.004186135716736317, + 0.0019349236972630024, + 0.0047461590729653835, + 0.009453848004341125, + 0.007798030041158199, + 0.010614949278533459, + 0.016239894554018974, + 0.008004476316273212, + 0.016490554437041283, + 0.016210472211241722, + 0.016283826902508736, + 0.010382471606135368, + 0.011936979368329048, + 0.01508837565779686, + 0.016943924129009247, + 0.012876840308308601, + 0.026512598618865013, + 0.022260403260588646, + 0.01848236471414566, + 0.015124262310564518, + 0.017153644934296608, + 0.019443940371274948, + 0.024127701297402382, + 0.0150721101090312, + 0.027481691911816597, + 0.025583788752555847, + 0.031235916540026665, + 0.004368480760604143, + 0.002520456677302718, + 0.003885365556925535, + 0.004209645092487335, + 0.002982695819810033, + 0.004729940555989742, + 0.006655437406152487, + 0.0034915937576442957, + 0.006567452568560839, + 0.005960424430668354, + 0.00691125588491559, + 0.00899967085570097, + 0.00720088230445981, + 0.01586099900305271, + 0.01000961847603321, + 0.00914986152201891, + 0.009172537364065647, + 0.007909552194178104, + 0.011272741481661797, + 0.012011135928332806, + 0.01366416085511446, + 0.033866796642541885, + 0.020742889493703842, + 0.01539343036711216, + 0.016732770949602127, + 0.011380154639482498, + 0.020432084798812866, + 0.02779252640902996, + 0.015138976275920868, + 0.028619101271033287, + 0.03494888171553612, + 0.01709427684545517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.033687494695186615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1945115178823471, + 0.09926676005125046, + 0.06441908329725266, + 0.08046354353427887, + 0.050716761499643326, + 0.051114220172166824, + 0.06686750054359436, + 0.09390252828598022, + 0.17320625483989716, + 0.12553173303604126 + ], + "credit_map": [ + 0.020200297236442566, + 0.03137051314115524, + 0.021762587130069733, + 0.02714899554848671, + 0.02058199606835842, + 0.025652015581727028, + 0.02353939600288868, + 0.02425573766231537, + 0.00901989545673132, + 0.007962527684867382, + 0.010279403999447823, + 0.009497118182480335, + 0.011262593790888786, + 0.021614158526062965, + 0.010241905227303505, + 0.01938915066421032, + 0.008800285868346691, + 0.0058286977000534534, + 0.009588344022631645, + 0.006146369036287069, + 0.0067422883585095406, + 0.011644572019577026, + 0.007657228037714958, + 0.008011301048099995, + 0.006461253855377436, + 0.009828352369368076, + 0.009694994427263737, + 0.013760400004684925, + 0.006959561724215746, + 0.01369955763220787, + 0.01093867514282465, + 0.009120752103626728, + 0.005258254706859589, + 0.004627786576747894, + 0.0049750483594834805, + 0.007494108751416206, + 0.00456314766779542, + 0.006964528001844883, + 0.006700417492538691, + 0.0101334722712636, + 0.004324509762227535, + 0.007147927302867174, + 0.005668817088007927, + 0.0073340218514204025, + 0.005639512091875076, + 0.007120214402675629, + 0.006003998219966888, + 0.007875222712755203, + 0.005965210497379303, + 0.007078996859490871, + 0.008893398568034172, + 0.005722043104469776, + 0.006427580025047064, + 0.015015775337815285, + 0.005455490667372942, + 0.012309006415307522, + 0.0052797202952206135, + 0.013254445977509022, + 0.009368937462568283, + 0.014721093699336052, + 0.007867377251386642, + 0.02116847038269043, + 0.010809765197336674, + 0.011432720348238945, + 0.023668795824050903, + 0.014017200097441673, + 0.016696225851774216, + 0.022334124892950058, + 0.018105771392583847, + 0.02878383733332157, + 0.031050367280840874, + 0.018549926578998566, + 0.01277837436646223, + 0.005942109506577253, + 0.020483974367380142, + 0.00683972192928195, + 0.014845750294625759, + 0.03745380416512489, + 0.01503431610763073, + 0.012153679504990578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.002816402819007635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, still photograph is tightly focused on a car's interior rearview mirror, encased in a matte black plastic casing. The faint ambient glow from the unseen dashboard barely captures the edges of the mirror's frame and the black fabric of the car's roof above. The entire drama unfolds within the mirror's glass surface, which reflects the nighttime multi-lane highway, rapidly receding into a pitch-black void. The bright white headlights of the following car in the distance are initially captured as distinct balls of light, but they quickly expand into thin, distorted streaks, their glow smeared on the glass by the car's immense speed. Then, these streaks of light rapidly shrink and dim, from vibrant lines to faint, isolated points of light that flicker unsteadily against the overwhelming darkness before vanishing entirely. The painted lane dividers on the asphalt appear to converge at an accelerating pace, swallowed by vanishing points as the car surges forward, creating a powerful and melancholic vision of departure and irreversible distance, immersed in the cool colors of deep blue and absolute black.", + "seed": 8520, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_15/index_7043_seed_8520_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": -0.09830693015828729, + "vlm_logprob": -0.09830693015828729, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.09830693015828729, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01826830767095089 + ], + "opposite_logprob_tokens": [ + -4.14326810836792 + ], + "logprob_reward": -0.01826830767095089, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01826830767095089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038705192506313324, + 0.07976068556308746, + 0.1531512439250946, + 0.013721014373004436, + 0.025050267577171326, + 0.016300221905112267, + 0.010333227925002575, + 0.5473389625549316, + 0.016538934782147408, + 0.09910020977258682 + ], + "credit_map": [ + 0.0033866320736706257, + 0.0041749123483896255, + 0.004621740896254778, + 0.0037872823886573315, + 0.008311409503221512, + 0.006352668162435293, + 0.004709095694124699, + 0.003361449809744954, + 0.005891363602131605, + 0.005538031458854675, + 0.0053481655195355415, + 0.005266645923256874, + 0.022627482190728188, + 0.02122853696346283, + 0.007010299246758223, + 0.006850159261375666, + 0.00921688973903656, + 0.012064334936439991, + 0.010565721429884434, + 0.007666056044399738, + 0.043384850025177, + 0.03730038180947304, + 0.01526512112468481, + 0.01768788881599903, + 0.001294099958613515, + 0.0014239551965147257, + 0.0010230609914287925, + 0.0011845800327137113, + 0.002301817527040839, + 0.0028532722499221563, + 0.0016301277792081237, + 0.0020101002883166075, + 0.0020901523530483246, + 0.004311223980039358, + 0.0021219279151409864, + 0.0038379691541194916, + 0.0038162381388247013, + 0.004292339086532593, + 0.002713769441470504, + 0.001866646809503436, + 0.0018549265805631876, + 0.001931355451233685, + 0.0013108552666381001, + 0.0012770458124578, + 0.003139697713777423, + 0.003509145462885499, + 0.0014215217670425773, + 0.001855672337114811, + 0.0010623176349326968, + 0.0011862816754728556, + 0.0010191904148086905, + 0.0011646937346085906, + 0.002104365499690175, + 0.0014476868091151118, + 0.0011401313822716475, + 0.0012085612397640944, + 0.013883369043469429, + 0.02202637679874897, + 0.005017545539885759, + 0.007108543999493122, + 0.24237403273582458, + 0.23307819664478302, + 0.007242023944854736, + 0.01660890132188797, + 0.0010908591793850064, + 0.0014956970699131489, + 0.0010231597116217017, + 0.0007201912812888622, + 0.003599520307034254, + 0.004582999739795923, + 0.0016059669433161616, + 0.0024205415975302458, + 0.00814605038613081, + 0.01250493060797453, + 0.007619847543537617, + 0.004868634976446629, + 0.010784504935145378, + 0.02945137210190296, + 0.005867102183401585, + 0.019857769832015038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.01826830767095089 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637585513293743 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637585513293743, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637585513293743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0817255824804306, + 0.04168930649757385, + 0.1475106030702591, + 0.03451540693640709, + 0.02420847862958908, + 0.01384750660508871, + 0.021846953779459, + 0.5962229371070862, + 0.02581406943500042, + 0.012619058601558208 + ], + "credit_map": [ + 0.007169333752244711, + 0.009936832822859287, + 0.005131256766617298, + 0.005234691314399242, + 0.023453209549188614, + 0.01891828514635563, + 0.004989765118807554, + 0.006892211269587278, + 0.004029011819511652, + 0.003309200517833233, + 0.00487699918448925, + 0.0026508362498134375, + 0.00902639701962471, + 0.0071320016868412495, + 0.004048981238156557, + 0.006615876220166683, + 0.00873116310685873, + 0.011323841288685799, + 0.009776195511221886, + 0.007398051675409079, + 0.04231671616435051, + 0.0361359566450119, + 0.014186214655637741, + 0.017642466351389885, + 0.003445025533437729, + 0.004321379587054253, + 0.003936468623578548, + 0.0028117552865296602, + 0.005077926442027092, + 0.00564604764804244, + 0.0046019768342375755, + 0.004674830008298159, + 0.001898722373880446, + 0.004324703477323055, + 0.0012733392650261521, + 0.0028903218917548656, + 0.004227448254823685, + 0.0039427983574569225, + 0.002596447942778468, + 0.00305469729937613, + 0.0012338226661086082, + 0.0013470230624079704, + 0.0010749284410849214, + 0.0010204321006312966, + 0.0030630805995315313, + 0.002919579390436411, + 0.0014537194510921836, + 0.001734921126626432, + 0.0019854740239679813, + 0.0028301191050559282, + 0.0021096684504300356, + 0.002809658646583557, + 0.003700489178299904, + 0.002926149871200323, + 0.0026093479245901108, + 0.002876044949516654, + 0.01500016450881958, + 0.02408452332019806, + 0.005373143125325441, + 0.007798993960022926, + 0.26574763655662537, + 0.253923624753952, + 0.007559956982731819, + 0.016734937205910683, + 0.002135978313162923, + 0.0023159903939813375, + 0.002252521924674511, + 0.0023032051976770163, + 0.004345677327364683, + 0.0060176425613462925, + 0.0020246447529643774, + 0.0044184112921357155, + 0.0017399467760697007, + 0.002067022956907749, + 0.001342741772532463, + 0.0017442911630496383, + 0.0015000798739492893, + 0.0017581345746293664, + 0.0011641514720395207, + 0.0013026908272877336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.007637585513293743 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486221730709076 + ], + "opposite_logprob_tokens": [ + -1.22362220287323 + ], + "logprob_reward": -0.3486221730709076, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486221730709076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018725859001278877, + 0.05232822522521019, + 0.03437535837292671, + 0.020731976255774498, + 0.04399991035461426, + 0.023582953959703445, + 0.057474520057439804, + 0.6961043477058411, + 0.018322112038731575, + 0.03435469791293144 + ], + "credit_map": [ + 0.0016608403529971838, + 0.002524069743230939, + 0.001949023804627359, + 0.002169516170397401, + 0.0029433981981128454, + 0.0030705174431204796, + 0.0017739325994625688, + 0.0026345611549913883, + 0.0037978459149599075, + 0.0037689213640987873, + 0.003644293872639537, + 0.002895965240895748, + 0.011866232380270958, + 0.01312296837568283, + 0.004377619829028845, + 0.00885438360273838, + 0.0026452329475432634, + 0.003505690023303032, + 0.0025049529504030943, + 0.002576420083642006, + 0.007924075238406658, + 0.005743767134845257, + 0.003889998886734247, + 0.005585222505033016, + 0.002287766197696328, + 0.002156892092898488, + 0.001803734921850264, + 0.0020005491096526384, + 0.004059378989040852, + 0.003540060017257929, + 0.0025808594655245543, + 0.0023027360439300537, + 0.003397397929802537, + 0.010325715877115726, + 0.0025660586543381214, + 0.004660616163164377, + 0.005222607869654894, + 0.007363190874457359, + 0.004549086093902588, + 0.005915236193686724, + 0.002827997086569667, + 0.0023834207095205784, + 0.0015520346350967884, + 0.001980700995773077, + 0.004414084367454052, + 0.005017850548028946, + 0.0024167003575712442, + 0.0029901661910116673, + 0.0055422112345695496, + 0.006889150943607092, + 0.005616213660687208, + 0.007223954424262047, + 0.010824511758983135, + 0.0077178399078547955, + 0.0069419690407812595, + 0.006718670018017292, + 0.018085455521941185, + 0.028635617345571518, + 0.0071168444119393826, + 0.010073108598589897, + 0.310311883687973, + 0.2946684956550598, + 0.009214081801474094, + 0.017998913303017616, + 0.0015446472680196166, + 0.0023058762308210135, + 0.0019454800058156252, + 0.002258972031995654, + 0.0032783299684524536, + 0.002885946072638035, + 0.001955546671524644, + 0.002147314138710499, + 0.002692775335162878, + 0.0038133258931338787, + 0.002224452793598175, + 0.00292355939745903, + 0.0026299315504729748, + 0.008138451725244522, + 0.0034241341054439545, + 0.008508069440722466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.3486221730709076 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291488260030746 + ], + "opposite_logprob_tokens": [ + -2.237914800643921 + ], + "logprob_reward": -0.11291488260030746, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291488260030746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21015508472919464, + 0.08840282261371613, + 0.04645325243473053, + 0.01957324892282486, + 0.03911731019616127, + 0.03476022556424141, + 0.035963576287031174, + 0.48066970705986023, + 0.0120346425101161, + 0.0328700914978981 + ], + "credit_map": [ + 0.014977146871387959, + 0.018793797120451927, + 0.01034244243055582, + 0.007705507799983025, + 0.06948655843734741, + 0.052651528269052505, + 0.02077575959265232, + 0.015422353520989418, + 0.009354940615594387, + 0.007174803409725428, + 0.010679596103727818, + 0.006880144588649273, + 0.017260896041989326, + 0.01550737302750349, + 0.008707619272172451, + 0.012837446294724941, + 0.0035201632417738438, + 0.004806759301573038, + 0.00332903815433383, + 0.0034698261879384518, + 0.009668479673564434, + 0.010309042409062386, + 0.006186303216964006, + 0.005163639783859253, + 0.0017501151887699962, + 0.0027139647863805294, + 0.0020285379141569138, + 0.0018852736102417111, + 0.002514840569347143, + 0.003371665719896555, + 0.0030785310082137585, + 0.00223032059147954, + 0.0043243784457445145, + 0.004632971249520779, + 0.0033215449657291174, + 0.0038361521437764168, + 0.00777795584872365, + 0.00618670042604208, + 0.004675460979342461, + 0.004362144973129034, + 0.003918771166354418, + 0.003292983630672097, + 0.0032017785124480724, + 0.002927674911916256, + 0.005618004128336906, + 0.004936286248266697, + 0.0042657251469790936, + 0.0065990025177598, + 0.0026803393848240376, + 0.00334964320063591, + 0.002964701736345887, + 0.002317665843293071, + 0.00684871943667531, + 0.00510009191930294, + 0.007120416034013033, + 0.005581999197602272, + 0.012477833777666092, + 0.019733814522624016, + 0.004564682953059673, + 0.006043226458132267, + 0.21737642586231232, + 0.2044980525970459, + 0.005975967273116112, + 0.009999675676226616, + 0.0008175971452146769, + 0.0012801692355424166, + 0.0016119665233418345, + 0.001244628569111228, + 0.0014356133760884404, + 0.002394401468336582, + 0.0011461434187367558, + 0.0021041235886514187, + 0.002705673221498728, + 0.004623310174793005, + 0.0022090112324804068, + 0.0025153798051178455, + 0.003964546136558056, + 0.007780036423355341, + 0.00266136578284204, + 0.006410770583897829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.11291488260030746 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up view of a car's interior rearview mirror with a matte black plastic casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the reflection show a nighttime multi-lane highway with bright white headlights that appear as distinct balls of light before expanding into distorted streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the streaks of light from the headlights rapidly shrink and dim into faint, isolated points of light that flicker unsteadily before vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the painted lane dividers on the asphalt appear to converge at an accelerating pace towards a vanishing point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cool colors of deep blue and absolute black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091701935976744 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091701935976744, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091701935976744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09099812060594559, + 0.025031834840774536, + 0.037393052130937576, + 0.04740986227989197, + 0.02307458594441414, + 0.016787726432085037, + 0.008213324472308159, + 0.7176677584648132, + 0.014085975475609303, + 0.019337806850671768 + ], + "credit_map": [ + 0.006755456328392029, + 0.008011367172002792, + 0.005138517357409, + 0.00373309338465333, + 0.030027495697140694, + 0.02439122647047043, + 0.007692065089941025, + 0.005248897708952427, + 0.0026606949977576733, + 0.002602065447717905, + 0.0032095389906316996, + 0.001982718473300338, + 0.004574754741042852, + 0.003815473522990942, + 0.002395347459241748, + 0.00379124004393816, + 0.0022617860231548548, + 0.002811793005093932, + 0.0025066586676985025, + 0.0018552106339484453, + 0.010736455209553242, + 0.009006090462207794, + 0.003351873718202114, + 0.004863187205046415, + 0.004833253566175699, + 0.005756346043199301, + 0.005197248421609402, + 0.0038253283128142357, + 0.006946125067770481, + 0.007274139206856489, + 0.007240281905978918, + 0.006337141618132591, + 0.0021104030311107635, + 0.0038061337545514107, + 0.0017617245903238654, + 0.0020263788755983114, + 0.003798305755481124, + 0.004079392179846764, + 0.0026220695581287146, + 0.002870177384465933, + 0.0027883227448910475, + 0.002522614784538746, + 0.0011402745731174946, + 0.0012041343143209815, + 0.0032204692251980305, + 0.002647319808602333, + 0.001593496766872704, + 0.0016710947966203094, + 0.0009083537152037024, + 0.0011135316453874111, + 0.0008604725007899106, + 0.0007235188968479633, + 0.0014190028887242079, + 0.0012327339500188828, + 0.0011319175828248262, + 0.000823792302981019, + 0.018587691709399223, + 0.029105570167303085, + 0.007109914440661669, + 0.009606614708900452, + 0.3133556544780731, + 0.30353498458862305, + 0.010395596735179424, + 0.025971708819270134, + 0.0009872677037492394, + 0.0013438811292871833, + 0.0011336650932207704, + 0.0012175593292340636, + 0.002291150391101837, + 0.0037014316767454147, + 0.0010872301645576954, + 0.002323789754882455, + 0.0017806475516408682, + 0.002046008128672838, + 0.0009464497561566532, + 0.0014082937268540263, + 0.0019854961428791285, + 0.0059961010701954365, + 0.0008527758764103055, + 0.004322034306824207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.004091701935976744 + } + ], + "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/0000052/group_2/index_tensor([3650, 3650])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_2/index_tensor([3650, 3650])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d6dcd8c631f876ed795200926844145116498db0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_2/index_tensor([3650, 3650])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic aerial time-lapse shot, captured from a high vantage point, surveys a vast expanse of golden desert, its rolling dunes casting sharp, long shadows under the scorching afternoon sun. As the camera slowly pulls back, a tiny point in the center of the frame undergoes a magical transformation. A vibrant emerald green radiates outward at a visible speed, seemingly devouring the surrounding yellow sand. First, tender green grass breaks through the sand, forming an ever-expanding circular carpet. Shortly thereafter, low shrubs and vibrant wildflowers quickly intersperse the landscape. Finally, towering palm trees and leafy broadleaf trees rise from the ground, their canopies unfurling in seconds. Simultaneously, a clear stream appears out of nowhere, winding through the new vegetation and converging into a shimmering central lake, its waters reflecting the azure sky. The entire expansion process is smooth and shocking, dramatically transforming the monotonous wasteland into a thriving ecosystem full of life, creating an atmosphere of grand, magical and awe-inspiring miracle of life.", + "seed": 12960, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_2/index_3650_seed_12960_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": -0.08732529722619801, + "vlm_logprob": -0.08732529722619801, + "vlm_yes_no_margin": 5.65, + "vlm_token_credit": -0.08732529722619801, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.0033004595898091793 + ], + "opposite_logprob_tokens": [ + -7.128300666809082 + ], + "logprob_reward": -0.0033004595898091793, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0033004595898091793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2858862578868866, + 0.20315609872341156, + 0.19423547387123108, + 0.06745649874210358, + 0.01875910349190235, + 0.08678086847066879, + 0.04067745804786682, + 0.01450990978628397, + 0.03415928781032562, + 0.05437908694148064 + ], + "credit_map": [ + 0.038016438484191895, + 0.019969511777162552, + 0.05406007543206215, + 0.03378027305006981, + 0.024402860552072525, + 0.018156500533223152, + 0.05556344985961914, + 0.04193711280822754, + 0.005420259665697813, + 0.006612818688154221, + 0.07831116020679474, + 0.020488066598773003, + 0.00476817786693573, + 0.0051759593188762665, + 0.06968439370393753, + 0.012695251032710075, + 0.010826028883457184, + 0.02200082689523697, + 0.041151903569698334, + 0.022071588784456253, + 0.012619218789041042, + 0.01840774528682232, + 0.020754419267177582, + 0.04640373960137367, + 0.00896932277828455, + 0.007342311087995768, + 0.008433314971625805, + 0.010273058898746967, + 0.011143692769110203, + 0.005043284967541695, + 0.008019800297915936, + 0.008231709711253643, + 0.0014669715892523527, + 0.002477750414982438, + 0.0033288251142948866, + 0.002862282795831561, + 0.002226301236078143, + 0.001956300577148795, + 0.0026392429135739803, + 0.0018014266388490796, + 0.0052780406549572945, + 0.007036705035716295, + 0.010903180576860905, + 0.02066337876021862, + 0.009217150509357452, + 0.006316474173218012, + 0.016621090471744537, + 0.010744838044047356, + 0.0020996013190597296, + 0.002275117440149188, + 0.0056389146484434605, + 0.015000646002590656, + 0.001962744165211916, + 0.0016549075953662395, + 0.004421252757310867, + 0.007624268997460604, + 0.0010862714843824506, + 0.0014688625233247876, + 0.0019534507300704718, + 0.002721637487411499, + 0.0012062402674928308, + 0.001243990147486329, + 0.0023582950234413147, + 0.002471159677952528, + 0.0034239653032273054, + 0.003016280010342598, + 0.006606066599488258, + 0.007669705431908369, + 0.002455055946484208, + 0.002379348035901785, + 0.004282302688807249, + 0.004326558671891689, + 0.006053052376955748, + 0.003910068888217211, + 0.007616197690367699, + 0.007778183091431856, + 0.0035562070552259684, + 0.009306215681135654, + 0.008416236378252506, + 0.0077429176308214664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0033004595898091793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.4289824068546295 + ], + "opposite_logprob_tokens": [ + -1.0539823770523071 + ], + "logprob_reward": -0.4289824068546295, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289824068546295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3442338705062866, + 0.13433010876178741, + 0.11820676922798157, + 0.09311258047819138, + 0.04039753973484039, + 0.0715765655040741, + 0.03852648660540581, + 0.044973041862249374, + 0.08415783196687698, + 0.030485212802886963 + ], + "credit_map": [ + 0.05153986066579819, + 0.028023164719343185, + 0.06321652978658676, + 0.04502713307738304, + 0.03246968612074852, + 0.028690535575151443, + 0.05136558786034584, + 0.04390140250325203, + 0.004128101747483015, + 0.006442319136112928, + 0.04860539734363556, + 0.012930518947541714, + 0.00454155495390296, + 0.007469504605978727, + 0.04259290173649788, + 0.007619828451424837, + 0.005432226229459047, + 0.0063843191601336, + 0.015078717842698097, + 0.02201523818075657, + 0.015155715867877007, + 0.01394268311560154, + 0.020039500668644905, + 0.020158370956778526, + 0.014790358953177929, + 0.009721526876091957, + 0.009746675379574299, + 0.014498199336230755, + 0.014508774504065514, + 0.009284541942179203, + 0.0092088608071208, + 0.011353648267686367, + 0.005643278360366821, + 0.004124497529119253, + 0.005155099555850029, + 0.00539207411929965, + 0.004560580477118492, + 0.006138075143098831, + 0.004689075518399477, + 0.004694863688200712, + 0.005710445810109377, + 0.006490899715572596, + 0.014074848033487797, + 0.01406433992087841, + 0.008101046085357666, + 0.0049355532974004745, + 0.01162315346300602, + 0.006576285697519779, + 0.0027257022447884083, + 0.0022040794137865305, + 0.005276822950690985, + 0.01005723886191845, + 0.0027859178371727467, + 0.002206498524174094, + 0.006145113613456488, + 0.007125115487724543, + 0.0026955371722579002, + 0.00324152410030365, + 0.006587367504835129, + 0.011833131313323975, + 0.004070940427482128, + 0.003860386786982417, + 0.006259238347411156, + 0.006424923427402973, + 0.00941927544772625, + 0.009485212154686451, + 0.013446193188428879, + 0.010529639199376106, + 0.009578930214047432, + 0.007059438154101372, + 0.01104702241718769, + 0.013592127710580826, + 0.003279538359493017, + 0.0022231561597436666, + 0.0055917659774422646, + 0.003440750762820244, + 0.0019874488934874535, + 0.00532549200579524, + 0.004104228224605322, + 0.004532833583652973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.4289824068546295 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.00034421717282384634 + ], + "opposite_logprob_tokens": [ + -8.000344276428223 + ], + "logprob_reward": -0.00034421717282384634, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.00034421717282384634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2942822277545929, + 0.16087281703948975, + 0.1695835441350937, + 0.06696067005395889, + 0.016925480216741562, + 0.05385655537247658, + 0.05650758370757103, + 0.0747927576303482, + 0.03581738844513893, + 0.07040096819400787 + ], + "credit_map": [ + 0.04332993924617767, + 0.020569512620568275, + 0.05014067143201828, + 0.039030518382787704, + 0.022988520562648773, + 0.020422611385583878, + 0.06093498691916466, + 0.0368654690682888, + 0.0038474127650260925, + 0.003228799905627966, + 0.06564868986606598, + 0.018683219328522682, + 0.004040195140987635, + 0.004070212133228779, + 0.05244866758584976, + 0.008905608206987381, + 0.005744494032114744, + 0.013550354167819023, + 0.04156781733036041, + 0.020081087946891785, + 0.012697053141891956, + 0.009569868445396423, + 0.021971814334392548, + 0.044401057064533234, + 0.010754410177469254, + 0.005630509927868843, + 0.011144300922751427, + 0.007658529561012983, + 0.00931297056376934, + 0.005856645293533802, + 0.007682451047003269, + 0.008920852094888687, + 0.001347400015220046, + 0.002178021939471364, + 0.001966359792277217, + 0.0037623574025928974, + 0.002282625762745738, + 0.0016280303243547678, + 0.0023366697132587433, + 0.001424014917574823, + 0.0040791030041873455, + 0.003892641980201006, + 0.009663725271821022, + 0.01109875924885273, + 0.005453954916447401, + 0.004425118211656809, + 0.00920261349529028, + 0.006040637847036123, + 0.003255503484979272, + 0.0029372896533459425, + 0.005670153070241213, + 0.022689878940582275, + 0.0023621434811502695, + 0.002394024981185794, + 0.006368912756443024, + 0.010829676873981953, + 0.00349487760104239, + 0.002892564283683896, + 0.004857352003455162, + 0.011244378983974457, + 0.003326882142573595, + 0.0025695739313960075, + 0.02660178579390049, + 0.019805341958999634, + 0.003618833376094699, + 0.004395935218781233, + 0.0049050417728722095, + 0.006284572649747133, + 0.004376763943582773, + 0.003013762878254056, + 0.004031156189739704, + 0.005191322416067123, + 0.007232211995869875, + 0.0048444089479744434, + 0.010328245349228382, + 0.013521571643650532, + 0.004439501091837883, + 0.008252328261733055, + 0.012571544386446476, + 0.009211162105202675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00034421717282384634 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.001510432455688715 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510432455688715, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510432455688715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14404180645942688, + 0.09471648186445236, + 0.0721236988902092, + 0.17587324976921082, + 0.03092537261545658, + 0.056477829813957214, + 0.03653084114193916, + 0.15906493365764618, + 0.049016598612070084, + 0.18122923374176025 + ], + "credit_map": [ + 0.021228939294815063, + 0.013704298064112663, + 0.026708444580435753, + 0.01827475056052208, + 0.015598119236528873, + 0.010281985625624657, + 0.020322518423199654, + 0.01792275719344616, + 0.0066458722576498985, + 0.01376483216881752, + 0.022551093250513077, + 0.02100400999188423, + 0.005749824456870556, + 0.00574088841676712, + 0.011589065194129944, + 0.007670895662158728, + 0.003600680734962225, + 0.00466102035716176, + 0.014803014695644379, + 0.014261170290410519, + 0.0036189586389809847, + 0.004990277346223593, + 0.014337289147078991, + 0.011851287446916103, + 0.02153816446661949, + 0.021793443709611893, + 0.014670135453343391, + 0.02543296478688717, + 0.03835723176598549, + 0.014079933986067772, + 0.017868230119347572, + 0.022133126854896545, + 0.0024463837035000324, + 0.005080665461719036, + 0.005326189566403627, + 0.004855849780142307, + 0.002762986347079277, + 0.0028816841077059507, + 0.004622214008122683, + 0.0029493982437998056, + 0.003964904695749283, + 0.005401617847383022, + 0.00721551850438118, + 0.013150207698345184, + 0.004213519394397736, + 0.0037663786206394434, + 0.010670287534594536, + 0.008095393888652325, + 0.002567429793998599, + 0.0024398353416472673, + 0.0032846746034920216, + 0.013628492131829262, + 0.0020664851181209087, + 0.002231351798400283, + 0.004124938044697046, + 0.006187637336552143, + 0.007416436914354563, + 0.005410847719758749, + 0.008982596918940544, + 0.01894640177488327, + 0.006224541924893856, + 0.005366397090256214, + 0.060273684561252594, + 0.04644401744008064, + 0.003927708137780428, + 0.0061315200291574, + 0.006811398081481457, + 0.005515993572771549, + 0.008245742879807949, + 0.004801872652024031, + 0.006369109731167555, + 0.007213255390524864, + 0.01842588558793068, + 0.010590417310595512, + 0.024177277460694313, + 0.031467556953430176, + 0.012106197886168957, + 0.02746996283531189, + 0.03238251805305481, + 0.024609418585896492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.001510432455688715 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.0024889700580388308 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.0024889700580388308, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024889700580388308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15950749814510345, + 0.1345786303281784, + 0.1542617529630661, + 0.06179478392004967, + 0.0481678768992424, + 0.16839861869812012, + 0.027974659577012062, + 0.11165045201778412, + 0.0444762147963047, + 0.08918946981430054 + ], + "credit_map": [ + 0.017767284065485, + 0.011551564559340477, + 0.03967737779021263, + 0.015514425933361053, + 0.013900562189519405, + 0.00829824898391962, + 0.03684912994503975, + 0.015948891639709473, + 0.011223619803786278, + 0.01437685452401638, + 0.019217610359191895, + 0.0280240960419178, + 0.00997045449912548, + 0.009453048929572105, + 0.02700943313539028, + 0.015303508378565311, + 0.008006238378584385, + 0.011127736419439316, + 0.024576576426625252, + 0.03169681131839752, + 0.007640589494258165, + 0.011502939276397228, + 0.03584371879696846, + 0.023867134004831314, + 0.006734931841492653, + 0.006058231461793184, + 0.00920859631150961, + 0.00975379254668951, + 0.0068403878249228, + 0.006622607354074717, + 0.008483534678816795, + 0.008092702366411686, + 0.0036080467980355024, + 0.006991095840930939, + 0.007171256002038717, + 0.007474241778254509, + 0.004772490356117487, + 0.006470612715929747, + 0.006685043219476938, + 0.004995089489966631, + 0.011808390729129314, + 0.01336305495351553, + 0.02884739078581333, + 0.03634271025657654, + 0.01859879493713379, + 0.012153019197285175, + 0.028752319514751434, + 0.018532944843173027, + 0.001883070101030171, + 0.003207265865057707, + 0.0033408741001039743, + 0.007078814785927534, + 0.0031651491299271584, + 0.001784594962373376, + 0.004094935487955809, + 0.003419955028221011, + 0.005644077435135841, + 0.0032800198532640934, + 0.007236794102936983, + 0.013859815895557404, + 0.005083094350993633, + 0.005178246181458235, + 0.03970484063029289, + 0.03166356682777405, + 0.0057557509280741215, + 0.007302762474864721, + 0.005223830230534077, + 0.004415715113282204, + 0.004269510041922331, + 0.005383014213293791, + 0.005801995750516653, + 0.006323633249849081, + 0.008097042329609394, + 0.006659748498350382, + 0.015660518780350685, + 0.01764390990138054, + 0.004113851115107536, + 0.009454915300011635, + 0.016218069940805435, + 0.01134141068905592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0024889700580388308 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic aerial time-lapse shot, captured from a high vantage point, surveys a vast expanse of golden desert, its rolling dunes casting sharp, long shadows under the scorching afternoon sun. As the camera slowly pulls back, a tiny point in the center of the frame undergoes a magical transformation. A vibrant emerald green radiates outward at a visible speed, seemingly devouring the surrounding yellow sand. First, tender green grass breaks through the sand, forming an ever-expanding circular carpet. Shortly thereafter, low shrubs and vibrant wildflowers quickly intersperse the landscape. Finally, towering palm trees and leafy broadleaf trees rise from the ground, their canopies unfurling in seconds. Simultaneously, a clear stream appears out of nowhere, winding through the new vegetation and converging into a shimmering central lake, its waters reflecting the azure sky. The entire expansion process is smooth and shocking, dramatically transforming the monotonous wasteland into a thriving ecosystem full of life, creating an atmosphere of grand, magical and awe-inspiring miracle of life.", + "seed": 12961, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_2/index_3650_seed_12961_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": -0.0911320474348031, + "vlm_logprob": -0.0911320474348031, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.0911320474348031, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.002719872398301959 + ], + "opposite_logprob_tokens": [ + -8.37771987915039 + ], + "logprob_reward": -0.002719872398301959, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.002719872398301959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14114339649677277, + 0.2233303338289261, + 0.07698392868041992, + 0.10657579451799393, + 0.0693388432264328, + 0.06717448681592941, + 0.07086939364671707, + 0.055737145245075226, + 0.054605595767498016, + 0.13424111902713776 + ], + "credit_map": [ + 0.030800512060523033, + 0.012571739964187145, + 0.017865903675556183, + 0.021029550582170486, + 0.01137159951031208, + 0.011751987971365452, + 0.021229665726423264, + 0.014522428624331951, + 0.024473020806908607, + 0.013645474798977375, + 0.03388645499944687, + 0.04929569736123085, + 0.018399512395262718, + 0.0159994438290596, + 0.043901845812797546, + 0.02372887171804905, + 0.008645005524158478, + 0.013541675172746181, + 0.008461588993668556, + 0.009316347539424896, + 0.012206077575683594, + 0.011293267831206322, + 0.007870567962527275, + 0.005649401340633631, + 0.011113660410046577, + 0.00712725892663002, + 0.016313526779413223, + 0.02052694745361805, + 0.01064645778387785, + 0.009407889097929, + 0.017995525151491165, + 0.01344453264027834, + 0.008355916477739811, + 0.005340089555829763, + 0.008841920644044876, + 0.013697759248316288, + 0.00593145377933979, + 0.006167386192828417, + 0.012392796576023102, + 0.008611522614955902, + 0.008123056963086128, + 0.007401208393275738, + 0.007205852307379246, + 0.009937561117112637, + 0.007711359299719334, + 0.006030607037246227, + 0.014056449756026268, + 0.006708397530019283, + 0.008767891675233841, + 0.01232868991792202, + 0.006128364708274603, + 0.010307894088327885, + 0.009977889247238636, + 0.006926691625267267, + 0.010128675028681755, + 0.006303297821432352, + 0.004131307825446129, + 0.007911034859716892, + 0.007699355483055115, + 0.009232358075678349, + 0.009599574841558933, + 0.004722690675407648, + 0.005685282871127129, + 0.006755536887794733, + 0.004366558976471424, + 0.010097753256559372, + 0.007508495822548866, + 0.009839551523327827, + 0.006683763116598129, + 0.004929512273520231, + 0.006160759832710028, + 0.005019201897084713, + 0.007678786292672157, + 0.007218638435006142, + 0.0399966835975647, + 0.012330266647040844, + 0.008478997275233269, + 0.008713023737072945, + 0.010692951269447803, + 0.0391317680478096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.002719872398301959 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.42897331714630127 + ], + "opposite_logprob_tokens": [ + -1.0539733171463013 + ], + "logprob_reward": -0.42897331714630127, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42897331714630127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08809104561805725, + 0.10193699598312378, + 0.0828963965177536, + 0.12158551812171936, + 0.046172045171260834, + 0.06462530791759491, + 0.14377321302890778, + 0.07053671777248383, + 0.14656321704387665, + 0.13381950557231903 + ], + "credit_map": [ + 0.01777598448097706, + 0.007968469522893429, + 0.013894359581172466, + 0.009742108173668385, + 0.007357614114880562, + 0.00796480942517519, + 0.015352911315858364, + 0.008034788072109222, + 0.012679300270974636, + 0.007462018635123968, + 0.01752498932182789, + 0.01752684824168682, + 0.009016476571559906, + 0.0076622068881988525, + 0.016527779400348663, + 0.013537372462451458, + 0.005146384704858065, + 0.00752206752076745, + 0.01499810442328453, + 0.019887534901499748, + 0.006950328126549721, + 0.007226142566651106, + 0.015828851610422134, + 0.005336979869753122, + 0.008456526324152946, + 0.00929268542677164, + 0.01801038719713688, + 0.032253194600343704, + 0.008684367872774601, + 0.00878765806555748, + 0.021013710647821426, + 0.015086994506418705, + 0.00746398139744997, + 0.0033953420352190733, + 0.006597470957785845, + 0.00879710540175438, + 0.004220618866384029, + 0.003577688941732049, + 0.006754211150109768, + 0.005365625489503145, + 0.011537477374076843, + 0.008225379511713982, + 0.006126672029495239, + 0.008941547013819218, + 0.008272253908216953, + 0.005849897395819426, + 0.00884610041975975, + 0.006825985386967659, + 0.02087946981191635, + 0.02892226167023182, + 0.013824572786688805, + 0.014484957791864872, + 0.017136208713054657, + 0.01409786194562912, + 0.01976681500673294, + 0.014661073684692383, + 0.0049188025295734406, + 0.00864200945943594, + 0.00846441276371479, + 0.011316100135445595, + 0.012697884812951088, + 0.00690858531743288, + 0.009118088521063328, + 0.008470839820802212, + 0.012128259986639023, + 0.034222204238176346, + 0.017565812915563583, + 0.024042338132858276, + 0.016136471182107925, + 0.011870634742081165, + 0.0187275018543005, + 0.011869993060827255, + 0.005245820619165897, + 0.00442694453522563, + 0.04565446451306343, + 0.018481353297829628, + 0.003927452024072409, + 0.004424065351486206, + 0.007991652935743332, + 0.0436677448451519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.42897331714630127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.000563224486541003 + ], + "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": [ + 0.10602657496929169, + 0.05725463107228279, + 0.12383168190717697, + 0.04499700665473938, + 0.09180457144975662, + 0.08301421254873276, + 0.04014721140265465, + 0.05920245498418808, + 0.0855712890625, + 0.30815041065216064 + ], + "credit_map": [ + 0.022543782368302345, + 0.01188789401203394, + 0.015176910907030106, + 0.010850616730749607, + 0.00875162798911333, + 0.008678062818944454, + 0.015776997432112694, + 0.012360670603811741, + 0.006076267454773188, + 0.005065874196588993, + 0.008467916399240494, + 0.008135530166327953, + 0.007145969197154045, + 0.008833782747387886, + 0.007831567898392677, + 0.005697713699191809, + 0.0067198858596384525, + 0.008818022906780243, + 0.024693937972187996, + 0.03730901703238487, + 0.009967910125851631, + 0.006618311628699303, + 0.02090410701930523, + 0.008800476789474487, + 0.005548061802983284, + 0.0032365405932068825, + 0.006778149865567684, + 0.009562424384057522, + 0.0042395503260195255, + 0.003098604269325733, + 0.0071332985535264015, + 0.005400374066084623, + 0.006875826511532068, + 0.008419519290328026, + 0.015020993538200855, + 0.016936669126152992, + 0.007794524542987347, + 0.008199050091207027, + 0.01669151894748211, + 0.01186645869165659, + 0.020809104666113853, + 0.012592087499797344, + 0.0068008070811629295, + 0.008692611008882523, + 0.009556652046740055, + 0.008103934116661549, + 0.007037749979645014, + 0.009421257302165031, + 0.005027052480727434, + 0.006608345080167055, + 0.0043098474852740765, + 0.005444594193249941, + 0.005002845544368029, + 0.003935621585696936, + 0.005334325600415468, + 0.004484573844820261, + 0.005863415077328682, + 0.008294863626360893, + 0.009233062155544758, + 0.008589426055550575, + 0.007450461853295565, + 0.005366641562432051, + 0.00872150156646967, + 0.00568307563662529, + 0.0058133904822170734, + 0.01684068702161312, + 0.011381634511053562, + 0.014502362348139286, + 0.007426153402775526, + 0.00927689764648676, + 0.011522306129336357, + 0.008807841688394547, + 0.01170189119875431, + 0.012466887943446636, + 0.11621098220348358, + 0.02242150716483593, + 0.008404127322137356, + 0.01035705953836441, + 0.01248867530375719, + 0.11409926414489746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.000563224486541003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.002814976265653968 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.002814976265653968, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002814976265653968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0767969861626625, + 0.04574856534600258, + 0.014467118307948112, + 0.03807509317994118, + 0.026076246052980423, + 0.03963704779744148, + 0.026329312473535538, + 0.021488580852746964, + 0.12195239961147308, + 0.5894286632537842 + ], + "credit_map": [ + 0.012235559523105621, + 0.005087982397526503, + 0.011038420721888542, + 0.010763276368379593, + 0.00830322876572609, + 0.005155564285814762, + 0.014556821435689926, + 0.009656129404902458, + 0.005241308826953173, + 0.0047416090965271, + 0.005735284648835659, + 0.005658717826008797, + 0.006277845706790686, + 0.008348250761628151, + 0.004755539819598198, + 0.004990008659660816, + 0.0012349130120128393, + 0.001260002376511693, + 0.002245861804112792, + 0.003174315672367811, + 0.0015858726110309362, + 0.001499958336353302, + 0.0023059919476509094, + 0.0011602017330005765, + 0.0034031032118946314, + 0.0028940713964402676, + 0.005670120939612389, + 0.010058642365038395, + 0.003283171448856592, + 0.002672638976946473, + 0.004817087668925524, + 0.005276254843920469, + 0.0021655121818184853, + 0.0020827895496040583, + 0.004217044450342655, + 0.005864189472049475, + 0.001771929208189249, + 0.0019562679808586836, + 0.004380454309284687, + 0.003638056805357337, + 0.01061532273888588, + 0.006040709558874369, + 0.003331915009766817, + 0.003968065604567528, + 0.0038999514654278755, + 0.003489343449473381, + 0.0038152923807501793, + 0.004476443398743868, + 0.0033778513316065073, + 0.004963680170476437, + 0.0025405704509466887, + 0.0038929616566747427, + 0.002866788301616907, + 0.002401866251602769, + 0.004186732228845358, + 0.0020988613832741976, + 0.0021707462146878242, + 0.0026656354311853647, + 0.0032850729767233133, + 0.003186458256095648, + 0.0031194102484732866, + 0.0019257360836490989, + 0.0033240611664950848, + 0.0018114587292075157, + 0.009065613150596619, + 0.03024373948574066, + 0.01387919019907713, + 0.016460945829749107, + 0.013199460692703724, + 0.013157849200069904, + 0.015096187591552734, + 0.010849405080080032, + 0.022434046491980553, + 0.02088264748454094, + 0.22467879951000214, + 0.044144317507743835, + 0.01203813124448061, + 0.020330756902694702, + 0.024033047258853912, + 0.2208869308233261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.002814976265653968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.020588846877217293 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588846877217293, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588846877217293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07092378288507462, + 0.07368995994329453, + 0.035022009164094925, + 0.09678909182548523, + 0.028225857764482498, + 0.040010955184698105, + 0.015002429485321045, + 0.06500241160392761, + 0.08421984314918518, + 0.49111366271972656 + ], + "credit_map": [ + 0.010349784977734089, + 0.008589502424001694, + 0.008115737698972225, + 0.0111779710277915, + 0.00887030828744173, + 0.005877163726836443, + 0.010802268981933594, + 0.007141046226024628, + 0.00798429548740387, + 0.007032772991806269, + 0.00939376000314951, + 0.00968869123607874, + 0.009684482589364052, + 0.011410363018512726, + 0.008847774937748909, + 0.009647820144891739, + 0.002077902667224407, + 0.0028384251054376364, + 0.006592416204512119, + 0.010029221884906292, + 0.0030929604545235634, + 0.003693737555295229, + 0.00460535753518343, + 0.0020919893868267536, + 0.007593314163386822, + 0.006270622834563255, + 0.016177810728549957, + 0.02959621511399746, + 0.006554706487804651, + 0.005429128184914589, + 0.012145412154495716, + 0.013021880760788918, + 0.002412904752418399, + 0.0032217085827142, + 0.0042686848901212215, + 0.0032576043158769608, + 0.0033806145656853914, + 0.003105947282165289, + 0.005160162691026926, + 0.003418230451643467, + 0.003455568104982376, + 0.00573852239176631, + 0.0038777776062488556, + 0.004497443791478872, + 0.004237282555550337, + 0.003599903779104352, + 0.009543598629534245, + 0.005060857627540827, + 0.00131927989423275, + 0.0019842644687741995, + 0.0018648625118657947, + 0.002096228301525116, + 0.0019146885024383664, + 0.0018200146732851863, + 0.002505221636965871, + 0.001497869030572474, + 0.0064787534065544605, + 0.008498577401041985, + 0.007044549100100994, + 0.010627029463648796, + 0.010775587521493435, + 0.0058348048478364944, + 0.009050944820046425, + 0.006692169234156609, + 0.00674133887514472, + 0.02083510346710682, + 0.00938851572573185, + 0.013764184899628162, + 0.013136271387338638, + 0.005154576618224382, + 0.008864234201610088, + 0.006335625424981117, + 0.018757330253720284, + 0.014626147225499153, + 0.19556039571762085, + 0.023873567581176758, + 0.01042911596596241, + 0.017347630113363266, + 0.01527965534478426, + 0.19523979723453522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.020588846877217293 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic aerial time-lapse shot, captured from a high vantage point, surveys a vast expanse of golden desert, its rolling dunes casting sharp, long shadows under the scorching afternoon sun. As the camera slowly pulls back, a tiny point in the center of the frame undergoes a magical transformation. A vibrant emerald green radiates outward at a visible speed, seemingly devouring the surrounding yellow sand. First, tender green grass breaks through the sand, forming an ever-expanding circular carpet. Shortly thereafter, low shrubs and vibrant wildflowers quickly intersperse the landscape. Finally, towering palm trees and leafy broadleaf trees rise from the ground, their canopies unfurling in seconds. Simultaneously, a clear stream appears out of nowhere, winding through the new vegetation and converging into a shimmering central lake, its waters reflecting the azure sky. The entire expansion process is smooth and shocking, dramatically transforming the monotonous wasteland into a thriving ecosystem full of life, creating an atmosphere of grand, magical and awe-inspiring miracle of life.", + "seed": 12962, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_2/index_3650_seed_12962_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": -0.15881529520847834, + "vlm_logprob": -0.15881529520847834, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.15881529520847834, + "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 initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.0048020039685070515 + ], + "opposite_logprob_tokens": [ + -5.629802227020264 + ], + "logprob_reward": -0.0048020039685070515, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0048020039685070515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18719147145748138, + 0.05257994309067726, + 0.09382513165473938, + 0.13816142082214355, + 0.22349965572357178, + 0.054685063660144806, + 0.06513204425573349, + 0.03095345012843609, + 0.12569981813430786, + 0.02827203832566738 + ], + "credit_map": [ + 0.016670716926455498, + 0.010257948189973831, + 0.028747662901878357, + 0.026356292888522148, + 0.01930852234363556, + 0.030148912221193314, + 0.03433718532323837, + 0.021364208310842514, + 0.005088586360216141, + 0.006233762949705124, + 0.006194898393005133, + 0.006441507488489151, + 0.0055727893486619, + 0.009454895742237568, + 0.006272134371101856, + 0.0073213642463088036, + 0.007659859023988247, + 0.01358813140541315, + 0.011286605149507523, + 0.009665490128099918, + 0.018299199640750885, + 0.013209966011345387, + 0.007447032257914543, + 0.012668835930526257, + 0.00781803298741579, + 0.0117805702611804, + 0.02296275459229946, + 0.018095726147294044, + 0.011604066006839275, + 0.010185930877923965, + 0.011698228307068348, + 0.04401608556509018, + 0.018551234155893326, + 0.02219940721988678, + 0.048124928027391434, + 0.034533631056547165, + 0.018036218360066414, + 0.01787639781832695, + 0.032353319227695465, + 0.0318244993686676, + 0.004480311647057533, + 0.005904443562030792, + 0.009898858144879341, + 0.007753409445285797, + 0.005860119592398405, + 0.005727256648242474, + 0.007217346224933863, + 0.007843315601348877, + 0.005748356226831675, + 0.007475446909666061, + 0.010657984763383865, + 0.010481745935976505, + 0.0062936036847531796, + 0.007286299951374531, + 0.009969250299036503, + 0.007219349499791861, + 0.0022144964896142483, + 0.002544490620493889, + 0.0037658452056348324, + 0.0051420219242572784, + 0.0030108322389423847, + 0.005030957981944084, + 0.004153018817305565, + 0.005091782659292221, + 0.013099820353090763, + 0.0289476178586483, + 0.016249386593699455, + 0.01078453753143549, + 0.01277985144406557, + 0.021253550425171852, + 0.011731901206076145, + 0.010853143408894539, + 0.0020592128857970238, + 0.0036081294529139996, + 0.003110758727416396, + 0.003741856897249818, + 0.0028299791738390923, + 0.00530679477378726, + 0.003339864546433091, + 0.004275438375771046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0048020039685070515 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578887939453125 + ], + "opposite_logprob_tokens": [ + -0.6328887939453125 + ], + "logprob_reward": -0.7578887939453125, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578887939453125, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15976561605930328, + 0.038220278918743134, + 0.11960061639547348, + 0.10513405501842499, + 0.12907809019088745, + 0.061155155301094055, + 0.05834897607564926, + 0.09628356248140335, + 0.0798952728509903, + 0.15251831710338593 + ], + "credit_map": [ + 0.01417278777807951, + 0.015005198307335377, + 0.023528238758444786, + 0.01793608069419861, + 0.01730905845761299, + 0.030881185084581375, + 0.019894035533070564, + 0.021039040759205818, + 0.002837039064615965, + 0.00425903033465147, + 0.006909437943249941, + 0.005354930646717548, + 0.0031590554863214493, + 0.006807957775890827, + 0.004271984566003084, + 0.004620841704308987, + 0.009212233126163483, + 0.0148762296885252, + 0.012279563583433628, + 0.013031916692852974, + 0.0294592697173357, + 0.021286580711603165, + 0.009812370873987675, + 0.00964245293289423, + 0.005152881611138582, + 0.006117422133684158, + 0.017204105854034424, + 0.013822492212057114, + 0.006031491328030825, + 0.006248415447771549, + 0.011838041245937347, + 0.03871919587254524, + 0.005453009158372879, + 0.005251363385468721, + 0.04439452663064003, + 0.04785594344139099, + 0.004355633165687323, + 0.0036170154344290495, + 0.011447749100625515, + 0.006702844984829426, + 0.002566832583397627, + 0.0065381648018956184, + 0.018174290657043457, + 0.01584506966173649, + 0.004092146642506123, + 0.0034881101455539465, + 0.006518491078168154, + 0.00393205089494586, + 0.004467685706913471, + 0.006618895102292299, + 0.00850797537714243, + 0.009301820769906044, + 0.003957680426537991, + 0.006088231224566698, + 0.012593827210366726, + 0.006812861189246178, + 0.011583046987652779, + 0.0074689798057079315, + 0.01738065853714943, + 0.01465046126395464, + 0.00881643034517765, + 0.008682629093527794, + 0.015164881944656372, + 0.012536480091512203, + 0.0048116096295416355, + 0.00821561086922884, + 0.010026641190052032, + 0.0133743304759264, + 0.007460844703018665, + 0.008136330172419548, + 0.013135693967342377, + 0.01473421510308981, + 0.010001782327890396, + 0.01867240108549595, + 0.016728827729821205, + 0.029874539002776146, + 0.009992225095629692, + 0.020969955250620842, + 0.016958797350525856, + 0.02931978367269039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.7578887939453125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.0003912875254172832 + ], + "opposite_logprob_tokens": [ + -7.875391483306885 + ], + "logprob_reward": -0.0003912875254172832, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003912875254172832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0780879557132721, + 0.02582443505525589, + 0.05935155227780342, + 0.06700650602579117, + 0.09887503832578659, + 0.05003911256790161, + 0.06591451913118362, + 0.07240031659603119, + 0.3650350868701935, + 0.11746542155742645 + ], + "credit_map": [ + 0.0068315621465444565, + 0.009164266288280487, + 0.0098423445597291, + 0.007454216945916414, + 0.0073896958492696285, + 0.021374864503741264, + 0.007572808302938938, + 0.008458191528916359, + 0.002008341019973159, + 0.004951173439621925, + 0.002850189106538892, + 0.002745207631960511, + 0.0024971680250018835, + 0.005390768405050039, + 0.0021782158873975277, + 0.003203373169526458, + 0.004421767313033342, + 0.007777554914355278, + 0.006148096174001694, + 0.006245115306228399, + 0.013261607848107815, + 0.009420769289135933, + 0.0047622667625546455, + 0.007314376998692751, + 0.0031744472216814756, + 0.005139240063726902, + 0.0111553268507123, + 0.008987734094262123, + 0.004074317403137684, + 0.005047437269240618, + 0.006156047806143761, + 0.0232719536870718, + 0.0023122155107557774, + 0.003506945911794901, + 0.03695514053106308, + 0.04016425088047981, + 0.0020205534528940916, + 0.0018926127813756466, + 0.007877087220549583, + 0.004146229475736618, + 0.002917648758739233, + 0.004291304387152195, + 0.007379894144833088, + 0.011883588507771492, + 0.0036677198950201273, + 0.0031301588751375675, + 0.011059834621846676, + 0.005708965472877026, + 0.006199815776199102, + 0.009317637421190739, + 0.011413609609007835, + 0.011250870302319527, + 0.0067355213686823845, + 0.0059149498119950294, + 0.008111187256872654, + 0.0069709219969809055, + 0.007521630730479956, + 0.005657967180013657, + 0.015316884033381939, + 0.006183383986353874, + 0.00835086964070797, + 0.009030510671436787, + 0.007170157972723246, + 0.01316890586167574, + 0.03785993158817291, + 0.0975218117237091, + 0.04476283863186836, + 0.02865799143910408, + 0.038273777812719345, + 0.05350295454263687, + 0.03645322844386101, + 0.028002550825476646, + 0.00738596310839057, + 0.021832572296261787, + 0.015188165940344334, + 0.017526162788271904, + 0.009742253459990025, + 0.012911779806017876, + 0.015573903918266296, + 0.017304614186286926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.0003912875254172832 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.0046355403028428555 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.0046355403028428555, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046355403028428555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0695195347070694, + 0.02461320534348488, + 0.07350634038448334, + 0.04598328471183777, + 0.28595438599586487, + 0.024884682148694992, + 0.053348109126091, + 0.06953788548707962, + 0.27794143557548523, + 0.07471121102571487 + ], + "credit_map": [ + 0.0052081444300711155, + 0.006014146842062473, + 0.0076090870425105095, + 0.008470297791063786, + 0.0076595512218773365, + 0.01303822547197342, + 0.009061574004590511, + 0.01245851255953312, + 0.0025296462699770927, + 0.0030739966314285994, + 0.004442584700882435, + 0.0035790090914815664, + 0.002268184209242463, + 0.0038023621309548616, + 0.00212073908187449, + 0.0027966841589659452, + 0.005702347029000521, + 0.011012875474989414, + 0.007520092651247978, + 0.0077554769814014435, + 0.016166526824235916, + 0.010739495977759361, + 0.0056676361709833145, + 0.008941884152591228, + 0.0024578887969255447, + 0.0032525144051760435, + 0.007805518340319395, + 0.006704897154122591, + 0.0034682736732065678, + 0.0034124376252293587, + 0.005855882540345192, + 0.013025871478021145, + 0.005939417518675327, + 0.010197265073657036, + 0.11312548816204071, + 0.11657755821943283, + 0.005600308999419212, + 0.005535243544727564, + 0.019191432744264603, + 0.009787687100470066, + 0.0019755468238145113, + 0.003173321019858122, + 0.003301930380985141, + 0.0044698165729641914, + 0.0026421642396599054, + 0.002129415748640895, + 0.0035190251655876637, + 0.00367346266284585, + 0.005479184910655022, + 0.00789735745638609, + 0.007232931908220053, + 0.008871383033692837, + 0.005045102443546057, + 0.005591009743511677, + 0.0076858424581587315, + 0.0055452981032431126, + 0.006124265491962433, + 0.00548273092135787, + 0.010232146829366684, + 0.00928493496030569, + 0.011122333817183971, + 0.009409273974597454, + 0.007426643744111061, + 0.010455555282533169, + 0.01489214226603508, + 0.0245117899030447, + 0.020504269748926163, + 0.06822460889816284, + 0.034293342381715775, + 0.0222774725407362, + 0.027310620993375778, + 0.0659271702170372, + 0.006190595217049122, + 0.0076436372473835945, + 0.008845100179314613, + 0.011516992934048176, + 0.005797774530947208, + 0.014061125926673412, + 0.007404384668916464, + 0.013251597993075848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.0046355403028428555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.026358850300312042 + ], + "opposite_logprob_tokens": [ + -3.6513588428497314 + ], + "logprob_reward": -0.026358850300312042, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358850300312042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09801285713911057, + 0.03698430210351944, + 0.11974459141492844, + 0.04784534126520157, + 0.1003611832857132, + 0.07243598997592926, + 0.05506404861807823, + 0.12803322076797485, + 0.1932404637336731, + 0.14827796816825867 + ], + "credit_map": [ + 0.010719938203692436, + 0.00863408762961626, + 0.013403182849287987, + 0.016004426404833794, + 0.010331586003303528, + 0.013347452506422997, + 0.00937135424464941, + 0.016200829297304153, + 0.003326319856569171, + 0.003670618636533618, + 0.007207521237432957, + 0.006138009484857321, + 0.003927127458155155, + 0.0053049540147185326, + 0.0037067595403641462, + 0.0037029937375336885, + 0.008231726475059986, + 0.01602756232023239, + 0.011936875060200691, + 0.013178754597902298, + 0.029176635667681694, + 0.01946047507226467, + 0.00939817912876606, + 0.012334386818110943, + 0.004191362764686346, + 0.0035083447583019733, + 0.00621140468865633, + 0.007532992400228977, + 0.0034405477344989777, + 0.0036969915963709354, + 0.006488118786364794, + 0.012775575742125511, + 0.00400960398837924, + 0.004095857031643391, + 0.032628633081912994, + 0.03292030096054077, + 0.004771078005433083, + 0.0027461734134703875, + 0.010400520637631416, + 0.008789010345935822, + 0.005414934363216162, + 0.009512576274573803, + 0.009261643514037132, + 0.008767202496528625, + 0.00887447688728571, + 0.005024787504225969, + 0.018273908644914627, + 0.0073064640164375305, + 0.0045885383151471615, + 0.005167353432625532, + 0.007407094351947308, + 0.01399959996342659, + 0.0041334484703838825, + 0.004117215052247047, + 0.009987368248403072, + 0.0056634326465427876, + 0.012831233441829681, + 0.015878528356552124, + 0.01514950580894947, + 0.013411757536232471, + 0.02011064440011978, + 0.021732274442911148, + 0.014094616286456585, + 0.014824659563601017, + 0.014460480771958828, + 0.03483384847640991, + 0.018993934616446495, + 0.01775902323424816, + 0.03608327731490135, + 0.025708317756652832, + 0.0274057537317276, + 0.017995838075876236, + 0.011035865172743797, + 0.013946395367383957, + 0.01796010509133339, + 0.020524874329566956, + 0.014066931791603565, + 0.03289172798395157, + 0.01334434188902378, + 0.024507716298103333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.026358850300312042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic aerial time-lapse shot, captured from a high vantage point, surveys a vast expanse of golden desert, its rolling dunes casting sharp, long shadows under the scorching afternoon sun. As the camera slowly pulls back, a tiny point in the center of the frame undergoes a magical transformation. A vibrant emerald green radiates outward at a visible speed, seemingly devouring the surrounding yellow sand. First, tender green grass breaks through the sand, forming an ever-expanding circular carpet. Shortly thereafter, low shrubs and vibrant wildflowers quickly intersperse the landscape. Finally, towering palm trees and leafy broadleaf trees rise from the ground, their canopies unfurling in seconds. Simultaneously, a clear stream appears out of nowhere, winding through the new vegetation and converging into a shimmering central lake, its waters reflecting the azure sky. The entire expansion process is smooth and shocking, dramatically transforming the monotonous wasteland into a thriving ecosystem full of life, creating an atmosphere of grand, magical and awe-inspiring miracle of life.", + "seed": 12963, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_2/index_3650_seed_12963_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": -0.08119837495032697, + "vlm_logprob": -0.08119837495032697, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.08119837495032697, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.002755299676209688 + ], + "opposite_logprob_tokens": [ + -7.127755165100098 + ], + "logprob_reward": -0.002755299676209688, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.002755299676209688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3095812499523163, + 0.10042083263397217, + 0.07894878834486008, + 0.0744694247841835, + 0.06277862191200256, + 0.09350332617759705, + 0.061742328107357025, + 0.07893981039524078, + 0.0872102677822113, + 0.05240532010793686 + ], + "credit_map": [ + 0.026847675442695618, + 0.03807704895734787, + 0.07131301611661911, + 0.03824521601200104, + 0.027703270316123962, + 0.03792466223239899, + 0.03215118125081062, + 0.03731921315193176, + 0.009087292477488518, + 0.006370891816914082, + 0.021556518971920013, + 0.01755952648818493, + 0.010071912780404091, + 0.011647309176623821, + 0.011556283570826054, + 0.012571101076900959, + 0.0049676415510475636, + 0.005217396188527346, + 0.012831051833927631, + 0.012066969648003578, + 0.007134046871215105, + 0.006689972709864378, + 0.018916398286819458, + 0.011125312186777592, + 0.004170689731836319, + 0.0057382541708648205, + 0.014173141680657864, + 0.01102987490594387, + 0.010159912519156933, + 0.0070704189129173756, + 0.012336541898548603, + 0.00979059748351574, + 0.003345624078065157, + 0.006255829241126776, + 0.009633587673306465, + 0.012088540941476822, + 0.0062151579186320305, + 0.0066675725392997265, + 0.009793267585337162, + 0.00877904798835516, + 0.013326170854270458, + 0.013758696615695953, + 0.014485610648989677, + 0.010893060825765133, + 0.010329426266252995, + 0.009684868156909943, + 0.01057338621467352, + 0.01045211311429739, + 0.00638841837644577, + 0.010985424742102623, + 0.00853990949690342, + 0.007427604869008064, + 0.005914886482059956, + 0.008462339639663696, + 0.005770299583673477, + 0.008253447711467743, + 0.010184423066675663, + 0.016706524416804314, + 0.010273562744259834, + 0.0036802326794713736, + 0.00584446731954813, + 0.019758671522140503, + 0.004754082765430212, + 0.00773785263299942, + 0.00865121465176344, + 0.02269877679646015, + 0.015156318433582783, + 0.005657947156578302, + 0.007318093907088041, + 0.012860449030995369, + 0.00658908486366272, + 0.008278390392661095, + 0.003979666158556938, + 0.005890196654945612, + 0.00940709002315998, + 0.005465925205498934, + 0.012364204972982407, + 0.0036498988047242165, + 0.006323664914816618, + 0.0053246766328811646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.002755299676209688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.38721343874931335 + ], + "opposite_logprob_tokens": [ + -1.1372134685516357 + ], + "logprob_reward": -0.38721343874931335, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38721343874931335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17817145586013794, + 0.08645851165056229, + 0.06800100952386856, + 0.0517159141600132, + 0.07407553493976593, + 0.07318553328514099, + 0.07575953006744385, + 0.18938516080379486, + 0.06880175322294235, + 0.13444560766220093 + ], + "credit_map": [ + 0.012193193659186363, + 0.022077951580286026, + 0.02798818238079548, + 0.024884507060050964, + 0.018895993009209633, + 0.024730151519179344, + 0.02645234204828739, + 0.020949121564626694, + 0.007031314540654421, + 0.007961343973875046, + 0.014273677952587605, + 0.013004555366933346, + 0.008059213869273663, + 0.010685188695788383, + 0.0123886214569211, + 0.01305459812283516, + 0.005309076514095068, + 0.007428858894854784, + 0.012638341635465622, + 0.00927451066672802, + 0.007260923273861408, + 0.005530213937163353, + 0.011618820950388908, + 0.00894027017056942, + 0.005660511553287506, + 0.004271212965250015, + 0.010145541280508041, + 0.0069440933875739574, + 0.004337938968092203, + 0.00389108806848526, + 0.00838566105812788, + 0.008079863153398037, + 0.0049469005316495895, + 0.009084552526473999, + 0.012125441804528236, + 0.008064888417720795, + 0.007788202725350857, + 0.008206472732126713, + 0.009868040680885315, + 0.01399103831499815, + 0.006976628210395575, + 0.010035853832960129, + 0.01196574792265892, + 0.005868158303201199, + 0.011310325004160404, + 0.01161606889218092, + 0.007880798541009426, + 0.007531955372542143, + 0.007005711551755667, + 0.011815693229436874, + 0.008272886276245117, + 0.012928216718137264, + 0.006885374430567026, + 0.012019309215247631, + 0.00793919712305069, + 0.008893142454326153, + 0.029028192162513733, + 0.04932577908039093, + 0.023993363603949547, + 0.01509317196905613, + 0.012667056173086166, + 0.028689123690128326, + 0.011743955314159393, + 0.018844522535800934, + 0.010922488756477833, + 0.011904466897249222, + 0.008982408791780472, + 0.0070629543624818325, + 0.005601832177489996, + 0.008878454566001892, + 0.006192547734826803, + 0.009256595745682716, + 0.0085782241076231, + 0.010768606327474117, + 0.041574928909540176, + 0.006469713058322668, + 0.020044274628162384, + 0.0072551304474473, + 0.03139825165271759, + 0.008356471545994282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.38721343874931335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.0003070359816774726 + ], + "opposite_logprob_tokens": [ + -8.125307083129883 + ], + "logprob_reward": -0.0003070359816774726, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0003070359816774726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11433779448270798, + 0.03855103999376297, + 0.05855582281947136, + 0.019120296463370323, + 0.038557518273591995, + 0.06015950068831444, + 0.029665693640708923, + 0.055259816348552704, + 0.35478368401527405, + 0.2310088872909546 + ], + "credit_map": [ + 0.00710487924516201, + 0.008017472922801971, + 0.01693478412926197, + 0.016715075820684433, + 0.00715785613283515, + 0.010230282321572304, + 0.023035135120153427, + 0.02514229714870453, + 0.0027279171627014875, + 0.003017127513885498, + 0.009323441423475742, + 0.005726293195039034, + 0.005003843456506729, + 0.003910051193088293, + 0.004549079108983278, + 0.004293282516300678, + 0.0027511355001479387, + 0.004796737805008888, + 0.0057525173760950565, + 0.006862184498459101, + 0.004050272982567549, + 0.004574383143335581, + 0.016796618700027466, + 0.012971962802112103, + 0.001164706889539957, + 0.0015845703892409801, + 0.003170978743582964, + 0.00250695226714015, + 0.0020735247526317835, + 0.0031209310982376337, + 0.003249453380703926, + 0.0022491752170026302, + 0.001869970466941595, + 0.004625534638762474, + 0.005331433843821287, + 0.007642318494617939, + 0.0036688465625047684, + 0.005106484051793814, + 0.005365550052374601, + 0.004947375506162643, + 0.00431531947106123, + 0.007717766799032688, + 0.009096642956137657, + 0.0085989311337471, + 0.007355585694313049, + 0.0073593659326434135, + 0.008438250981271267, + 0.007277628406882286, + 0.0022854001726955175, + 0.006053507328033447, + 0.0029961580876260996, + 0.003937923815101385, + 0.004446379374712706, + 0.004742117132991552, + 0.0020144011359661818, + 0.0031898028682917356, + 0.005599165800958872, + 0.009087879210710526, + 0.005242824554443359, + 0.0054021370597183704, + 0.003972315229475498, + 0.013221454806625843, + 0.004473702982068062, + 0.0082603320479393, + 0.031325049698352814, + 0.10861282795667648, + 0.0807071328163147, + 0.01782301999628544, + 0.0282694473862648, + 0.034996528178453445, + 0.021211277693510056, + 0.031838368624448776, + 0.008598349057137966, + 0.02389048971235752, + 0.01657966710627079, + 0.008023210801184177, + 0.0814756229519844, + 0.015459680929780006, + 0.06260552257299423, + 0.014376325532793999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0003070359816774726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.0015124559868127108 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015124559868127108, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015124559868127108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14172564446926117, + 0.04574328660964966, + 0.060209374874830246, + 0.02411341294646263, + 0.05298002064228058, + 0.08220415562391281, + 0.08030354231595993, + 0.05231669917702675, + 0.21400953829288483, + 0.24639426171779633 + ], + "credit_map": [ + 0.01130749937146902, + 0.015324254520237446, + 0.03186997398734093, + 0.01483506802469492, + 0.01310508418828249, + 0.016451319679617882, + 0.016850313171744347, + 0.021982137113809586, + 0.0036949978675693274, + 0.0035734111443161964, + 0.007887572981417179, + 0.007248949259519577, + 0.005396266467869282, + 0.007706894539296627, + 0.005147636402398348, + 0.005087558180093765, + 0.0032913314644247293, + 0.006235317327082157, + 0.006620905827730894, + 0.007617258001118898, + 0.004982937127351761, + 0.004936323035508394, + 0.01568029448390007, + 0.010845007374882698, + 0.001986690564081073, + 0.0024116872809827328, + 0.003597530070692301, + 0.003438919549807906, + 0.002729241969063878, + 0.003173484932631254, + 0.0033489526249468327, + 0.0034269047901034355, + 0.004136293660849333, + 0.005226540379226208, + 0.006781372241675854, + 0.00748246256262064, + 0.006605956703424454, + 0.010310770943760872, + 0.006879206746816635, + 0.005557420663535595, + 0.006853960454463959, + 0.010342567227780819, + 0.019403917714953423, + 0.009921371936798096, + 0.007255436386913061, + 0.008946571499109268, + 0.010617958381772041, + 0.008862370625138283, + 0.005846664775162935, + 0.018737487494945526, + 0.007113840896636248, + 0.012103701010346413, + 0.012998994439840317, + 0.008878156542778015, + 0.006724136415868998, + 0.007900560274720192, + 0.0058208610862493515, + 0.009887241758406162, + 0.00613003596663475, + 0.004084577783942223, + 0.004978295881301165, + 0.010287721641361713, + 0.0034281727857887745, + 0.007699795067310333, + 0.024821067228913307, + 0.05862552300095558, + 0.05092708393931389, + 0.019558265805244446, + 0.014699200168251991, + 0.013032784685492516, + 0.017481187358498573, + 0.014864424243569374, + 0.011794541962444782, + 0.028260501101613045, + 0.07075680047273636, + 0.011614061892032623, + 0.052351437509059906, + 0.014106345362961292, + 0.04045630618929863, + 0.017054270952939987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0015124559868127108 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene a vast expanse of golden desert with rolling dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start from a high vantage point and slowly pull back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green radiate outward from the center, causing grass, shrubs, wildflowers, and trees to grow rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a clear stream appear and wind through the new vegetation to form a shimmering central lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the lake's waters reflecting an azure 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": [ + -0.01420364435762167 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.01420364435762167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420364435762167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0556194968521595, + 0.03134681284427643, + 0.03341922163963318, + 0.015066919848322868, + 0.029346762225031853, + 0.028364790603518486, + 0.01652185246348381, + 0.008720665238797665, + 0.049795810133218765, + 0.7317976355552673 + ], + "credit_map": [ + 0.0036176301073282957, + 0.005041889846324921, + 0.006903635337948799, + 0.004455499351024628, + 0.004634644836187363, + 0.005140083376318216, + 0.011806940659880638, + 0.014019178226590157, + 0.002062247833237052, + 0.0021394342184066772, + 0.003967904020100832, + 0.003984863869845867, + 0.003162396140396595, + 0.0038880843203514814, + 0.006570661906152964, + 0.005571222398430109, + 0.001240396173670888, + 0.0015641923528164625, + 0.003140814136713743, + 0.004604918882250786, + 0.0021881265565752983, + 0.0017876296769827604, + 0.009818724356591702, + 0.009074417874217033, + 0.002141233766451478, + 0.0017394308233633637, + 0.002933642128482461, + 0.0025567361153662205, + 0.0012312685139477253, + 0.0012656995095312595, + 0.001617926056496799, + 0.0015809834003448486, + 0.0024461322464048862, + 0.002823059679940343, + 0.0037484823260456324, + 0.0037261827383190393, + 0.003854637034237385, + 0.00531152356415987, + 0.004198848269879818, + 0.0032378975301980972, + 0.0026925764977931976, + 0.004902074579149485, + 0.005663441028445959, + 0.0026788420509546995, + 0.002982313511893153, + 0.0026904784608632326, + 0.0027234696317464113, + 0.004031593911349773, + 0.00160796579439193, + 0.0031265404541045427, + 0.0017398473573848605, + 0.0017962345154955983, + 0.0022202059626579285, + 0.0029647930059581995, + 0.0012492523528635502, + 0.0018170124385505915, + 0.0008118821424432099, + 0.001023353892378509, + 0.0013657184317708015, + 0.0008429557783529162, + 0.0007304835598915815, + 0.0019343363819643855, + 0.0008835689513944089, + 0.0011283664498478174, + 0.004935944452881813, + 0.014866985380649567, + 0.011314800940454006, + 0.0030844667926430702, + 0.00370256369933486, + 0.004458706825971603, + 0.0036208208184689283, + 0.0038115191273391247, + 0.03196791931986809, + 0.07146964222192764, + 0.21895073354244232, + 0.030648214742541313, + 0.21349208056926727, + 0.043837327510118484, + 0.07422465831041336, + 0.047207072377204895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01420364435762167 + } + ], + "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/0000052/group_3/index_tensor([3134, 3134])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_3/index_tensor([3134, 3134])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e5ba4c8efff9a7bed42490cf92908914a56890ef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_3/index_tensor([3134, 3134])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot captures a flowing, off-white, sheer silk curtain hung in front of a large, open window in a minimalist interior. Beneath it, the polished, deep espresso brown hardwood floor gleams with a high-gloss, mirror-like finish, reflecting the soft ambient light. The entire scene is bathed in bright, diffused natural morning light, creating a tranquil, high-key atmosphere. A gentle, yet deliberate breeze energizes the fabric, causing it to billow gracefully inward. The delicate silk expands and ripples, creating a series of fluid, undulating waves that dance briefly in the air. As the curtain moves, a sharp, brilliant streak of golden sunlight slants across the dark wooden floor, momentarily illuminating the tiny dust particles suspended in its path. This moment is fleeting; as the wind subsides, the curtain slowly and weightlessly returns to its original position, and the sharp streak of sunlight on the floor disappears as quickly as it appeared, leaving behind only a soft, even glow. The entire sequence unfolds with a quiet, poetic rhythm, enhancing the elegant, tranquil, and minimalist atmosphere of the space.", + "seed": 22061, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_3/index_3134_seed_22061_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": -0.040791518054902554, + "vlm_logprob": -0.040791518054902554, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.040791518054902554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008686495013535023 + ], + "opposite_logprob_tokens": [ + -4.883686542510986 + ], + "logprob_reward": -0.008686495013535023, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008686495013535023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22545869648456573, + 0.24392582476139069, + 0.02544696256518364, + 0.019769497215747833, + 0.02671101875603199, + 0.17601899802684784, + 0.12579698860645294, + 0.0955917164683342, + 0.03802480548620224, + 0.023255417123436928 + ], + "credit_map": [ + 0.0187233816832304, + 0.07117653638124466, + 0.012686329893767834, + 0.018259119242429733, + 0.06152459606528282, + 0.01690431870520115, + 0.013601768761873245, + 0.012582636438310146, + 0.009678714908659458, + 0.09624578803777695, + 0.006508252117782831, + 0.005802283063530922, + 0.10208795964717865, + 0.010740382596850395, + 0.005498972721397877, + 0.007363479118794203, + 0.0033045418094843626, + 0.0032241421286016703, + 0.0038870107382535934, + 0.0024056818801909685, + 0.0030376468785107136, + 0.004060454200953245, + 0.001695993822067976, + 0.0038314901757985353, + 0.0023903404362499714, + 0.0023145482409745455, + 0.0031569097191095352, + 0.002196740126237273, + 0.0026758338790386915, + 0.002376838121563196, + 0.0017701571341603994, + 0.002888129558414221, + 0.002992333611473441, + 0.005655978806316853, + 0.00287318741902709, + 0.0015545766800642014, + 0.00632321834564209, + 0.0032753197010606527, + 0.0018466650508344173, + 0.002189739141613245, + 0.017061764374375343, + 0.025100689381361008, + 0.017928702756762505, + 0.01629260554909706, + 0.03568987920880318, + 0.028079500421881676, + 0.01116503681987524, + 0.024700822308659554, + 0.018519937992095947, + 0.01724657602608204, + 0.01743309386074543, + 0.008220522664487362, + 0.0259600467979908, + 0.01605808734893799, + 0.006718180608004332, + 0.015640554949641228, + 0.008702916093170643, + 0.014505906961858273, + 0.0047540608793497086, + 0.0020181562285870314, + 0.04036366939544678, + 0.02086908370256424, + 0.0018655345775187016, + 0.0025123865343630314, + 0.0060427226126194, + 0.005251044407486916, + 0.0034744085278362036, + 0.0029070947784930468, + 0.009359286166727543, + 0.005549093708395958, + 0.0021505597978830338, + 0.0032905982807278633, + 0.003974366467446089, + 0.0029510168824344873, + 0.002671200083568692, + 0.0032485725823789835, + 0.003009213600307703, + 0.002338988706469536, + 0.0024387803860008717, + 0.00262327934615314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.008686495013535023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036250948905945 + ], + "opposite_logprob_tokens": [ + -2.350362539291382 + ], + "logprob_reward": -0.10036250948905945, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036250948905945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10168196260929108, + 0.25598660111427307, + 0.01401427760720253, + 0.007939361967146397, + 0.030177906155586243, + 0.031485624611377716, + 0.01264582946896553, + 0.42724189162254333, + 0.06351649761199951, + 0.05531010404229164 + ], + "credit_map": [ + 0.010820367373526096, + 0.020520031452178955, + 0.012985517270863056, + 0.009021230973303318, + 0.024883611127734184, + 0.00820185523480177, + 0.00773407518863678, + 0.007515264209359884, + 0.009593340568244457, + 0.10635553300380707, + 0.00456235883757472, + 0.003979160916060209, + 0.11140306293964386, + 0.011063267476856709, + 0.0041703032329678535, + 0.004859562497586012, + 0.0016326057957485318, + 0.0020213399548083544, + 0.0023393614683300257, + 0.0015961205353960395, + 0.0018040527356788516, + 0.001642731949687004, + 0.001197240431793034, + 0.0017808245029300451, + 0.0009125115466304123, + 0.002031862735748291, + 0.0009613280999474227, + 0.0007643940625712276, + 0.001195025397464633, + 0.0009822703432291746, + 0.00045351978042162955, + 0.0006384489242918789, + 0.0023463943507522345, + 0.007719927467405796, + 0.0032070553861558437, + 0.0019012812990695238, + 0.008665736764669418, + 0.0021802971605211496, + 0.0013348781503736973, + 0.002822332317009568, + 0.003023226745426655, + 0.004236959386616945, + 0.0034320135600864887, + 0.003352512139827013, + 0.006093811709433794, + 0.004824732430279255, + 0.0019183089025318623, + 0.004604056943207979, + 0.001519199926406145, + 0.0012768157757818699, + 0.0018329204758629203, + 0.0012036000844091177, + 0.0016833595000207424, + 0.0014394843019545078, + 0.0010365727357566357, + 0.002653876319527626, + 0.03477330505847931, + 0.16113458573818207, + 0.011070300824940205, + 0.009112873114645481, + 0.03681228309869766, + 0.1490316092967987, + 0.009241420775651932, + 0.01606549695134163, + 0.015175405889749527, + 0.009178023785352707, + 0.005399390589445829, + 0.004155724309384823, + 0.009457557462155819, + 0.0108600789681077, + 0.004369084723293781, + 0.004921225365251303, + 0.010598649270832539, + 0.009403576143085957, + 0.00512215681374073, + 0.004070469178259373, + 0.006797808688133955, + 0.009116776287555695, + 0.00418544327840209, + 0.006015218328684568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.10036250948905945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866420477628708 + ], + "opposite_logprob_tokens": [ + -3.048664093017578 + ], + "logprob_reward": -0.04866420477628708, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866420477628708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2480613887310028, + 0.4622156023979187, + 0.01814820058643818, + 0.02094547264277935, + 0.021641777828335762, + 0.049508266150951385, + 0.03758437931537628, + 0.059227105230093, + 0.05406202748417854, + 0.028605788946151733 + ], + "credit_map": [ + 0.01739921048283577, + 0.08433595299720764, + 0.011561994440853596, + 0.016359541565179825, + 0.07875656336545944, + 0.01668793521821499, + 0.012595422565937042, + 0.010364754125475883, + 0.019079681485891342, + 0.17676270008087158, + 0.015064995735883713, + 0.013918320648372173, + 0.1858590990304947, + 0.021060489118099213, + 0.011950054205954075, + 0.018520256504416466, + 0.0028420116286724806, + 0.0023739133030176163, + 0.002496921457350254, + 0.0017848190618678927, + 0.002737657865509391, + 0.0022022912744432688, + 0.0014077435480430722, + 0.0023028419818729162, + 0.002228343393653631, + 0.0028844757471233606, + 0.0032524995040148497, + 0.0024625060614198446, + 0.0025727387983351946, + 0.002776441164314747, + 0.0015830392949283123, + 0.0031854265835136175, + 0.001477150828577578, + 0.005138715263456106, + 0.0015496957348659635, + 0.0011131154606118798, + 0.007024635560810566, + 0.002110274275764823, + 0.0011364861857146025, + 0.0020917030051350594, + 0.005461031571030617, + 0.007365095429122448, + 0.0053734309040009975, + 0.003265094244852662, + 0.011156134307384491, + 0.008506000973284245, + 0.0036643394269049168, + 0.004717134404927492, + 0.00545046990737319, + 0.005122437607496977, + 0.006659721489995718, + 0.0026000584475696087, + 0.005841356702148914, + 0.004136164207011461, + 0.0031399494037032127, + 0.004634220153093338, + 0.0037571496795862913, + 0.008533447980880737, + 0.002501674462109804, + 0.0012898193672299385, + 0.02507634274661541, + 0.014806356281042099, + 0.0014973515644669533, + 0.0017649609362706542, + 0.008614483289420605, + 0.00750957103446126, + 0.005056881345808506, + 0.005004420876502991, + 0.011565846391022205, + 0.008630183525383472, + 0.003629897255450487, + 0.004050743766129017, + 0.007076866924762726, + 0.002736948197707534, + 0.002766634104773402, + 0.0024450302589684725, + 0.0034477964509278536, + 0.003518122946843505, + 0.003044274402782321, + 0.003570111468434334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.04866420477628708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11845682561397552, + 0.2760985791683197, + 0.1149977520108223, + 0.014257988892495632, + 0.06338793784379959, + 0.05287931114435196, + 0.07548104226589203, + 0.1577291339635849, + 0.07030440121889114, + 0.05640704184770584 + ], + "credit_map": [ + 0.013681430369615555, + 0.022978000342845917, + 0.00938168540596962, + 0.011582759208977222, + 0.02580568753182888, + 0.014691298827528954, + 0.010427662171423435, + 0.009908291511237621, + 0.01131172850728035, + 0.10442747175693512, + 0.010098584927618504, + 0.007470909971743822, + 0.11193813383579254, + 0.013295162469148636, + 0.00646133441478014, + 0.011095241643488407, + 0.01278853788971901, + 0.014109806157648563, + 0.01865018904209137, + 0.015628602355718613, + 0.01474054716527462, + 0.0094089200720191, + 0.012718881480395794, + 0.016952255740761757, + 0.0021167632658034563, + 0.0016291373176500201, + 0.0023785613011568785, + 0.00119927863124758, + 0.0021564848721027374, + 0.0014406931586563587, + 0.001235166098922491, + 0.002101902151480317, + 0.0059148636646568775, + 0.01625906117260456, + 0.0031269420869648457, + 0.0027931828517466784, + 0.023481009528040886, + 0.005906610284000635, + 0.0026766122318804264, + 0.003229646710678935, + 0.0054082260467112064, + 0.007398761808872223, + 0.005252353381365538, + 0.004913242533802986, + 0.010194842703640461, + 0.00808767881244421, + 0.00387443695217371, + 0.0077497633174061775, + 0.009829293936491013, + 0.012660565786063671, + 0.009660028852522373, + 0.005532237235456705, + 0.013592506758868694, + 0.011048748157918453, + 0.004325886722654104, + 0.008831769227981567, + 0.006245729047805071, + 0.018071051687002182, + 0.004439173266291618, + 0.0036596504505723715, + 0.0563361681997776, + 0.061039142310619354, + 0.003017476527020335, + 0.0049207271076738834, + 0.008434670977294445, + 0.011913674883544445, + 0.004635467194020748, + 0.00408369954675436, + 0.02145276963710785, + 0.012259860523045063, + 0.003942401148378849, + 0.0035818491596728563, + 0.0060483673587441444, + 0.007630824577063322, + 0.00484815239906311, + 0.00551265012472868, + 0.011165349744260311, + 0.009798121638596058, + 0.005679059773683548, + 0.005724508315324783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "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 a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699020743370056 + ], + "opposite_logprob_tokens": [ + -3.4086990356445312 + ], + "logprob_reward": -0.033699020743370056, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699020743370056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12935452163219452, + 0.2975509464740753, + 0.0175912044942379, + 0.009456411935389042, + 0.015908872708678246, + 0.026578061282634735, + 0.024646127596497536, + 0.4291051924228668, + 0.018982870504260063, + 0.030825812369585037 + ], + "credit_map": [ + 0.010152378119528294, + 0.03702893480658531, + 0.008059417828917503, + 0.010647516697645187, + 0.03610389679670334, + 0.012928653508424759, + 0.00706561841070652, + 0.007368093356490135, + 0.010301951318979263, + 0.12422186881303787, + 0.0040551344864070415, + 0.003705662675201893, + 0.13187620043754578, + 0.01499105803668499, + 0.004002650734037161, + 0.004396399948745966, + 0.0024066318292170763, + 0.003375038504600525, + 0.0027203611098229885, + 0.0020387752447277308, + 0.002229344565421343, + 0.0018921131268143654, + 0.0012364444555714726, + 0.0016924955416470766, + 0.0012647417606785893, + 0.00220268196426332, + 0.0010020977351814508, + 0.0008863331167958677, + 0.0014099999098107219, + 0.0012760248500853777, + 0.0005481153493747115, + 0.0008664167835377157, + 0.0015429522609338164, + 0.003674492472782731, + 0.0014410470612347126, + 0.0009306026040576398, + 0.00455509265884757, + 0.001838795025832951, + 0.000729296647477895, + 0.0011965949088335037, + 0.003881644457578659, + 0.003251145826652646, + 0.0030134241096675396, + 0.002272806828841567, + 0.004790014587342739, + 0.0029528019949793816, + 0.0026511065661907196, + 0.0037651173770427704, + 0.003235420910641551, + 0.003830195404589176, + 0.0025859219022095203, + 0.0020643766038119793, + 0.0047096326015889645, + 0.0032783583737909794, + 0.0015251688892021775, + 0.003417053958401084, + 0.03164152055978775, + 0.1662687361240387, + 0.01004792284220457, + 0.008038864471018314, + 0.048417408019304276, + 0.14067254960536957, + 0.008794358931481838, + 0.015223849564790726, + 0.0036707990802824497, + 0.00315444590523839, + 0.0015679089119657874, + 0.0012365690199658275, + 0.003972437232732773, + 0.0027053006924688816, + 0.0011957463575527072, + 0.001479662605561316, + 0.006942762993276119, + 0.004820997826755047, + 0.002178889000788331, + 0.0021427455358207226, + 0.005963845178484917, + 0.004285043105483055, + 0.002094762399792671, + 0.0023967681918293238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.033699020743370056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot captures a flowing, off-white, sheer silk curtain hung in front of a large, open window in a minimalist interior. Beneath it, the polished, deep espresso brown hardwood floor gleams with a high-gloss, mirror-like finish, reflecting the soft ambient light. The entire scene is bathed in bright, diffused natural morning light, creating a tranquil, high-key atmosphere. A gentle, yet deliberate breeze energizes the fabric, causing it to billow gracefully inward. The delicate silk expands and ripples, creating a series of fluid, undulating waves that dance briefly in the air. As the curtain moves, a sharp, brilliant streak of golden sunlight slants across the dark wooden floor, momentarily illuminating the tiny dust particles suspended in its path. This moment is fleeting; as the wind subsides, the curtain slowly and weightlessly returns to its original position, and the sharp streak of sunlight on the floor disappears as quickly as it appeared, leaving behind only a soft, even glow. The entire sequence unfolds with a quiet, poetic rhythm, enhancing the elegant, tranquil, and minimalist atmosphere of the space.", + "seed": 22062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_3/index_3134_seed_22062_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": -0.05965949073433876, + "vlm_logprob": -0.05965949073433876, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.05965949073433876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013729188591241837 + ], + "opposite_logprob_tokens": [ + -4.388729095458984 + ], + "logprob_reward": -0.013729188591241837, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013729188591241837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26099875569343567, + 0.17014633119106293, + 0.07234127819538116, + 0.07200046628713608, + 0.11215493828058243, + 0.04801734909415245, + 0.11727897822856903, + 0.05173434317111969, + 0.043495628982782364, + 0.051831986755132675 + ], + "credit_map": [ + 0.027364108711481094, + 0.038259394466876984, + 0.026525890454649925, + 0.030060850083827972, + 0.05920550599694252, + 0.02794751338660717, + 0.028026588261127472, + 0.023608865216374397, + 0.017545511946082115, + 0.026869017630815506, + 0.019606124609708786, + 0.01559509802609682, + 0.03957737237215042, + 0.02056594006717205, + 0.01245683804154396, + 0.0179304126650095, + 0.008716017939150333, + 0.008421029895544052, + 0.007036152761429548, + 0.00821898877620697, + 0.011416424997150898, + 0.008775766007602215, + 0.013564154505729675, + 0.006192736327648163, + 0.008978006429970264, + 0.010790135711431503, + 0.005807843059301376, + 0.0065145231783390045, + 0.01230471208691597, + 0.013637191615998745, + 0.009222677908837795, + 0.0047453721053898335, + 0.014127613045275211, + 0.026491980999708176, + 0.008143933489918709, + 0.00909243244677782, + 0.02769743837416172, + 0.009418265894055367, + 0.007520012091845274, + 0.009663254022598267, + 0.006145381834357977, + 0.006954062730073929, + 0.0052920132875442505, + 0.005735538899898529, + 0.006686259992420673, + 0.006686655804514885, + 0.0041634575463831425, + 0.006353974808007479, + 0.013244347646832466, + 0.02282077446579933, + 0.011764046736061573, + 0.01351628452539444, + 0.016253404319286346, + 0.01981642097234726, + 0.009541764855384827, + 0.010321924462914467, + 0.0074117667973041534, + 0.0058448296040296555, + 0.006602118723094463, + 0.00721701281145215, + 0.007183175068348646, + 0.006013772916048765, + 0.0060315243899822235, + 0.005430140066891909, + 0.0044994656927883625, + 0.005656470078974962, + 0.00523734325543046, + 0.005599434021860361, + 0.007811093702912331, + 0.0071531557478010654, + 0.004060870967805386, + 0.003477792488411069, + 0.00457004364579916, + 0.013355817645788193, + 0.002529639983549714, + 0.003073285100981593, + 0.01689017005264759, + 0.004531953949481249, + 0.0032166552264243364, + 0.003664417425170541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.013729188591241837 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869871959090233 + ], + "opposite_logprob_tokens": [ + -3.048698663711548 + ], + "logprob_reward": -0.04869871959090233, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869871959090233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2154935896396637, + 0.11399567127227783, + 0.0804772824048996, + 0.06354282051324844, + 0.11696752905845642, + 0.05349723994731903, + 0.16371041536331177, + 0.05303013697266579, + 0.05718506872653961, + 0.08210023492574692 + ], + "credit_map": [ + 0.020788857713341713, + 0.030490560457110405, + 0.02510792762041092, + 0.02500978298485279, + 0.04219518601894379, + 0.02869885228574276, + 0.020492909476161003, + 0.02270950749516487, + 0.007214295212179422, + 0.016426365822553635, + 0.010456660762429237, + 0.009256526827812195, + 0.03534785658121109, + 0.016952838748693466, + 0.00801415927708149, + 0.010326968505978584, + 0.009209150448441505, + 0.015546377748250961, + 0.005810873117297888, + 0.00781246367841959, + 0.016034595668315887, + 0.011749467812478542, + 0.009270258247852325, + 0.005044093355536461, + 0.006722248159348965, + 0.008852435275912285, + 0.004963632207363844, + 0.006597722414880991, + 0.010612976737320423, + 0.009994466789066792, + 0.00932991597801447, + 0.006469422951340675, + 0.008969814516603947, + 0.028700657188892365, + 0.008038967847824097, + 0.008754229173064232, + 0.03666067123413086, + 0.01045508123934269, + 0.005852141417562962, + 0.009535962715744972, + 0.006447615567594767, + 0.008625633083283901, + 0.005631946958601475, + 0.0068332296796143055, + 0.009179042652249336, + 0.006589589174836874, + 0.004699284210801125, + 0.005490901879966259, + 0.018164033070206642, + 0.0304929967969656, + 0.017775485292077065, + 0.01915891095995903, + 0.021598774939775467, + 0.027213478460907936, + 0.014951436780393124, + 0.014355302788317204, + 0.0066154408268630505, + 0.008914750069379807, + 0.005528399255126715, + 0.004998528398573399, + 0.00981510616838932, + 0.006126727443188429, + 0.005683549214154482, + 0.005347630009055138, + 0.005907217040657997, + 0.011532970704138279, + 0.004165667109191418, + 0.005539057310670614, + 0.013652052730321884, + 0.007513862103223801, + 0.003830524394288659, + 0.005043718498200178, + 0.00466386741027236, + 0.016083210706710815, + 0.007040807045996189, + 0.013450868427753448, + 0.020193617790937424, + 0.006366519723087549, + 0.00821177288889885, + 0.0060895695351064205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.04869871959090233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280438423156738 + ], + "opposite_logprob_tokens": [ + -2.0178043842315674 + ], + "logprob_reward": -0.14280438423156738, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280438423156738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15679524838924408, + 0.14790450036525726, + 0.08790940791368484, + 0.0675356462597847, + 0.06104658916592598, + 0.055527932941913605, + 0.21647955477237701, + 0.082033172249794, + 0.06756394356489182, + 0.057203952223062515 + ], + "credit_map": [ + 0.018877636641263962, + 0.02455057017505169, + 0.01717190630733967, + 0.011952110566198826, + 0.03418876230716705, + 0.017797568812966347, + 0.016128044575452805, + 0.0161286573857069, + 0.016153724864125252, + 0.02117093838751316, + 0.01844058744609356, + 0.021133558824658394, + 0.025793803855776787, + 0.01409873366355896, + 0.012780639342963696, + 0.018332524225115776, + 0.008310160599648952, + 0.013096018694341183, + 0.010044432245194912, + 0.01365192886441946, + 0.013225564733147621, + 0.008884327486157417, + 0.013159726746380329, + 0.007537255063652992, + 0.007744944654405117, + 0.008775289170444012, + 0.006684012711048126, + 0.007634215988218784, + 0.008976388722658157, + 0.008159847930073738, + 0.013621218502521515, + 0.0059397355653345585, + 0.008516563102602959, + 0.012064036913216114, + 0.005356020759791136, + 0.004762966651469469, + 0.013213584199547768, + 0.006276206113398075, + 0.004265679512172937, + 0.006591539364308119, + 0.00573583273217082, + 0.008336304686963558, + 0.005909820552915335, + 0.004978380631655455, + 0.010185198858380318, + 0.0068851797841489315, + 0.005556479096412659, + 0.007940741255879402, + 0.023333599790930748, + 0.03628434240818024, + 0.026838093996047974, + 0.028608985245227814, + 0.02579049952328205, + 0.0325973741710186, + 0.022190602496266365, + 0.02083607204258442, + 0.011691183783113956, + 0.00799616426229477, + 0.01207729335874319, + 0.011550728231668472, + 0.009695201180875301, + 0.009436986409127712, + 0.012185120955109596, + 0.0074004982598125935, + 0.005334051791578531, + 0.01184152439236641, + 0.007462252397090197, + 0.008856447413563728, + 0.01539174560457468, + 0.0072499546222388744, + 0.005765923298895359, + 0.005662049166858196, + 0.00362687767483294, + 0.013670817948877811, + 0.0035563402343541384, + 0.006719416473060846, + 0.016718436032533646, + 0.003698715940117836, + 0.005057628732174635, + 0.004155727569013834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.14280438423156738 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13121531903743744, + 0.06163232773542404, + 0.04998365044593811, + 0.06492920219898224, + 0.05141279473900795, + 0.06507883220911026, + 0.44112586975097656, + 0.05637593939900398, + 0.016430094838142395, + 0.061816003173589706 + ], + "credit_map": [ + 0.013142265379428864, + 0.01542297750711441, + 0.014317227527499199, + 0.025919945910573006, + 0.013250682502985, + 0.011022266931831837, + 0.025551652535796165, + 0.012588302604854107, + 0.007385373581200838, + 0.009446844458580017, + 0.0062200697138905525, + 0.0062492419965565205, + 0.014603568241000175, + 0.008587003685534, + 0.004102251958101988, + 0.005037974566221237, + 0.00327301979996264, + 0.009852337650954723, + 0.0031691608019173145, + 0.005201741121709347, + 0.01148457545787096, + 0.004342924803495407, + 0.009360489435493946, + 0.003299406263977289, + 0.007613311987370253, + 0.009779197163879871, + 0.005388221703469753, + 0.00457670958712697, + 0.010611491277813911, + 0.011638134717941284, + 0.009983744472265244, + 0.0053383903577923775, + 0.0061745247803628445, + 0.007677736226469278, + 0.005563730373978615, + 0.0073329247534275055, + 0.006998938508331776, + 0.006494150497019291, + 0.004479162860661745, + 0.006691629532724619, + 0.007096680346876383, + 0.008220234885811806, + 0.006770235951989889, + 0.00814772117882967, + 0.013574621640145779, + 0.006628493312746286, + 0.006465696729719639, + 0.008175145834684372, + 0.04717155545949936, + 0.07659447938203812, + 0.05208095163106918, + 0.05534340441226959, + 0.05526131018996239, + 0.06953161209821701, + 0.04424035921692848, + 0.040902163833379745, + 0.006502295378595591, + 0.007286685053259134, + 0.0077813598327338696, + 0.008154584094882011, + 0.007942073047161102, + 0.005138044711202383, + 0.008198270574212074, + 0.005372622516006231, + 0.0020497250370681286, + 0.002056014258414507, + 0.0016158812213689089, + 0.0024641957134008408, + 0.002708976622670889, + 0.002196467947214842, + 0.0015733501641079783, + 0.0017654836410656571, + 0.005298242438584566, + 0.015315794385969639, + 0.0029687427449971437, + 0.006057283841073513, + 0.0188955869525671, + 0.00559511361643672, + 0.004258620087057352, + 0.003426620736718178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.03811221942305565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054952941834926605 + ], + "opposite_logprob_tokens": [ + -2.92995285987854 + ], + "logprob_reward": -0.054952941834926605, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054952941834926605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.166973277926445, + 0.14705336093902588, + 0.041973862797021866, + 0.04538635537028313, + 0.08484630286693573, + 0.01849888637661934, + 0.358812153339386, + 0.04794490337371826, + 0.01665806770324707, + 0.07185277342796326 + ], + "credit_map": [ + 0.01909354329109192, + 0.02192877233028412, + 0.01623346470296383, + 0.027211228385567665, + 0.023511283099651337, + 0.016911491751670837, + 0.02734454534947872, + 0.01473897136747837, + 0.013641434721648693, + 0.026338525116443634, + 0.015033491887152195, + 0.01617399975657463, + 0.03711899742484093, + 0.01772315427660942, + 0.010198311880230904, + 0.010825463570654392, + 0.005006701219826937, + 0.005815737880766392, + 0.003344298806041479, + 0.0041347285732626915, + 0.0072952997870743275, + 0.005574438720941544, + 0.007594936992973089, + 0.0032077247742563486, + 0.003983973525464535, + 0.00678037665784359, + 0.003938779234886169, + 0.004690556321293116, + 0.00700412830337882, + 0.008054704405367374, + 0.007142440415918827, + 0.0037913983687758446, + 0.01333307009190321, + 0.010794045403599739, + 0.009774626232683659, + 0.010481368750333786, + 0.011378934606909752, + 0.009966718964278698, + 0.007389351259917021, + 0.011728205718100071, + 0.002823122777044773, + 0.0026181465946137905, + 0.0021728691644966602, + 0.001928805373609066, + 0.0029350307304412127, + 0.0022224141284823418, + 0.0015097629511728883, + 0.0022887387312948704, + 0.038499534130096436, + 0.06464873254299164, + 0.03884430602192879, + 0.04204489290714264, + 0.048204533755779266, + 0.05958583950996399, + 0.034490298479795456, + 0.03249404951930046, + 0.006664665881544352, + 0.005245896056294441, + 0.006759545300155878, + 0.006636607926338911, + 0.00660553527995944, + 0.005909280385822058, + 0.006427853833884001, + 0.0036955266259610653, + 0.0019810793455690145, + 0.0027473149821162224, + 0.001742302207276225, + 0.001578064518980682, + 0.0031491087283939123, + 0.002621962921693921, + 0.0013950434513390064, + 0.0014431931776925921, + 0.00744421873241663, + 0.015152143314480782, + 0.004142767749726772, + 0.00797627680003643, + 0.019812865182757378, + 0.0071459305472671986, + 0.005502086132764816, + 0.004676496610045433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.054952941834926605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot captures a flowing, off-white, sheer silk curtain hung in front of a large, open window in a minimalist interior. Beneath it, the polished, deep espresso brown hardwood floor gleams with a high-gloss, mirror-like finish, reflecting the soft ambient light. The entire scene is bathed in bright, diffused natural morning light, creating a tranquil, high-key atmosphere. A gentle, yet deliberate breeze energizes the fabric, causing it to billow gracefully inward. The delicate silk expands and ripples, creating a series of fluid, undulating waves that dance briefly in the air. As the curtain moves, a sharp, brilliant streak of golden sunlight slants across the dark wooden floor, momentarily illuminating the tiny dust particles suspended in its path. This moment is fleeting; as the wind subsides, the curtain slowly and weightlessly returns to its original position, and the sharp streak of sunlight on the floor disappears as quickly as it appeared, leaving behind only a soft, even glow. The entire sequence unfolds with a quiet, poetic rhythm, enhancing the elegant, tranquil, and minimalist atmosphere of the space.", + "seed": 22063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_3/index_3134_seed_22063_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": -0.040633033215999606, + "vlm_logprob": -0.040633033215999606, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.040633033215999606, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009247920475900173 + ], + "opposite_logprob_tokens": [ + -5.009247779846191 + ], + "logprob_reward": -0.009247920475900173, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009247920475900173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06333128362894058, + 0.08283748477697372, + 0.16874726116657257, + 0.03486327826976776, + 0.05017993226647377, + 0.04962237551808357, + 0.03519832715392113, + 0.023766016587615013, + 0.37073206901550293, + 0.12072187662124634 + ], + "credit_map": [ + 0.004764166194945574, + 0.008424818515777588, + 0.006349833682179451, + 0.007391775492578745, + 0.011489085853099823, + 0.0072815995663404465, + 0.010255202651023865, + 0.007374811917543411, + 0.0071474225260317326, + 0.012156134471297264, + 0.0053713382221758366, + 0.015731459483504295, + 0.0105568403378129, + 0.010686291381716728, + 0.014534163288772106, + 0.006653844844549894, + 0.011550644412636757, + 0.019689267501235008, + 0.01870749145746231, + 0.021561149507761, + 0.01985803246498108, + 0.022974058985710144, + 0.027821185067296028, + 0.026585452258586884, + 0.002145997481420636, + 0.006618238054215908, + 0.002317548729479313, + 0.006670937407761812, + 0.005123639013618231, + 0.0050309342332184315, + 0.0043668728321790695, + 0.0025891170371323824, + 0.0028177122585475445, + 0.007509049028158188, + 0.0042808144353330135, + 0.0072375331073999405, + 0.005872353445738554, + 0.00567568140104413, + 0.011489995755255222, + 0.005296800285577774, + 0.0037749928887933493, + 0.006365577690303326, + 0.004534172359853983, + 0.007081644609570503, + 0.008554519154131413, + 0.00659954734146595, + 0.006738451775163412, + 0.005973477382212877, + 0.004232896491885185, + 0.004693553317338228, + 0.004798606503754854, + 0.0029515333008021116, + 0.007336558308452368, + 0.004174749832600355, + 0.0037309115286916494, + 0.0032795225270092487, + 0.0022698899265378714, + 0.004668121691793203, + 0.0025031347759068012, + 0.0021651885472238064, + 0.0035324953496456146, + 0.003299237694591284, + 0.002419164404273033, + 0.0029087872244417667, + 0.009181824512779713, + 0.15518610179424286, + 0.00550454156473279, + 0.00855982955545187, + 0.01523913349956274, + 0.1624370962381363, + 0.006615632213652134, + 0.008007947355508804, + 0.006893121171742678, + 0.020892823114991188, + 0.013266364112496376, + 0.01254542451351881, + 0.016392884775996208, + 0.014250725507736206, + 0.025680260732769966, + 0.010800284333527088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.009247920475900173 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08911111205816269 + ], + "opposite_logprob_tokens": [ + -2.464111089706421 + ], + "logprob_reward": -0.08911111205816269, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08911111205816269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11672824621200562, + 0.15716099739074707, + 0.19405505061149597, + 0.03299589082598686, + 0.10270103067159653, + 0.01882232166826725, + 0.032204266637563705, + 0.019372262060642242, + 0.28835663199424744, + 0.03760329633951187 + ], + "credit_map": [ + 0.008839743211865425, + 0.015800796449184418, + 0.010469933971762657, + 0.014839818701148033, + 0.019915316253900528, + 0.02030295506119728, + 0.016605697572231293, + 0.009953983128070831, + 0.009105334989726543, + 0.03606428578495979, + 0.009490616619586945, + 0.022119535133242607, + 0.03023410029709339, + 0.017830446362495422, + 0.020551791414618492, + 0.01176487561315298, + 0.012966351583600044, + 0.026120876893401146, + 0.02225872501730919, + 0.024557065218687057, + 0.023411348462104797, + 0.026337143033742905, + 0.030068203806877136, + 0.028335317969322205, + 0.0023782222997397184, + 0.006897835060954094, + 0.0025687492452561855, + 0.004216344561427832, + 0.005157241597771645, + 0.005466153379529715, + 0.003178179729729891, + 0.0031331623904407024, + 0.004239065572619438, + 0.011298511177301407, + 0.006980185862630606, + 0.02043011598289013, + 0.011993742547929287, + 0.009656702168285847, + 0.02705192007124424, + 0.011050782166421413, + 0.001449365052394569, + 0.0025781223084777594, + 0.0013741620350629091, + 0.0023241841699928045, + 0.0030452089849859476, + 0.002606883179396391, + 0.002716651652008295, + 0.002727742772549391, + 0.0033131518866866827, + 0.003810465568676591, + 0.003646246390417218, + 0.002867958275601268, + 0.008624470792710781, + 0.004467722959816456, + 0.0027383030392229557, + 0.0027359446976333857, + 0.0016377585707232356, + 0.0034210148733109236, + 0.002671212889254093, + 0.0016356646083295345, + 0.0019334966782480478, + 0.0023978487588465214, + 0.002021168824285269, + 0.0036540934816002846, + 0.007091938052326441, + 0.11971350759267807, + 0.004575359169393778, + 0.00767928222194314, + 0.012070094235241413, + 0.1255512833595276, + 0.0053132642060518265, + 0.006361899431794882, + 0.0023815606255084276, + 0.005528891924768686, + 0.0046665058471262455, + 0.0038817126769572496, + 0.005540965124964714, + 0.003927899524569511, + 0.007922442629933357, + 0.0037533154245465994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.08911111205816269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002592831850052 + ], + "opposite_logprob_tokens": [ + -2.69502592086792 + ], + "logprob_reward": -0.07002592831850052, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002592831850052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03136920928955078, + 0.05791005864739418, + 0.05073747783899307, + 0.048886459320783615, + 0.09102780371904373, + 0.09231434017419815, + 0.020771387964487076, + 0.021515313535928726, + 0.39430662989616394, + 0.19116130471229553 + ], + "credit_map": [ + 0.0027396201621741056, + 0.0041053504683077335, + 0.0031767163891345263, + 0.0037637101486325264, + 0.004912313539534807, + 0.004398967605084181, + 0.00459210854023695, + 0.003680423367768526, + 0.004890278913080692, + 0.011117590591311455, + 0.003334942040964961, + 0.010386474430561066, + 0.00858944933861494, + 0.00599929504096508, + 0.009513737633824348, + 0.004078295547515154, + 0.003556435229256749, + 0.0052627623081207275, + 0.005981056485325098, + 0.005928628612309694, + 0.006342150270938873, + 0.007394592277705669, + 0.00856840144842863, + 0.007703451905399561, + 0.002671519760042429, + 0.009090930223464966, + 0.003296022769063711, + 0.009559829719364643, + 0.006754997652024031, + 0.007467289920896292, + 0.0067694117315113544, + 0.003276453586295247, + 0.006869475357234478, + 0.013122949749231339, + 0.007257937919348478, + 0.010528496466577053, + 0.013553883880376816, + 0.013907322660088539, + 0.013389315456151962, + 0.012398420833051205, + 0.00670652324333787, + 0.010913017205893993, + 0.007856144569814205, + 0.01240306906402111, + 0.01746613346040249, + 0.01348339207470417, + 0.013617804273962975, + 0.009868262335658073, + 0.0020154910162091255, + 0.0030374410562217236, + 0.002562912879511714, + 0.0016647323500365019, + 0.004280367400497198, + 0.0025089820846915245, + 0.0023863413371145725, + 0.0023151191417127848, + 0.0019560912624001503, + 0.002707261126488447, + 0.0021829467732459307, + 0.0019038552418351173, + 0.0024132155813276768, + 0.0024483478628098965, + 0.002835726598277688, + 0.005067870020866394, + 0.009555255062878132, + 0.16231770813465118, + 0.006504235323518515, + 0.012294286862015724, + 0.01687774620950222, + 0.17007069289684296, + 0.007859899662435055, + 0.00882680993527174, + 0.010061590000987053, + 0.027876151725649834, + 0.021709740161895752, + 0.02192115969955921, + 0.028036708012223244, + 0.01825530454516411, + 0.043808259069919586, + 0.019492365419864655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.07002592831850052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199883677065372 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199883677065372, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199883677065372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060282956808805466, + 0.09563563764095306, + 0.09732368588447571, + 0.06322275102138519, + 0.08876241743564606, + 0.048447735607624054, + 0.0686408281326294, + 0.25521159172058105, + 0.05301225557923317, + 0.16946013271808624 + ], + "credit_map": [ + 0.004857892636209726, + 0.007327074650675058, + 0.008240236900746822, + 0.007567299529910088, + 0.009051181375980377, + 0.008814431726932526, + 0.006403523497283459, + 0.008021313697099686, + 0.0077931820414960384, + 0.020283276215195656, + 0.0057604508474469185, + 0.01500472892075777, + 0.016388840973377228, + 0.011319111101329327, + 0.012519906274974346, + 0.006566139869391918, + 0.0057086641900241375, + 0.018880784511566162, + 0.007820588536560535, + 0.008669553324580193, + 0.022496314719319344, + 0.011491463519632816, + 0.01056260708719492, + 0.011693711392581463, + 0.004831902217119932, + 0.012953833676874638, + 0.006328766234219074, + 0.006844987627118826, + 0.008989432826638222, + 0.010678177699446678, + 0.006765523459762335, + 0.005830130074173212, + 0.0048459661193192005, + 0.012079560197889805, + 0.006300546228885651, + 0.013261438347399235, + 0.01394371036440134, + 0.011602730490267277, + 0.01642363704741001, + 0.010304822586476803, + 0.003901178715750575, + 0.007331878412514925, + 0.00474215392023325, + 0.005450576078146696, + 0.008541830815374851, + 0.00609837519004941, + 0.005718905478715897, + 0.006662835832685232, + 0.007093287538737059, + 0.009129089303314686, + 0.008531449362635612, + 0.005353596992790699, + 0.016240278258919716, + 0.008448423817753792, + 0.007382390554994345, + 0.006462313700467348, + 0.02309131994843483, + 0.040300216525793076, + 0.04245579242706299, + 0.020723624154925346, + 0.023484159260988235, + 0.03012128546833992, + 0.029794277623295784, + 0.04524092376232147, + 0.0017423303797841072, + 0.018058964982628822, + 0.0023782532662153244, + 0.004226175602525473, + 0.004057989455759525, + 0.018691791221499443, + 0.0023228938225656748, + 0.001533861388452351, + 0.010188297368586063, + 0.03011234849691391, + 0.018635861575603485, + 0.018719227984547615, + 0.02149871177971363, + 0.01991933025419712, + 0.0355277881026268, + 0.014858563430607319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.014199883677065372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580321550369263 + ], + "opposite_logprob_tokens": [ + -3.895580291748047 + ], + "logprob_reward": -0.020580321550369263, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580321550369263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037237346172332764, + 0.0680079534649849, + 0.13596704602241516, + 0.04957592859864235, + 0.11701293289661407, + 0.03601100668311119, + 0.029158862307667732, + 0.024810798466205597, + 0.4747776389122009, + 0.027440542355179787 + ], + "credit_map": [ + 0.003245726227760315, + 0.005075742956250906, + 0.0037563962396234274, + 0.005133399274200201, + 0.004624248947948217, + 0.0053164782002568245, + 0.005343638826161623, + 0.004741716664284468, + 0.006199894472956657, + 0.01424667239189148, + 0.004780603107064962, + 0.009350086562335491, + 0.00970478542149067, + 0.010320927016437054, + 0.007905527949333191, + 0.005499451886862516, + 0.008158338256180286, + 0.020627206191420555, + 0.013227618299424648, + 0.012969151139259338, + 0.027172105386853218, + 0.01656596176326275, + 0.01853085309267044, + 0.018715811893343925, + 0.004075226839631796, + 0.009661705233156681, + 0.004427688661962748, + 0.006171252112835646, + 0.008808139711618423, + 0.008048121817409992, + 0.004561807028949261, + 0.0038219839334487915, + 0.005508771166205406, + 0.017013369128108025, + 0.008352727629244328, + 0.021367274224758148, + 0.016116516664624214, + 0.01396932452917099, + 0.021434234455227852, + 0.013250726275146008, + 0.0021224787924438715, + 0.007134794257581234, + 0.002755712950602174, + 0.004439667332917452, + 0.0062225451692938805, + 0.006876285653561354, + 0.0031484239734709263, + 0.0033110985532402992, + 0.0032618839759379625, + 0.004391370341181755, + 0.003979366272687912, + 0.002726558595895767, + 0.0046976227313280106, + 0.003376826411113143, + 0.0032592874485999346, + 0.0034659449011087418, + 0.0024484428577125072, + 0.004800369031727314, + 0.0024568894878029823, + 0.002308167750015855, + 0.0033489868510514498, + 0.0033285573590546846, + 0.0030929783824831247, + 0.003026406979188323, + 0.011840902268886566, + 0.1987762451171875, + 0.007054126355797052, + 0.01072768960148096, + 0.019557012245059013, + 0.20811840891838074, + 0.00854385644197464, + 0.010159403085708618, + 0.0019039100734516978, + 0.0029033406171947718, + 0.0034904449712485075, + 0.0032502533867955208, + 0.004190972540527582, + 0.0032791802659630775, + 0.005149963311851025, + 0.0032724745105952024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.020580321550369263 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot captures a flowing, off-white, sheer silk curtain hung in front of a large, open window in a minimalist interior. Beneath it, the polished, deep espresso brown hardwood floor gleams with a high-gloss, mirror-like finish, reflecting the soft ambient light. The entire scene is bathed in bright, diffused natural morning light, creating a tranquil, high-key atmosphere. A gentle, yet deliberate breeze energizes the fabric, causing it to billow gracefully inward. The delicate silk expands and ripples, creating a series of fluid, undulating waves that dance briefly in the air. As the curtain moves, a sharp, brilliant streak of golden sunlight slants across the dark wooden floor, momentarily illuminating the tiny dust particles suspended in its path. This moment is fleeting; as the wind subsides, the curtain slowly and weightlessly returns to its original position, and the sharp streak of sunlight on the floor disappears as quickly as it appeared, leaving behind only a soft, even glow. The entire sequence unfolds with a quiet, poetic rhythm, enhancing the elegant, tranquil, and minimalist atmosphere of the space.", + "seed": 22064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_3/index_3134_seed_22064_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": -0.07299974206835032, + "vlm_logprob": -0.07299974206835032, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.07299974206835032, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012603395618498325 + ], + "opposite_logprob_tokens": [ + -4.512603282928467 + ], + "logprob_reward": -0.012603395618498325, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012603395618498325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22371938824653625, + 0.10077270865440369, + 0.13528035581111908, + 0.050957951694726944, + 0.04505642503499985, + 0.08338190615177155, + 0.060806646943092346, + 0.2025519460439682, + 0.04708898440003395, + 0.0503837876021862 + ], + "credit_map": [ + 0.029381873086094856, + 0.03147595375776291, + 0.01699737459421158, + 0.03313023969531059, + 0.038156528025865555, + 0.03494725748896599, + 0.0235628392547369, + 0.01606729067862034, + 0.009392487816512585, + 0.02218376100063324, + 0.005717604421079159, + 0.007182726636528969, + 0.02430400252342224, + 0.016515647992491722, + 0.006829786114394665, + 0.008646674454212189, + 0.013632592745125294, + 0.022771798074245453, + 0.009757280349731445, + 0.012904932722449303, + 0.025871239602565765, + 0.021523986011743546, + 0.017687810584902763, + 0.011130714789032936, + 0.0031984562519937754, + 0.008479863405227661, + 0.002700903220102191, + 0.004469568375498056, + 0.011850501410663128, + 0.009800481610000134, + 0.0052696713246405125, + 0.005188499111682177, + 0.003250115318223834, + 0.007990949787199497, + 0.0021440736018121243, + 0.002674596616998315, + 0.010776123963296413, + 0.010081569664180279, + 0.0034723165445029736, + 0.004666677676141262, + 0.0057558114640414715, + 0.012220515869557858, + 0.005624945275485516, + 0.009436695836484432, + 0.011244534514844418, + 0.018309710547327995, + 0.01126326248049736, + 0.009526421315968037, + 0.005224244203418493, + 0.008937031961977482, + 0.005338084883987904, + 0.01049063540995121, + 0.00881512463092804, + 0.009272200055420399, + 0.006951508112251759, + 0.005777810234576464, + 0.010941023007035255, + 0.04138173535466194, + 0.009633523412048817, + 0.014842495322227478, + 0.042513780295848846, + 0.048834871500730515, + 0.020253067836165428, + 0.0141514353454113, + 0.006030485965311527, + 0.006692560389637947, + 0.004513544030487537, + 0.00534053472802043, + 0.005688623059540987, + 0.006345133297145367, + 0.006076785270124674, + 0.006401311606168747, + 0.0038408376276493073, + 0.008339992724359035, + 0.0028666453436017036, + 0.005691806320101023, + 0.00977071188390255, + 0.010578413493931293, + 0.005052372347563505, + 0.004243007395416498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.012603395618498325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20174114406108856 + ], + "opposite_logprob_tokens": [ + -1.701741099357605 + ], + "logprob_reward": -0.20174114406108856, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20174114406108856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23434199392795563, + 0.08858541399240494, + 0.04351559281349182, + 0.23196293413639069, + 0.04445130005478859, + 0.041803229600191116, + 0.07124508917331696, + 0.19223903119564056, + 0.01622466742992401, + 0.035630807280540466 + ], + "credit_map": [ + 0.023257819935679436, + 0.04608575999736786, + 0.011967440135776997, + 0.02721412293612957, + 0.057472214102745056, + 0.038822241127491, + 0.014993495307862759, + 0.014528888277709484, + 0.007033581845462322, + 0.018935518339276314, + 0.00585772143676877, + 0.0066452063620090485, + 0.0159195214509964, + 0.015050661750137806, + 0.008370722644031048, + 0.010772474110126495, + 0.0035974967759102583, + 0.008924477733671665, + 0.0025241703260689974, + 0.003412834368646145, + 0.006852096412330866, + 0.008588054217398167, + 0.0048246001824736595, + 0.004791858606040478, + 0.017879081889986992, + 0.04586603119969368, + 0.008912261575460434, + 0.015936270356178284, + 0.048780325800180435, + 0.058285947889089584, + 0.015836616978049278, + 0.020466385409235954, + 0.003291719127446413, + 0.007604480255395174, + 0.002550398698076606, + 0.0020400912035256624, + 0.009009343571960926, + 0.011445440351963043, + 0.0038734888657927513, + 0.004636333789676428, + 0.002847092691808939, + 0.007103456649929285, + 0.0029908977448940277, + 0.0038485853001475334, + 0.005292573478072882, + 0.01077423244714737, + 0.004593892954289913, + 0.004352493677288294, + 0.004671267233788967, + 0.010511848144233227, + 0.0052588325925171375, + 0.009522062726318836, + 0.010929251089692116, + 0.015583676286041737, + 0.006920352578163147, + 0.007847793400287628, + 0.010231129825115204, + 0.043348412960767746, + 0.008935489691793919, + 0.012368240393698215, + 0.0428776741027832, + 0.043873824179172516, + 0.019032560288906097, + 0.011571682058274746, + 0.0019004716305062175, + 0.001954900799319148, + 0.0011371254222467542, + 0.0016532838344573975, + 0.0020651959348469973, + 0.003610428422689438, + 0.002020265208557248, + 0.001882994663901627, + 0.00305097340606153, + 0.007918463088572025, + 0.001293286681175232, + 0.002566698705777526, + 0.006294730119407177, + 0.009798221290111542, + 0.0019753684755414724, + 0.0027330617886036634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.20174114406108856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2003348022699356, + 0.14092370867729187, + 0.0826820433139801, + 0.08359075337648392, + 0.0599990151822567, + 0.2308991253376007, + 0.08561410754919052, + 0.06178278103470802, + 0.034414369612932205, + 0.019759228453040123 + ], + "credit_map": [ + 0.020271582528948784, + 0.03050627000629902, + 0.01923246867954731, + 0.02828350104391575, + 0.02852095291018486, + 0.024143964052200317, + 0.02288338914513588, + 0.026492685079574585, + 0.012646578252315521, + 0.01634456403553486, + 0.011453952640295029, + 0.020941240713000298, + 0.015652280300855637, + 0.02439572848379612, + 0.01953367330133915, + 0.0199557077139616, + 0.008763453923165798, + 0.0163470059633255, + 0.004816189408302307, + 0.006880240514874458, + 0.015993105247616768, + 0.014075210317969322, + 0.008846985176205635, + 0.006959859281778336, + 0.006911323871463537, + 0.01341862790286541, + 0.006592917256057262, + 0.007860620506107807, + 0.0138348788022995, + 0.01783236488699913, + 0.008066843263804913, + 0.009073186665773392, + 0.005212122108787298, + 0.008839537389576435, + 0.003448027651757002, + 0.004503763280808926, + 0.010229778476059437, + 0.013087534345686436, + 0.006844751071184874, + 0.007833506911993027, + 0.016536502167582512, + 0.0337272547185421, + 0.01654508151113987, + 0.027193058282136917, + 0.024484967812895775, + 0.04529242217540741, + 0.0403265655040741, + 0.026793302968144417, + 0.0071892328560352325, + 0.011791897006332874, + 0.009012388065457344, + 0.009623068384826183, + 0.012604177929461002, + 0.016291532665491104, + 0.008464248850941658, + 0.010637574829161167, + 0.008158612996339798, + 0.007866430096328259, + 0.00568343373015523, + 0.007630081381648779, + 0.007800142280757427, + 0.008297167718410492, + 0.008260137401521206, + 0.00808678288012743, + 0.00349060189910233, + 0.007169670425355434, + 0.0023167631588876247, + 0.002101440681144595, + 0.008513524197041988, + 0.005719842854887247, + 0.0030174630228430033, + 0.002085070824250579, + 0.002223394578322768, + 0.0028583514504134655, + 0.001164102228358388, + 0.0018736995989456773, + 0.003182589542120695, + 0.004000068176537752, + 0.0021862557623535395, + 0.0022707697935402393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.012545360252261162 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077335104346275 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077335104346275, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077335104346275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1977357417345047, + 0.15542593598365784, + 0.07222109287977219, + 0.06480348855257034, + 0.11425992846488953, + 0.14802001416683197, + 0.08354727923870087, + 0.0972190797328949, + 0.02924472652375698, + 0.037522632628679276 + ], + "credit_map": [ + 0.018811309710144997, + 0.02598537690937519, + 0.013052795082330704, + 0.040634796023368835, + 0.027671854943037033, + 0.0272376649081707, + 0.027139324694871902, + 0.017202632501721382, + 0.01193880196660757, + 0.02232554368674755, + 0.015321885235607624, + 0.020878123119473457, + 0.018916208297014236, + 0.02192491479218006, + 0.020979663357138634, + 0.023140799254179, + 0.005722485017031431, + 0.01449197344481945, + 0.004318609833717346, + 0.006195195019245148, + 0.013293545693159103, + 0.012906654737889767, + 0.00799841620028019, + 0.007294211536645889, + 0.006994226947426796, + 0.012131907977163792, + 0.002584272064268589, + 0.00428243912756443, + 0.0127669433131814, + 0.014539806172251701, + 0.005558858159929514, + 0.005945035256445408, + 0.008695987984538078, + 0.01442598644644022, + 0.00975990854203701, + 0.012220300734043121, + 0.012920993380248547, + 0.01881963200867176, + 0.017555106431245804, + 0.019862016662955284, + 0.011212180368602276, + 0.02067149616777897, + 0.011649765074253082, + 0.018803678452968597, + 0.017291665077209473, + 0.032414358109235764, + 0.020743420347571373, + 0.01523345522582531, + 0.0074362377636134624, + 0.014378520660102367, + 0.008421393111348152, + 0.007292390801012516, + 0.015881551429629326, + 0.016643472015857697, + 0.0068928757682442665, + 0.006600835360586643, + 0.006608241703361273, + 0.012675853446125984, + 0.005926208104938269, + 0.008270562626421452, + 0.018139643594622612, + 0.024921195581555367, + 0.01171225868165493, + 0.008965115994215012, + 0.002557000145316124, + 0.003830761881545186, + 0.002348278183490038, + 0.0034502544440329075, + 0.0038646406028419733, + 0.00533785717561841, + 0.004215858411043882, + 0.003640076844021678, + 0.003139647888019681, + 0.008929453790187836, + 0.0019793289247900248, + 0.003811378264799714, + 0.006402901839464903, + 0.006708621513098478, + 0.0036181355826556683, + 0.002933165756985545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.011077335104346275 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a flowing, off-white, sheer silk curtain visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the curtain hanging in front of a large, open window in a minimalist interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor made of polished, deep espresso brown hardwood with a high-gloss, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp, brilliant streak of golden sunlight slant across the floor while dust particles are visible in its path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the curtain billow inward and ripple with fluid, undulating waves due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703147530555725 + ], + "opposite_logprob_tokens": [ + -2.1270315647125244 + ], + "logprob_reward": -0.12703147530555725, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703147530555725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2098291963338852, + 0.06277098506689072, + 0.10916636139154434, + 0.08399093151092529, + 0.04228688403964043, + 0.2642501890659332, + 0.049495402723550797, + 0.09032604098320007, + 0.03254276141524315, + 0.05534123256802559 + ], + "credit_map": [ + 0.021798046305775642, + 0.026324007660150528, + 0.01750754565000534, + 0.02909853123128414, + 0.03266165405511856, + 0.04277179390192032, + 0.01870332844555378, + 0.02096429094672203, + 0.005145740695297718, + 0.009756408631801605, + 0.005946887191385031, + 0.008076113648712635, + 0.008609415031969547, + 0.007431900128722191, + 0.009407504461705685, + 0.008397010155022144, + 0.010749155655503273, + 0.019278917461633682, + 0.008600804023444653, + 0.009943456389009953, + 0.019036035984754562, + 0.01813437230885029, + 0.01243837270885706, + 0.010985240340232849, + 0.004960407968610525, + 0.0190903190523386, + 0.002761451993137598, + 0.005806917790323496, + 0.014576256275177002, + 0.024312643334269524, + 0.005379817448556423, + 0.00710311671718955, + 0.003259139135479927, + 0.00817782711237669, + 0.0025827756617218256, + 0.0031408825889229774, + 0.007363615091890097, + 0.009746800176799297, + 0.0038370920810848475, + 0.004178747069090605, + 0.01895095780491829, + 0.04022670164704323, + 0.0194894690066576, + 0.028512952849268913, + 0.03562339395284653, + 0.0623435340821743, + 0.029525568708777428, + 0.029577605426311493, + 0.007049119099974632, + 0.007009278517216444, + 0.004620922729372978, + 0.006908881478011608, + 0.006818574853241444, + 0.006890563294291496, + 0.006364424247294664, + 0.003833634313195944, + 0.004488853737711906, + 0.02464240789413452, + 0.005816218443214893, + 0.005598666612058878, + 0.024709466844797134, + 0.013466794975101948, + 0.007026873994618654, + 0.004576744511723518, + 0.003333313623443246, + 0.0064604333601891994, + 0.0024053952656686306, + 0.0034161636140197515, + 0.005284550599753857, + 0.004497535526752472, + 0.0037994487211108208, + 0.00334591930732131, + 0.0036523074377328157, + 0.011175386607646942, + 0.0024681317154318094, + 0.00462924363091588, + 0.012890449725091457, + 0.011884485371410847, + 0.004450848326086998, + 0.004190376494079828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.12703147530555725 + } + ], + "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/0000052/group_4/index_tensor([7814, 7814])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_4/index_tensor([7814, 7814])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..04d4de81fb351af63de5193a86cfc40d6c9ad460 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_4/index_tensor([7814, 7814])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from an upward angle, instantly transports us into a fantastical, Art Nouveau-inspired metal forest. Here, massive \"trees\" of twisted bronze and brass rise from the ground, their branches spreading like vines, their ends terminating in soft, lily-shaped lights. At the center of the frame, a green-haired swordsman lowers his weight into an explosive attacking stance, his bright green, shoulder-length hair striking against the warm glow of the metallic vegetation. His expression is stern and focused, and in his mouth he clutches a renowned short sword, its hilt entwined with ancient runes, its blade gleaming with a cold gleam. He wears a dark green, skin-tight leather armor embroidered with golden vines, and in his hands he holds two sleek, silver scimitars in reverse. Directly before him stands a massive, ornate column crafted entirely of polished brass, its shaft carved with intricate, interwoven reliefs of vibrant curves. The swordsman exerted force from his waist, and with the force of thunder, his arms slashed out in two swift and deadly silver arcs. The blades precisely sliced into the middle of the brass pistil. With a sharp metallic cutting sound and a burst of dazzling sparks, the massive pistil was neatly bisected. As the severed upper half began to slowly and heavily tilt, finally collapsing to the ground with a loud thud, the camera experienced a sharp and brief tremor, perfectly conveying the unparalleled impact and weight of the giant's collapse.", + "seed": 12849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_4/index_7814_seed_12849_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": -0.05460863728076219, + "vlm_logprob": -0.05460863728076219, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05460863728076219, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010586884804069996 + ], + "opposite_logprob_tokens": [ + -5.385586738586426 + ], + "logprob_reward": -0.010586884804069996, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010586884804069996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17064520716667175, + 0.11078750342130661, + 0.06864023208618164, + 0.03520496189594269, + 0.12240763753652573, + 0.04749322310090065, + 0.04867170751094818, + 0.1397923082113266, + 0.10050661861896515, + 0.1558506041765213 + ], + "credit_map": [ + 0.023686116561293602, + 0.023076588287949562, + 0.024026835337281227, + 0.025482287630438805, + 0.015292149037122726, + 0.016049262136220932, + 0.017027484253048897, + 0.0260044913738966, + 0.01159253902733326, + 0.010058624669909477, + 0.018478967249393463, + 0.014093845151364803, + 0.012337748892605305, + 0.010950847528874874, + 0.016245735809206963, + 0.01702919416129589, + 0.007408938370645046, + 0.00743254367262125, + 0.00884084403514862, + 0.011268200352787971, + 0.007265473250299692, + 0.006115639116615057, + 0.009370481595396996, + 0.010938111692667007, + 0.005315867252647877, + 0.002944589126855135, + 0.004804778378456831, + 0.0054223258048295975, + 0.003380577079951763, + 0.0033539559226483107, + 0.004596812184900045, + 0.005386055447161198, + 0.014794617891311646, + 0.014496475458145142, + 0.017603253945708275, + 0.02048826590180397, + 0.01027620118111372, + 0.014958641491830349, + 0.014399189502000809, + 0.01539099682122469, + 0.004294176585972309, + 0.006382330320775509, + 0.006436222232878208, + 0.0076274750754237175, + 0.0035098446533083916, + 0.004687707871198654, + 0.008061935193836689, + 0.006493531633168459, + 0.005211745388805866, + 0.004341960418969393, + 0.005136203020811081, + 0.00822217296808958, + 0.005974145606160164, + 0.004890508484095335, + 0.00695054791867733, + 0.007944425567984581, + 0.010476290248334408, + 0.009343084879219532, + 0.020306266844272614, + 0.04466309770941734, + 0.007650554645806551, + 0.008399532176554203, + 0.011894828639924526, + 0.02705865353345871, + 0.011865411885082722, + 0.011634236201643944, + 0.013205413706600666, + 0.014718709513545036, + 0.006845477037131786, + 0.01449249405413866, + 0.014065920375287533, + 0.01367894746363163, + 0.012848584912717342, + 0.01146393921226263, + 0.020439423620700836, + 0.014398066326975822, + 0.03069823421537876, + 0.010357721708714962, + 0.02462007850408554, + 0.03102456033229828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.010586884804069996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607493683695793 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.01607493683695793, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607493683695793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16094811260700226, + 0.19911041855812073, + 0.11252420395612717, + 0.07769084721803665, + 0.03387727588415146, + 0.04736161604523659, + 0.03525901958346367, + 0.13836351037025452, + 0.04495394229888916, + 0.1499110758304596 + ], + "credit_map": [ + 0.019360028207302094, + 0.022805042564868927, + 0.024114778265357018, + 0.024677492678165436, + 0.015342922881245613, + 0.013639074750244617, + 0.024528920650482178, + 0.016479872167110443, + 0.017399264499545097, + 0.03451930359005928, + 0.026077184826135635, + 0.038052912801504135, + 0.02115562930703163, + 0.012301607988774776, + 0.027956951409578323, + 0.021647579967975616, + 0.008913789875805378, + 0.014644940383732319, + 0.017568357288837433, + 0.01970408298075199, + 0.007503153756260872, + 0.009285934269428253, + 0.017222177237272263, + 0.017681773751974106, + 0.0070943087339401245, + 0.009382763877511024, + 0.010305475443601608, + 0.018011808395385742, + 0.005686893127858639, + 0.00715644471347332, + 0.007491238880902529, + 0.012561921030282974, + 0.004205808974802494, + 0.003604728262871504, + 0.003988056909292936, + 0.006817854009568691, + 0.0030021683778613806, + 0.004201805219054222, + 0.0033204529900103807, + 0.00473640626296401, + 0.0038364960346370935, + 0.006293704267591238, + 0.007553306873887777, + 0.010565745644271374, + 0.003664578776806593, + 0.004728587344288826, + 0.004262876696884632, + 0.006456325761973858, + 0.0030708368867635727, + 0.0029578611720353365, + 0.003683623392134905, + 0.004918145947158337, + 0.004451758228242397, + 0.0044809384271502495, + 0.005110546015202999, + 0.0065853106789290905, + 0.008534262888133526, + 0.009169120341539383, + 0.015370984561741352, + 0.045101504772901535, + 0.008378512226045132, + 0.00821591168642044, + 0.014073255471885204, + 0.02951996400952339, + 0.003589938161894679, + 0.004560340661555529, + 0.004767670761793852, + 0.009527510963380337, + 0.004054203629493713, + 0.005946790333837271, + 0.0064368234016001225, + 0.006070667412132025, + 0.012179902754724026, + 0.010453137569129467, + 0.017337901517748833, + 0.01831504888832569, + 0.0322691835463047, + 0.010377093218266964, + 0.026539010927081108, + 0.022439805790781975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.01607493683695793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495181307196617 + ], + "opposite_logprob_tokens": [ + -2.9299519062042236 + ], + "logprob_reward": -0.05495181307196617, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495181307196617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32707372307777405, + 0.11801843345165253, + 0.05289044976234436, + 0.051516707986593246, + 0.0538257360458374, + 0.03927914425730705, + 0.03561757132411003, + 0.03573622182011604, + 0.058332644402980804, + 0.22770938277244568 + ], + "credit_map": [ + 0.03571910411119461, + 0.04622166231274605, + 0.040776535868644714, + 0.0546039342880249, + 0.032102618366479874, + 0.02759382873773575, + 0.042361464351415634, + 0.04769456759095192, + 0.014999509789049625, + 0.011352398432791233, + 0.017558319494128227, + 0.01901075430214405, + 0.011584103107452393, + 0.008960489183664322, + 0.01494363322854042, + 0.019609233364462852, + 0.004754533525556326, + 0.006685894448310137, + 0.00801131408661604, + 0.007498355582356453, + 0.0058100116439163685, + 0.005758623592555523, + 0.00782698392868042, + 0.006544732954353094, + 0.0052582016214728355, + 0.0052480921149253845, + 0.006162992212921381, + 0.00969886314123869, + 0.006148336920887232, + 0.0054617333225905895, + 0.006706359796226025, + 0.006832128390669823, + 0.005031547974795103, + 0.005992671940475702, + 0.006870834156870842, + 0.009933934547007084, + 0.004282395355403423, + 0.005824876483529806, + 0.006973578594624996, + 0.008915895596146584, + 0.0033798881340771914, + 0.004923558793962002, + 0.004582107532769442, + 0.0072140172123909, + 0.002977828960865736, + 0.0039045994635671377, + 0.005798985715955496, + 0.006498155649751425, + 0.0034438548609614372, + 0.0032634271774441004, + 0.004439059179276228, + 0.005807685200124979, + 0.00422983942553401, + 0.003840925171971321, + 0.0033710398711264133, + 0.007221742533147335, + 0.0022846711799502373, + 0.0030236756429076195, + 0.003753222757950425, + 0.008935624733567238, + 0.0022251103073358536, + 0.002588908188045025, + 0.003371545812115073, + 0.009553462266921997, + 0.00508081866428256, + 0.004489379934966564, + 0.007826464250683784, + 0.013781879097223282, + 0.005197826307266951, + 0.004692836198955774, + 0.009721020236611366, + 0.007542419712990522, + 0.01677108369767666, + 0.01798217184841633, + 0.019997654482722282, + 0.017078572884202003, + 0.05208129808306694, + 0.017355334013700485, + 0.04037800431251526, + 0.046065252274274826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.05495181307196617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864853248000145 + ], + "opposite_logprob_tokens": [ + -3.0486485958099365 + ], + "logprob_reward": -0.04864853248000145, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864853248000145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26733794808387756, + 0.12297412008047104, + 0.05240059271454811, + 0.053404550999403, + 0.046597570180892944, + 0.0452236533164978, + 0.09668228775262833, + 0.04419216141104698, + 0.085805743932724, + 0.18538135290145874 + ], + "credit_map": [ + 0.030596431344747543, + 0.03755030035972595, + 0.033573586493730545, + 0.04283209890127182, + 0.030876504257321358, + 0.02764979936182499, + 0.03547254577279091, + 0.028786703944206238, + 0.013893872499465942, + 0.014000678434967995, + 0.01856308989226818, + 0.019497990608215332, + 0.014263429678976536, + 0.008679342456161976, + 0.016374194994568825, + 0.017701534554362297, + 0.004162003751844168, + 0.005476485006511211, + 0.00996227003633976, + 0.007243869826197624, + 0.005352687556296587, + 0.007380567956715822, + 0.004568359348922968, + 0.00825435109436512, + 0.004885430913418531, + 0.008118956349790096, + 0.008585315197706223, + 0.008073164150118828, + 0.005862670484930277, + 0.0068246121518313885, + 0.003981956746429205, + 0.007072445936501026, + 0.004429834894835949, + 0.005878068972378969, + 0.005591639317572117, + 0.008930634707212448, + 0.004569195210933685, + 0.005907457787543535, + 0.0045547280460596085, + 0.006736009381711483, + 0.003500616643577814, + 0.0067540984600782394, + 0.0064238314516842365, + 0.008017146028578281, + 0.004291885066777468, + 0.005246609915047884, + 0.005063696764409542, + 0.0059257736429572105, + 0.00843239575624466, + 0.008060650900006294, + 0.010394877754151821, + 0.010954846628010273, + 0.013048246502876282, + 0.010522290132939816, + 0.011457188986241817, + 0.02381180226802826, + 0.00364607572555542, + 0.003172218333929777, + 0.006106958724558353, + 0.010689266957342625, + 0.0031195555347949266, + 0.0043075960129499435, + 0.004291081801056862, + 0.008859412744641304, + 0.009249560534954071, + 0.008533162996172905, + 0.008768166415393353, + 0.0180952325463295, + 0.008891323581337929, + 0.008378446102142334, + 0.013309292495250702, + 0.010580567643046379, + 0.01721573807299137, + 0.018514147028326988, + 0.017093494534492493, + 0.02037619799375534, + 0.028005432337522507, + 0.015606020577251911, + 0.04167461767792702, + 0.02689572423696518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.04864853248000145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278101921081543 + ], + "opposite_logprob_tokens": [ + -2.0177810192108154 + ], + "logprob_reward": -0.14278101921081543, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278101921081543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1516098976135254, + 0.08015822619199753, + 0.04579155519604683, + 0.04515272006392479, + 0.08706358075141907, + 0.06625137478113174, + 0.06699321419000626, + 0.07711438089609146, + 0.09222337603569031, + 0.2876416742801666 + ], + "credit_map": [ + 0.01624520681798458, + 0.014724794775247574, + 0.021894212812185287, + 0.018268870189785957, + 0.01335546188056469, + 0.018330033868551254, + 0.020222138613462448, + 0.028569186106324196, + 0.00852201133966446, + 0.006250416859984398, + 0.015147964470088482, + 0.012553572654724121, + 0.008054223842918873, + 0.010603434406220913, + 0.010128392837941647, + 0.008898211643099785, + 0.004726398270577192, + 0.004076044075191021, + 0.007830551825463772, + 0.0056016710586845875, + 0.004399466793984175, + 0.006178729236125946, + 0.004180131945759058, + 0.008798562921583652, + 0.004912642762064934, + 0.0043486738577485085, + 0.007298596203327179, + 0.006095851305872202, + 0.0042112283408641815, + 0.005790816154330969, + 0.004594632424414158, + 0.007900279015302658, + 0.010770375840365887, + 0.01026267558336258, + 0.01771301031112671, + 0.016322124749422073, + 0.00487072579562664, + 0.006243868265300989, + 0.00850867759436369, + 0.012372124008834362, + 0.006005878560245037, + 0.009831219911575317, + 0.007936165668070316, + 0.011251579970121384, + 0.0065141585655510426, + 0.006219347007572651, + 0.009362378157675266, + 0.009130647405982018, + 0.005991016514599323, + 0.006254115607589483, + 0.006586594041436911, + 0.007823800668120384, + 0.008467277511954308, + 0.00949608813971281, + 0.008436513133347034, + 0.013937806710600853, + 0.007108623161911964, + 0.007369405124336481, + 0.00866126362234354, + 0.013562035746872425, + 0.006537489127367735, + 0.00753300404176116, + 0.009723707102239132, + 0.016618847846984863, + 0.008142122067511082, + 0.010603279806673527, + 0.010338257998228073, + 0.01529382262378931, + 0.01008190494030714, + 0.012608756311237812, + 0.012392033822834492, + 0.01276319194585085, + 0.02907237410545349, + 0.02824671007692814, + 0.047390248626470566, + 0.02780192159116268, + 0.03412371873855591, + 0.027894211933016777, + 0.04865081608295441, + 0.0444616861641407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.14278101921081543 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from an upward angle, instantly transports us into a fantastical, Art Nouveau-inspired metal forest. Here, massive \"trees\" of twisted bronze and brass rise from the ground, their branches spreading like vines, their ends terminating in soft, lily-shaped lights. At the center of the frame, a green-haired swordsman lowers his weight into an explosive attacking stance, his bright green, shoulder-length hair striking against the warm glow of the metallic vegetation. His expression is stern and focused, and in his mouth he clutches a renowned short sword, its hilt entwined with ancient runes, its blade gleaming with a cold gleam. He wears a dark green, skin-tight leather armor embroidered with golden vines, and in his hands he holds two sleek, silver scimitars in reverse. Directly before him stands a massive, ornate column crafted entirely of polished brass, its shaft carved with intricate, interwoven reliefs of vibrant curves. The swordsman exerted force from his waist, and with the force of thunder, his arms slashed out in two swift and deadly silver arcs. The blades precisely sliced into the middle of the brass pistil. With a sharp metallic cutting sound and a burst of dazzling sparks, the massive pistil was neatly bisected. As the severed upper half began to slowly and heavily tilt, finally collapsing to the ground with a loud thud, the camera experienced a sharp and brief tremor, perfectly conveying the unparalleled impact and weight of the giant's collapse.", + "seed": 12850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_4/index_7814_seed_12850_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": -0.05290882494300604, + "vlm_logprob": -0.05290882494300604, + "vlm_yes_no_margin": 3.399999976158142, + "vlm_token_credit": -0.05290882494300604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01120524387806654 + ], + "opposite_logprob_tokens": [ + -5.136205196380615 + ], + "logprob_reward": -0.01120524387806654, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01120524387806654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17300787568092346, + 0.06035744771361351, + 0.07694917172193527, + 0.03878174349665642, + 0.03237466141581535, + 0.053743068128824234, + 0.1401924043893814, + 0.231421560049057, + 0.04945547506213188, + 0.14371658861637115 + ], + "credit_map": [ + 0.017371023073792458, + 0.02761448360979557, + 0.041306961327791214, + 0.01925509423017502, + 0.012422974221408367, + 0.026679717004299164, + 0.014257825911045074, + 0.01409979909658432, + 0.006270965561270714, + 0.011321417056024075, + 0.00640213442966342, + 0.007039589807391167, + 0.00551904970780015, + 0.007629137486219406, + 0.00931449607014656, + 0.006860655266791582, + 0.00568776112049818, + 0.005840763915330172, + 0.006519087124615908, + 0.02620891109108925, + 0.004427067004144192, + 0.004667248111218214, + 0.005498090758919716, + 0.018100237473845482, + 0.005631412845104933, + 0.003443048568442464, + 0.00558078708127141, + 0.00508735328912735, + 0.002987236948683858, + 0.0026854367461055517, + 0.0026880167424678802, + 0.010678448714315891, + 0.003049972467124462, + 0.002659421181306243, + 0.0023329213727265596, + 0.009929492138326168, + 0.003177030012011528, + 0.0021190568804740906, + 0.002150035696104169, + 0.0069567314349114895, + 0.005043622571974993, + 0.004092852119356394, + 0.003461450105533004, + 0.01752745732665062, + 0.00469461502507329, + 0.0024516796693205833, + 0.0037950898986309767, + 0.012676301412284374, + 0.011699063703417778, + 0.011117608286440372, + 0.009809136390686035, + 0.05210794880986214, + 0.005691026337444782, + 0.00491838064044714, + 0.009327183477580547, + 0.03552206978201866, + 0.02402505837380886, + 0.030043289065361023, + 0.034967198967933655, + 0.03752313554286957, + 0.020992690697312355, + 0.026924889534711838, + 0.023965585976839066, + 0.032979708164930344, + 0.007912223227322102, + 0.00324691622518003, + 0.004434117581695318, + 0.016300413757562637, + 0.0019116700859740376, + 0.002439816715195775, + 0.004819550085812807, + 0.008390765637159348, + 0.00941296387463808, + 0.01345518697053194, + 0.01025332324206829, + 0.04081687331199646, + 0.015533738769590855, + 0.011002459563314915, + 0.014472794719040394, + 0.02876923978328705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.01120524387806654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820521429181099 + ], + "opposite_logprob_tokens": [ + -4.018205165863037 + ], + "logprob_reward": -0.01820521429181099, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820521429181099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19947223365306854, + 0.10844944417476654, + 0.11848784238100052, + 0.060872748494148254, + 0.05462358519434929, + 0.06043683737516403, + 0.0824531689286232, + 0.09027304500341415, + 0.06194940581917763, + 0.16298167407512665 + ], + "credit_map": [ + 0.034559279680252075, + 0.013241668231785297, + 0.06598398089408875, + 0.019035158678889275, + 0.015822209417819977, + 0.027052894234657288, + 0.008949131704866886, + 0.014827918261289597, + 0.010515733622014523, + 0.02522299624979496, + 0.008580580353736877, + 0.01350304577499628, + 0.012558582238852978, + 0.012274758890271187, + 0.015577259473502636, + 0.010216486640274525, + 0.006032498087733984, + 0.008926562033593655, + 0.009089887142181396, + 0.04106607288122177, + 0.00425232807174325, + 0.00638276943936944, + 0.007425071205943823, + 0.035312652587890625, + 0.006448056548833847, + 0.0037455775309354067, + 0.005363308358937502, + 0.022500455379486084, + 0.0022646973375231028, + 0.0030540996231138706, + 0.004322375170886517, + 0.013174178078770638, + 0.004414318595081568, + 0.0040899463929235935, + 0.004263793118298054, + 0.01897127367556095, + 0.0024819145910441875, + 0.0031366893090307713, + 0.0036795863416045904, + 0.01358606293797493, + 0.004347167443484068, + 0.004264179151505232, + 0.004099549725651741, + 0.02082126773893833, + 0.003258225740864873, + 0.002596568316221237, + 0.004184271674603224, + 0.01686561107635498, + 0.006774003151804209, + 0.006549467798322439, + 0.007698019966483116, + 0.029308142140507698, + 0.003204440465196967, + 0.004103638231754303, + 0.00538366474211216, + 0.019431794062256813, + 0.008960717357695103, + 0.010631891898810863, + 0.012609407305717468, + 0.014364296570420265, + 0.009037654846906662, + 0.011858669109642506, + 0.009612046182155609, + 0.013198353350162506, + 0.006939293351024389, + 0.004407644271850586, + 0.005155777093023062, + 0.022887451574206352, + 0.00220765289850533, + 0.0022737248800694942, + 0.009873946197330952, + 0.008203918114304543, + 0.013201217167079449, + 0.01374690979719162, + 0.014170832931995392, + 0.03725787252187729, + 0.022928502410650253, + 0.014248127117753029, + 0.02395635098218918, + 0.023471862077713013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.01820521429181099 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097988814115524 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.038097988814115524, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097988814115524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14412660896778107, + 0.049750640988349915, + 0.20162038505077362, + 0.07471423596143723, + 0.05414016917347908, + 0.08557267487049103, + 0.018877997994422913, + 0.17883501946926117, + 0.06192849576473236, + 0.1304338425397873 + ], + "credit_map": [ + 0.018988225609064102, + 0.021582655608654022, + 0.03549608215689659, + 0.014210222288966179, + 0.015410901978611946, + 0.01781175285577774, + 0.009459390304982662, + 0.01116736140102148, + 0.004666250664740801, + 0.013830926269292831, + 0.0060221245512366295, + 0.00634921295568347, + 0.003389295656234026, + 0.0034770513884723186, + 0.005606647115200758, + 0.006409128196537495, + 0.013920089229941368, + 0.014351505786180496, + 0.013455303385853767, + 0.060809843242168427, + 0.01468962524086237, + 0.012972372584044933, + 0.01296843308955431, + 0.058453191071748734, + 0.005617458838969469, + 0.004845998249948025, + 0.006290275137871504, + 0.022013776004314423, + 0.0031857492867857218, + 0.006335115525871515, + 0.004921299405395985, + 0.021504556760191917, + 0.003681825241073966, + 0.0036575221456587315, + 0.004497947171330452, + 0.01666972041130066, + 0.0026000402867794037, + 0.0044154333882033825, + 0.003727348754182458, + 0.01489032618701458, + 0.005258696153759956, + 0.006482016760855913, + 0.006452897563576698, + 0.02637503668665886, + 0.006033489014953375, + 0.004232055973261595, + 0.005603137891739607, + 0.02513534016907215, + 0.0027470083441585302, + 0.001777150435373187, + 0.0019864412024617195, + 0.003940075635910034, + 0.0016892942367121577, + 0.0013087836559861898, + 0.0013334495015442371, + 0.004095793701708317, + 0.0175135787576437, + 0.023260124027729034, + 0.025722380727529526, + 0.02794824168086052, + 0.017276741564273834, + 0.021265242248773575, + 0.01976298913359642, + 0.026085708290338516, + 0.005052348133176565, + 0.004131145775318146, + 0.005659922957420349, + 0.022490669041872025, + 0.0030642375349998474, + 0.0034093663562089205, + 0.004288188647478819, + 0.013832605443894863, + 0.007300728466361761, + 0.009999813511967659, + 0.009347372688353062, + 0.04103425517678261, + 0.011369372718036175, + 0.008586387149989605, + 0.01200152188539505, + 0.030794380232691765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.038097988814115524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001281529664993 + ], + "opposite_logprob_tokens": [ + -2.6950128078460693 + ], + "logprob_reward": -0.07001281529664993, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001281529664993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16714084148406982, + 0.04738089069724083, + 0.18201391398906708, + 0.08481569588184357, + 0.05589660257101059, + 0.055927932262420654, + 0.06626501679420471, + 0.027049537748098373, + 0.05817848816514015, + 0.2553310990333557 + ], + "credit_map": [ + 0.03443372994661331, + 0.013371686451137066, + 0.05521165579557419, + 0.014505070634186268, + 0.012317381799221039, + 0.01668858341872692, + 0.007409390062093735, + 0.013203336857259274, + 0.0038221722934395075, + 0.007411264348775148, + 0.00613717595115304, + 0.005371416453272104, + 0.0053586894646286964, + 0.008793300949037075, + 0.004883831366896629, + 0.005603045225143433, + 0.011375010944902897, + 0.010537232272326946, + 0.01114205177873373, + 0.06922417879104614, + 0.0055252681486308575, + 0.008041045628488064, + 0.0113186901435256, + 0.05485042184591293, + 0.005378499161452055, + 0.005724618211388588, + 0.006449967622756958, + 0.02553405612707138, + 0.0049567753449082375, + 0.007472501136362553, + 0.005385000724345446, + 0.02391427755355835, + 0.0035927705466747284, + 0.0041052838787436485, + 0.004271169658750296, + 0.016965286806225777, + 0.0040011899545788765, + 0.004523701500147581, + 0.003555749775841832, + 0.014881451614201069, + 0.003387091215699911, + 0.00437190942466259, + 0.00456341402605176, + 0.016080567613244057, + 0.00470732431858778, + 0.003072147024795413, + 0.0037534027360379696, + 0.015992075204849243, + 0.0056394957937300205, + 0.005841013044118881, + 0.004933296237140894, + 0.023915262892842293, + 0.00372250284999609, + 0.00243988330475986, + 0.0043691424652934074, + 0.015404416248202324, + 0.0021858022082597017, + 0.003456349950283766, + 0.0030215869192034006, + 0.004393748473376036, + 0.0032177050597965717, + 0.003176380880177021, + 0.003798181191086769, + 0.0037997823674231768, + 0.0039555709809064865, + 0.0033339648507535458, + 0.005043542478233576, + 0.02129175327718258, + 0.0023284319322556257, + 0.00269176228903234, + 0.008258441463112831, + 0.011275021359324455, + 0.013320312835276127, + 0.013344908133149147, + 0.017238525673747063, + 0.08780669420957565, + 0.0198984332382679, + 0.0150682982057333, + 0.021986795589327812, + 0.06666714698076248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.07001281529664993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270228624343872 + ], + "opposite_logprob_tokens": [ + -2.1270227432250977 + ], + "logprob_reward": -0.1270228624343872, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270228624343872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3038617670536041, + 0.03985113650560379, + 0.09868350625038147, + 0.11469626426696777, + 0.02297314815223217, + 0.021390635520219803, + 0.02634485997259617, + 0.1648796945810318, + 0.052510038018226624, + 0.154808908700943 + ], + "credit_map": [ + 0.06412068754434586, + 0.053364697843790054, + 0.07350882142782211, + 0.03161215782165527, + 0.014156370423734188, + 0.01988723874092102, + 0.01991145871579647, + 0.027300333604216576, + 0.0033311834558844566, + 0.008522854186594486, + 0.0043843891471624374, + 0.005435615312308073, + 0.004130168817937374, + 0.004229930229485035, + 0.004651111084967852, + 0.0051658861339092255, + 0.006258148234337568, + 0.007121853064745665, + 0.006461995653808117, + 0.029108885675668716, + 0.007781559135764837, + 0.005540181417018175, + 0.00670655956491828, + 0.029704328626394272, + 0.009132028557360172, + 0.0077333152294158936, + 0.009009663946926594, + 0.03623585402965546, + 0.005242484621703625, + 0.007833401672542095, + 0.007618619129061699, + 0.031890902668237686, + 0.002089251996949315, + 0.002814090345054865, + 0.0035023444797843695, + 0.004542170558124781, + 0.002955074654892087, + 0.0023572142235934734, + 0.0018039665883406997, + 0.0029090361203998327, + 0.003264385275542736, + 0.001632173079997301, + 0.002174076857045293, + 0.002984727034345269, + 0.002683906350284815, + 0.0021161313634365797, + 0.001665023504756391, + 0.004870213568210602, + 0.0031817699782550335, + 0.0016602608375251293, + 0.002251469297334552, + 0.004673270042985678, + 0.0029481197707355022, + 0.003461570246145129, + 0.001973128644749522, + 0.00619527418166399, + 0.016425846144557, + 0.02299770899116993, + 0.023580413311719894, + 0.02638617530465126, + 0.015462708659470081, + 0.01915981061756611, + 0.017615042626857758, + 0.023252001032233238, + 0.007815206423401833, + 0.0032532361801713705, + 0.004414319060742855, + 0.018564114347100258, + 0.001993500627577305, + 0.0017796108731999993, + 0.009480413980782032, + 0.005209640134125948, + 0.008466684259474277, + 0.00903358031064272, + 0.010280884802341461, + 0.05489607900381088, + 0.011424258351325989, + 0.00722863245755434, + 0.011915789917111397, + 0.0415630005300045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.1270228624343872 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from an upward angle, instantly transports us into a fantastical, Art Nouveau-inspired metal forest. Here, massive \"trees\" of twisted bronze and brass rise from the ground, their branches spreading like vines, their ends terminating in soft, lily-shaped lights. At the center of the frame, a green-haired swordsman lowers his weight into an explosive attacking stance, his bright green, shoulder-length hair striking against the warm glow of the metallic vegetation. His expression is stern and focused, and in his mouth he clutches a renowned short sword, its hilt entwined with ancient runes, its blade gleaming with a cold gleam. He wears a dark green, skin-tight leather armor embroidered with golden vines, and in his hands he holds two sleek, silver scimitars in reverse. Directly before him stands a massive, ornate column crafted entirely of polished brass, its shaft carved with intricate, interwoven reliefs of vibrant curves. The swordsman exerted force from his waist, and with the force of thunder, his arms slashed out in two swift and deadly silver arcs. The blades precisely sliced into the middle of the brass pistil. With a sharp metallic cutting sound and a burst of dazzling sparks, the massive pistil was neatly bisected. As the severed upper half began to slowly and heavily tilt, finally collapsing to the ground with a loud thud, the camera experienced a sharp and brief tremor, perfectly conveying the unparalleled impact and weight of the giant's collapse.", + "seed": 12851, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_4/index_7814_seed_12851_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": -0.05756262820214033, + "vlm_logprob": -0.05756262820214033, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.05756262820214033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008259071037173271 + ], + "opposite_logprob_tokens": [ + -5.383259296417236 + ], + "logprob_reward": -0.008259071037173271, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.008259071037173271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10674293339252472, + 0.028158612549304962, + 0.022252660244703293, + 0.027575919404625893, + 0.033140551298856735, + 0.6020771861076355, + 0.03497282788157463, + 0.055507976561784744, + 0.06997808814048767, + 0.019593188539147377 + ], + "credit_map": [ + 0.010994775220751762, + 0.011684088036417961, + 0.015178516507148743, + 0.011524565517902374, + 0.02350832335650921, + 0.012154330499470234, + 0.0076437294483184814, + 0.014054625295102596, + 0.004063230473548174, + 0.0020527890883386135, + 0.00365042663179338, + 0.002441048389300704, + 0.006242308299988508, + 0.003349324455484748, + 0.002390185371041298, + 0.0039693028666079044, + 0.0026564074214547873, + 0.0019218978704884648, + 0.004324744921177626, + 0.0032386519014835358, + 0.0022106932010501623, + 0.001681030262261629, + 0.0016116616316139698, + 0.0046075740829110146, + 0.003689417615532875, + 0.0028491131961345673, + 0.00465012900531292, + 0.003086651209741831, + 0.005462528672069311, + 0.002812702674418688, + 0.002189746592193842, + 0.002835633931681514, + 0.00376342236995697, + 0.002608951646834612, + 0.008341329172253609, + 0.0034001432359218597, + 0.0044496688060462475, + 0.0034172385931015015, + 0.003248434280976653, + 0.003911369480192661, + 0.038750894367694855, + 0.03715040534734726, + 0.1907718926668167, + 0.0680273175239563, + 0.04709241911768913, + 0.0427187941968441, + 0.03595046326518059, + 0.14161506295204163, + 0.005919222719967365, + 0.004277215804904699, + 0.004419542849063873, + 0.004948379006236792, + 0.002129092812538147, + 0.005017162300646305, + 0.0030024321749806404, + 0.005259783938527107, + 0.006121666636317968, + 0.005882274359464645, + 0.005382626783102751, + 0.008017195388674736, + 0.004579077009111643, + 0.007422042544931173, + 0.011036140844225883, + 0.007066956255584955, + 0.00492051150649786, + 0.007908377796411514, + 0.00591912679374218, + 0.008436422795057297, + 0.01242409460246563, + 0.012339470908045769, + 0.011493363417685032, + 0.0065367273055016994, + 0.0021115075796842575, + 0.0027639097534120083, + 0.002172870095819235, + 0.002286561531946063, + 0.001608779188245535, + 0.0034201645758002996, + 0.0031429824884980917, + 0.00208641542121768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.008259071037173271 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087474413216114 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087474413216114, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087474413216114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2958155572414398, + 0.08203400671482086, + 0.04870874434709549, + 0.0642859935760498, + 0.03660750761628151, + 0.17881058156490326, + 0.06530816853046417, + 0.044672127813100815, + 0.07529722154140472, + 0.10846009850502014 + ], + "credit_map": [ + 0.025636982172727585, + 0.02180304005742073, + 0.028883161023259163, + 0.022535385563969612, + 0.08534093201160431, + 0.021669775247573853, + 0.017957568168640137, + 0.07198872417211533, + 0.009262208826839924, + 0.009809575974941254, + 0.01658490113914013, + 0.01551196351647377, + 0.008070528507232666, + 0.005910512991249561, + 0.0070055280812084675, + 0.009878777898848057, + 0.0072638350538909435, + 0.005978561472147703, + 0.007159641478210688, + 0.008009367622435093, + 0.00577545864507556, + 0.004179603885859251, + 0.00624736025929451, + 0.004094913136214018, + 0.005388868506997824, + 0.0042799594812095165, + 0.008358984254300594, + 0.007444105111062527, + 0.01678769290447235, + 0.005423775874078274, + 0.007461567874997854, + 0.009141044691205025, + 0.0041269659996032715, + 0.004811684135347605, + 0.0048419274389743805, + 0.005410656798630953, + 0.004383752588182688, + 0.004291222430765629, + 0.004420922137796879, + 0.0043203807435929775, + 0.011309240944683552, + 0.009958150796592236, + 0.05806558579206467, + 0.020256467163562775, + 0.012545892037451267, + 0.012795301154255867, + 0.010121157392859459, + 0.04375879094004631, + 0.007036630995571613, + 0.008916990831494331, + 0.01068356167525053, + 0.01037351693958044, + 0.008111868053674698, + 0.005494487471878529, + 0.006652490235865116, + 0.008038626052439213, + 0.004052971489727497, + 0.0038161040283739567, + 0.004691608715802431, + 0.008386574685573578, + 0.003942365292459726, + 0.0052737025544047356, + 0.00827884767204523, + 0.006229954306036234, + 0.007500488311052322, + 0.007600201293826103, + 0.00854479894042015, + 0.013484764844179153, + 0.011232594959437847, + 0.0054906196892261505, + 0.013210980221629143, + 0.008232765831053257, + 0.011086276732385159, + 0.012375022284686565, + 0.01268036849796772, + 0.014585096389055252, + 0.009464753791689873, + 0.016162872314453125, + 0.022188035771250725, + 0.009917665272951126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.011087474413216114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097184151411057 + ], + "opposite_logprob_tokens": [ + -3.2880971431732178 + ], + "logprob_reward": -0.038097184151411057, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097184151411057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2992779314517975, + 0.05568508431315422, + 0.03115130588412285, + 0.0451979823410511, + 0.025461936369538307, + 0.17547033727169037, + 0.08623873442411423, + 0.05186641216278076, + 0.09449566155672073, + 0.1351546198129654 + ], + "credit_map": [ + 0.019287746399641037, + 0.019527073949575424, + 0.028841927647590637, + 0.020370246842503548, + 0.09480028599500656, + 0.021368447691202164, + 0.01923448219895363, + 0.07584772258996964, + 0.006309774704277515, + 0.006143678445369005, + 0.010908549651503563, + 0.009903732687234879, + 0.005154540296643972, + 0.004274845123291016, + 0.005026719067245722, + 0.007963243871927261, + 0.0041476828046143055, + 0.005092047620564699, + 0.0048759193159639835, + 0.00523753184825182, + 0.0032262715976685286, + 0.0030192153062671423, + 0.0029488776344805956, + 0.0026037588249891996, + 0.005019962787628174, + 0.0063054547645151615, + 0.005800691898912191, + 0.007735887076705694, + 0.005945661570876837, + 0.003352029249072075, + 0.0071459002792835236, + 0.003892391687259078, + 0.0035555341746658087, + 0.0037133393343538046, + 0.00478120893239975, + 0.003962332382798195, + 0.0023501713294535875, + 0.0025688542518764734, + 0.0021239526104182005, + 0.002406542655080557, + 0.011394655331969261, + 0.010630344040691853, + 0.05862979590892792, + 0.0209891889244318, + 0.011751065030694008, + 0.012905015610158443, + 0.008262748830020428, + 0.04090752452611923, + 0.008455654606223106, + 0.010221580974757671, + 0.013964670710265636, + 0.018317464739084244, + 0.011721726506948471, + 0.005614969879388809, + 0.00814064871519804, + 0.009802020154893398, + 0.004918727092444897, + 0.006018404383212328, + 0.005229657981544733, + 0.011204870417714119, + 0.0041125062853097916, + 0.00669645331799984, + 0.007463816087692976, + 0.006221977062523365, + 0.011057705618441105, + 0.009745963849127293, + 0.011267488822340965, + 0.016958916559815407, + 0.010985179804265499, + 0.015756133943796158, + 0.011333574540913105, + 0.007390701677650213, + 0.01514330692589283, + 0.016829924657940865, + 0.01602832041680813, + 0.018233386799693108, + 0.012459362857043743, + 0.020357953384518623, + 0.02304716780781746, + 0.013055184856057167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.038097184151411057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002303749322891 + ], + "opposite_logprob_tokens": [ + -2.6950230598449707 + ], + "logprob_reward": -0.07002303749322891, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002303749322891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2261507213115692, + 0.07421804964542389, + 0.017825281247496605, + 0.08016334474086761, + 0.024246636778116226, + 0.2686397433280945, + 0.10654687881469727, + 0.05701473355293274, + 0.07306224852800369, + 0.07213239371776581 + ], + "credit_map": [ + 0.02029183879494667, + 0.016346564516425133, + 0.022674493491649628, + 0.016610581427812576, + 0.06402113288640976, + 0.016263455152511597, + 0.014314946718513966, + 0.05562771111726761, + 0.008406720124185085, + 0.01062053069472313, + 0.013339015655219555, + 0.014571218751370907, + 0.009514742530882359, + 0.005326162092387676, + 0.005169020965695381, + 0.0072706411592662334, + 0.0033472308423370123, + 0.0024287961423397064, + 0.0018274069298058748, + 0.00295258522965014, + 0.001893874374218285, + 0.0016704985173419118, + 0.0017498215893283486, + 0.001955067738890648, + 0.008259963244199753, + 0.00808974914252758, + 0.011950088664889336, + 0.012971639633178711, + 0.01237682905048132, + 0.005413932725787163, + 0.013278751634061337, + 0.007822388783097267, + 0.0038126029539853334, + 0.002125423401594162, + 0.004769416991621256, + 0.0027045824099332094, + 0.002806571079418063, + 0.00283634290099144, + 0.0019666925072669983, + 0.003225005231797695, + 0.01685008965432644, + 0.01374347135424614, + 0.0903063639998436, + 0.03157791867852211, + 0.01776561327278614, + 0.01923322305083275, + 0.012264817021787167, + 0.06689824908971786, + 0.009513822384178638, + 0.012691175565123558, + 0.01942063309252262, + 0.02217419072985649, + 0.01487206295132637, + 0.0075218090787529945, + 0.00846942886710167, + 0.011883758939802647, + 0.007029253989458084, + 0.008815758861601353, + 0.0058573163114488125, + 0.010514424182474613, + 0.003924309276044369, + 0.008103364147245884, + 0.006320868153125048, + 0.006449435837566853, + 0.009644228033721447, + 0.006627965718507767, + 0.007840784266591072, + 0.00930215884000063, + 0.011731753125786781, + 0.01123818103224039, + 0.010419332422316074, + 0.006257843691855669, + 0.007623143028467894, + 0.009747111238539219, + 0.008358152583241463, + 0.008993112482130527, + 0.0064696683548390865, + 0.01003152783960104, + 0.013904298655688763, + 0.007005379069596529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.07002303749322891 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603463739156723 + ], + "opposite_logprob_tokens": [ + -1.9103463888168335 + ], + "logprob_reward": -0.1603463739156723, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603463739156723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1393294483423233, + 0.1045520007610321, + 0.031225044280290604, + 0.023855818435549736, + 0.03643787279725075, + 0.1968139410018921, + 0.12379758805036545, + 0.15457406640052795, + 0.14133650064468384, + 0.048077698796987534 + ], + "credit_map": [ + 0.015644477680325508, + 0.011110587976872921, + 0.01335042156279087, + 0.014127044938504696, + 0.030302980914711952, + 0.0165697168558836, + 0.008284985087811947, + 0.029939234256744385, + 0.011749106459319592, + 0.011358731426298618, + 0.019728325307369232, + 0.018031427636742592, + 0.01047317311167717, + 0.012177323922514915, + 0.007896233350038528, + 0.01313768606632948, + 0.005720676388591528, + 0.0026656610425561666, + 0.005209591705352068, + 0.00416968297213316, + 0.0029369378462433815, + 0.004249300807714462, + 0.0024615100119262934, + 0.003811684437096119, + 0.002932110568508506, + 0.0019372499082237482, + 0.0031771145295351744, + 0.0026845235843211412, + 0.004184947814792395, + 0.002655247924849391, + 0.0026452753227204084, + 0.003639349015429616, + 0.004771084059029818, + 0.004823262337595224, + 0.006351127754896879, + 0.005031217820942402, + 0.0037273543421179056, + 0.0037943017669022083, + 0.004110353998839855, + 0.003829168388620019, + 0.013952121138572693, + 0.009187908843159676, + 0.04816877841949463, + 0.018599729984998703, + 0.017976973205804825, + 0.014359282329678535, + 0.023864487186074257, + 0.05070466548204422, + 0.012464030645787716, + 0.01903468929231167, + 0.017552824690937996, + 0.021594563499093056, + 0.021159019321203232, + 0.009736612439155579, + 0.010080518200993538, + 0.01217533741146326, + 0.01581445336341858, + 0.014169352129101753, + 0.01624518260359764, + 0.02613905817270279, + 0.013355329632759094, + 0.019179873168468475, + 0.030261054635047913, + 0.01940976269543171, + 0.011087078601121902, + 0.014499776996672153, + 0.012796965427696705, + 0.012908454053103924, + 0.013371744193136692, + 0.03368445858359337, + 0.016795020550489426, + 0.026193002238869667, + 0.004198512062430382, + 0.00567179499194026, + 0.004333339631557465, + 0.010034006088972092, + 0.0038471082225441933, + 0.005585165694355965, + 0.00996683444827795, + 0.004440934397280216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.1603463739156723 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from an upward angle, instantly transports us into a fantastical, Art Nouveau-inspired metal forest. Here, massive \"trees\" of twisted bronze and brass rise from the ground, their branches spreading like vines, their ends terminating in soft, lily-shaped lights. At the center of the frame, a green-haired swordsman lowers his weight into an explosive attacking stance, his bright green, shoulder-length hair striking against the warm glow of the metallic vegetation. His expression is stern and focused, and in his mouth he clutches a renowned short sword, its hilt entwined with ancient runes, its blade gleaming with a cold gleam. He wears a dark green, skin-tight leather armor embroidered with golden vines, and in his hands he holds two sleek, silver scimitars in reverse. Directly before him stands a massive, ornate column crafted entirely of polished brass, its shaft carved with intricate, interwoven reliefs of vibrant curves. The swordsman exerted force from his waist, and with the force of thunder, his arms slashed out in two swift and deadly silver arcs. The blades precisely sliced into the middle of the brass pistil. With a sharp metallic cutting sound and a burst of dazzling sparks, the massive pistil was neatly bisected. As the severed upper half began to slowly and heavily tilt, finally collapsing to the ground with a loud thud, the camera experienced a sharp and brief tremor, perfectly conveying the unparalleled impact and weight of the giant's collapse.", + "seed": 12852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_4/index_7814_seed_12852_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": -0.03918051421642303, + "vlm_logprob": -0.03918051421642303, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03918051421642303, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008736955933272839 + ], + "opposite_logprob_tokens": [ + -5.383737087249756 + ], + "logprob_reward": -0.008736955933272839, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.008736955933272839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13076162338256836, + 0.09693427383899689, + 0.06448356062173843, + 0.028589170426130295, + 0.24209648370742798, + 0.04654562100768089, + 0.1958385407924652, + 0.046863049268722534, + 0.09092966467142105, + 0.05695803835988045 + ], + "credit_map": [ + 0.01327136717736721, + 0.019351009279489517, + 0.013934936374425888, + 0.015910079702734947, + 0.017829755321145058, + 0.018002700060606003, + 0.01894466020166874, + 0.013517097570002079, + 0.010397572070360184, + 0.008379681967198849, + 0.02116283029317856, + 0.010098128579556942, + 0.019270701333880424, + 0.008599186316132545, + 0.010373888537287712, + 0.008652271702885628, + 0.00804835557937622, + 0.007154486607760191, + 0.007937856949865818, + 0.0059682209976017475, + 0.011234046891331673, + 0.008261183276772499, + 0.007539832964539528, + 0.00833957176655531, + 0.003043348202481866, + 0.003105090232565999, + 0.003911261912435293, + 0.0030613273847848177, + 0.0037699202075600624, + 0.0028803925961256027, + 0.004122538957744837, + 0.004695289768278599, + 0.021406671032309532, + 0.021473567932844162, + 0.06702520698308945, + 0.012121886946260929, + 0.07554274797439575, + 0.009079710580408573, + 0.019041158258914948, + 0.01640550047159195, + 0.005024965852499008, + 0.006340966559946537, + 0.00641496479511261, + 0.007388202007859945, + 0.005247159861028194, + 0.004502869211137295, + 0.006958141457289457, + 0.0046683489345014095, + 0.01792238838970661, + 0.00878014974296093, + 0.04360264167189598, + 0.01764303259551525, + 0.06135093793272972, + 0.010840076953172684, + 0.020163729786872864, + 0.015535564161837101, + 0.003908457700163126, + 0.004860643297433853, + 0.006846940610557795, + 0.007937098853290081, + 0.003713555634021759, + 0.006371402647346258, + 0.006185627076774836, + 0.007039323449134827, + 0.008921390399336815, + 0.013567253015935421, + 0.012698381207883358, + 0.01143006794154644, + 0.00767881516367197, + 0.010041642934083939, + 0.016232062131166458, + 0.010360048152506351, + 0.012498808093369007, + 0.007848375476896763, + 0.009507586248219013, + 0.00534374313428998, + 0.004780309274792671, + 0.005719031672924757, + 0.005259253084659576, + 0.006000927183777094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.008736955933272839 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263565294444561 + ], + "opposite_logprob_tokens": [ + -3.6513564586639404 + ], + "logprob_reward": -0.0263565294444561, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263565294444561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09237848222255707, + 0.20465166866779327, + 0.045524850487709045, + 0.04452800005674362, + 0.3593873679637909, + 0.04221763834357262, + 0.061580561101436615, + 0.021094847470521927, + 0.0745289996266365, + 0.05410761013627052 + ], + "credit_map": [ + 0.006328372284770012, + 0.012746714986860752, + 0.011549152433872223, + 0.011033873073756695, + 0.021776368841528893, + 0.008495833724737167, + 0.008650005795061588, + 0.011798156425356865, + 0.014102235436439514, + 0.03023075871169567, + 0.04035469889640808, + 0.020579811185598373, + 0.058972060680389404, + 0.009733016602694988, + 0.017302507534623146, + 0.013376566581428051, + 0.006658068858087063, + 0.004182735923677683, + 0.005944622214883566, + 0.004613231401890516, + 0.004766242112964392, + 0.007357478141784668, + 0.004881848115473986, + 0.007120621390640736, + 0.0034615518525242805, + 0.008096271194517612, + 0.005945598240941763, + 0.007091045845299959, + 0.007011090405285358, + 0.0036771974992007017, + 0.004705321509391069, + 0.004539918154478073, + 0.03136104717850685, + 0.031661003828048706, + 0.09775424003601074, + 0.01740008220076561, + 0.11638695746660233, + 0.012290416285395622, + 0.029713492840528488, + 0.022820113226771355, + 0.004463274497538805, + 0.009221839718520641, + 0.0061494638212025166, + 0.006429850123822689, + 0.004619416315108538, + 0.0034235622733831406, + 0.003629701677709818, + 0.004280525725334883, + 0.006034203339368105, + 0.004546827170997858, + 0.012032145634293556, + 0.006545378360897303, + 0.016989601776003838, + 0.0033969220239669085, + 0.0058876215480268, + 0.0061478521674871445, + 0.0018648244440555573, + 0.003239609766751528, + 0.0031386162154376507, + 0.004120420664548874, + 0.0013799667358398438, + 0.0027298531495034695, + 0.0018959029112011194, + 0.00272565265186131, + 0.007704455871134996, + 0.013568874448537827, + 0.010850364342331886, + 0.012926110997796059, + 0.006248826161026955, + 0.008153896778821945, + 0.007745109964162111, + 0.007331354543566704, + 0.007931992411613464, + 0.006409962195903063, + 0.007255755364894867, + 0.006085420958697796, + 0.007921603508293629, + 0.004968755412846804, + 0.006389904767274857, + 0.007144213188439608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0263565294444561 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419412437826395 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.01419412437826395, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419412437826395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08391554653644562, + 0.1477096527814865, + 0.04199348762631416, + 0.02905518375337124, + 0.2106499820947647, + 0.04850959777832031, + 0.10018687695264816, + 0.06370575726032257, + 0.06328931450843811, + 0.21098457276821136 + ], + "credit_map": [ + 0.00777652021497488, + 0.012994968332350254, + 0.012854577973484993, + 0.009550516493618488, + 0.014459090307354927, + 0.009163782000541687, + 0.007523372769355774, + 0.009592719376087189, + 0.010317876003682613, + 0.011199412867426872, + 0.03598592057824135, + 0.010551695711910725, + 0.04493481293320656, + 0.009848610498011112, + 0.013958349823951721, + 0.010912972502410412, + 0.0058872513473033905, + 0.005643496289849281, + 0.005431529600173235, + 0.007023466285318136, + 0.003683769376948476, + 0.004835300147533417, + 0.004228333942592144, + 0.005260339472442865, + 0.0027223024517297745, + 0.0037480955943465233, + 0.004619950894266367, + 0.0037923199124634266, + 0.0040182191878557205, + 0.002704466925933957, + 0.0037161782383918762, + 0.0037336477544158697, + 0.018049456179142, + 0.017721738666296005, + 0.05764441564679146, + 0.01101589947938919, + 0.06731344014406204, + 0.007298077456653118, + 0.017220988869667053, + 0.014385964721441269, + 0.004782881122082472, + 0.004353160038590431, + 0.006886005401611328, + 0.004952206276357174, + 0.008575394749641418, + 0.004363590385764837, + 0.0096730412915349, + 0.004923318512737751, + 0.0076439836993813515, + 0.007393297739326954, + 0.01896556094288826, + 0.010746777057647705, + 0.029476577416062355, + 0.007304229773581028, + 0.011952289380133152, + 0.00670416047796607, + 0.00589886587113142, + 0.007599035743623972, + 0.008646217174828053, + 0.01224596332758665, + 0.0038944275584071875, + 0.005615489091724157, + 0.011030437424778938, + 0.008775322698056698, + 0.005756394471973181, + 0.012085827998816967, + 0.0087236063554883, + 0.009718744084239006, + 0.005004563368856907, + 0.006907585076987743, + 0.009344808757305145, + 0.005747782066464424, + 0.03174704313278198, + 0.027836034074425697, + 0.025879954919219017, + 0.02645365335047245, + 0.02850246988236904, + 0.02453642711043358, + 0.021064821630716324, + 0.02496415190398693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.01419412437826395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033697523176670074 + ], + "opposite_logprob_tokens": [ + -3.4086976051330566 + ], + "logprob_reward": -0.033697523176670074, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033697523176670074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2011663317680359, + 0.12303084135055542, + 0.062149904668331146, + 0.04710366576910019, + 0.10360236465930939, + 0.03237530216574669, + 0.07482177019119263, + 0.05630099028348923, + 0.08566807955503464, + 0.21378077566623688 + ], + "credit_map": [ + 0.009774256497621536, + 0.017074016854166985, + 0.019381942227482796, + 0.022213825955986977, + 0.06678285449743271, + 0.01686001382768154, + 0.030029572546482086, + 0.01904985122382641, + 0.010253671556711197, + 0.010816671885550022, + 0.026369791477918625, + 0.012751839123666286, + 0.032412830740213394, + 0.006824888288974762, + 0.014680545777082443, + 0.008920598775148392, + 0.007228298112750053, + 0.007449242286384106, + 0.011293800547719002, + 0.007548091467469931, + 0.008583982475101948, + 0.007289382629096508, + 0.006404458545148373, + 0.006352648604661226, + 0.0045248777605593204, + 0.005732065998017788, + 0.0067087276838719845, + 0.006318706087768078, + 0.0072664860635995865, + 0.004316126927733421, + 0.006241464521735907, + 0.005995211657136679, + 0.00864272378385067, + 0.008385411463677883, + 0.027060789987444878, + 0.00791979394853115, + 0.031550224870443344, + 0.003702521324157715, + 0.008328847587108612, + 0.008012047037482262, + 0.003657504916191101, + 0.004158730152994394, + 0.004528788384050131, + 0.0044166226871311665, + 0.005206394009292126, + 0.002401414792984724, + 0.004183411132544279, + 0.0038224358577281237, + 0.006907894741743803, + 0.0035804403014481068, + 0.013921346515417099, + 0.007934542372822762, + 0.020053664222359657, + 0.00583313824608922, + 0.010605483315885067, + 0.005985260475426912, + 0.00536209624260664, + 0.006187681574374437, + 0.007855030708014965, + 0.009250640869140625, + 0.004887204151600599, + 0.0072842747904360294, + 0.0072553944773972034, + 0.008218668401241302, + 0.008586934767663479, + 0.011181383393704891, + 0.014183994382619858, + 0.009978922083973885, + 0.007163546048104763, + 0.009752357378602028, + 0.013567598536610603, + 0.011253339238464832, + 0.042164791375398636, + 0.027849316596984863, + 0.03598487004637718, + 0.025690438225865364, + 0.018050840124487877, + 0.02136235497891903, + 0.020187698304653168, + 0.022490447387099266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.033697523176670074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 fantastical Art Nouveau-inspired metal forest with twisted bronze and brass trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-haired swordsman visible in a dark green leather armor holding two silver scimitars in reverse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman perform an explosive attacking stance where his arms slash out in two swift silver arcs to bisect a massive brass pistil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass pistil shown being neatly bisected with a burst of dazzling sparks and then collapsing to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience a sharp and brief tremor as the upper half of the brass pistil collapses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291743814945221 + ], + "opposite_logprob_tokens": [ + -2.237917423248291 + ], + "logprob_reward": -0.11291743814945221, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291743814945221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719481348991394, + 0.15431588888168335, + 0.04477572441101074, + 0.028894776478409767, + 0.07091782242059708, + 0.034517962485551834, + 0.048850975930690765, + 0.023195559158921242, + 0.08552677929401398, + 0.3370562791824341 + ], + "credit_map": [ + 0.00850832648575306, + 0.022346310317516327, + 0.011477145366370678, + 0.027326276525855064, + 0.05913274362683296, + 0.010419124737381935, + 0.01499706506729126, + 0.017741158604621887, + 0.008643753826618195, + 0.01021803542971611, + 0.0409393273293972, + 0.011164603754878044, + 0.050792429596185684, + 0.008205721154808998, + 0.013089598156511784, + 0.011262406595051289, + 0.005151964724063873, + 0.006035605911165476, + 0.0074560027569532394, + 0.0058945356868207455, + 0.005425355397164822, + 0.005102077033370733, + 0.004749536979943514, + 0.004960644990205765, + 0.0027125729247927666, + 0.004305995535105467, + 0.0036123974714428186, + 0.004777869209647179, + 0.0026028521824628115, + 0.0029385071247816086, + 0.004045714624226093, + 0.0038988690357655287, + 0.005442561116069555, + 0.005038069095462561, + 0.016819993034005165, + 0.006635854486376047, + 0.021878957748413086, + 0.002996945520862937, + 0.006474338937550783, + 0.005631101317703724, + 0.004531797952950001, + 0.005384373478591442, + 0.005395673681050539, + 0.004058674443513155, + 0.0031081875786185265, + 0.0029289855156093836, + 0.0048209442757070065, + 0.004289327189326286, + 0.005458609201014042, + 0.00524057075381279, + 0.0072543383575975895, + 0.006851116195321083, + 0.003914273343980312, + 0.005067650694400072, + 0.008426252752542496, + 0.006638168357312679, + 0.0022527463734149933, + 0.0031074851285666227, + 0.003643467789515853, + 0.0036388924345374107, + 0.0018409070326015353, + 0.0024122355971485376, + 0.003271290101110935, + 0.0030285362154245377, + 0.008113023824989796, + 0.012119442224502563, + 0.01077859103679657, + 0.012586496770381927, + 0.006437488365918398, + 0.0118996137753129, + 0.011842714622616768, + 0.01174941286444664, + 0.054206013679504395, + 0.034880008548498154, + 0.043393127620220184, + 0.030771950259804726, + 0.045624200254678726, + 0.04024944826960564, + 0.051211848855018616, + 0.03671968728303909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.11291743814945221 + } + ], + "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/0000052/group_5/index_tensor([8292, 8292])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_5/index_tensor([8292, 8292])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..876ea733eb4055518574774d44a100f38df1ad31 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_5/index_tensor([8292, 8292])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene opens with a high-angle, wide shot, the camera slowly tilting downward to follow a young, happy Asian child as he expertly rides a BMX bike through a park. The child, about seven years old, has an expression of pure, focused joy, thick, dark curls peeking out from under a vibrant orange beanie. They're dressed in a layered, comfortable style—a slightly oversized gray zip-up hoodie paired with a plain white T-shirt and dark, durable jeans. Their small frame rests on a compact, matte black BMX bike, which they expertly maneuver. By deliberately shifting their weight, the child leans back, arms extended, and lifts the front wheel off the ground in a slow, steady manual effort. The bike glides gracefully and steadily along the smooth concrete path. The entire park is bathed in the magical glow of dusk's golden hour. A low, warm sun slants through a thin, ethereal layer of mist, casting long, soft shadows that make the air itself seem to glow. The background reveals the soft outlines of tall trees and a park bench in the distance, all softened by a light mist, creating a serene, dreamy and deeply peaceful atmosphere that celebrates a moment of youthful freedom and skill.", + "seed": 2433, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_5/index_8292_seed_2433_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": -0.013329861452803016, + "vlm_logprob": -0.013329861452803016, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.013329861452803016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028596071526408195 + ], + "opposite_logprob_tokens": [ + -4.403595924377441 + ], + "logprob_reward": -0.028596071526408195, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.028596071526408195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07920275628566742, + 0.024020858108997345, + 0.034095145761966705, + 0.10092286765575409, + 0.05270316079258919, + 0.11659020185470581, + 0.42489299178123474, + 0.10288350284099579, + 0.0292664747685194, + 0.03542197868227959 + ], + "credit_map": [ + 0.0072805071249604225, + 0.008247829042375088, + 0.014900743030011654, + 0.007563839666545391, + 0.007859738543629646, + 0.011296221055090427, + 0.009921460412442684, + 0.012132416479289532, + 0.0032474861945956945, + 0.0029629936907440424, + 0.002673171693459153, + 0.0018536229617893696, + 0.0019570181611925364, + 0.006325359921902418, + 0.0020473902113735676, + 0.0029538150411099195, + 0.003491633338853717, + 0.0033626470249146223, + 0.004711728543043137, + 0.0022019187454134226, + 0.002910274313762784, + 0.0065633454360067844, + 0.002232972765341401, + 0.008620626293122768, + 0.010197935625910759, + 0.023348623886704445, + 0.013037131167948246, + 0.015630347654223442, + 0.012959307059645653, + 0.011507604271173477, + 0.006360930856317282, + 0.00788098108023405, + 0.0058903321623802185, + 0.005561257246881723, + 0.006441797129809856, + 0.006173725239932537, + 0.005486360751092434, + 0.007458813022822142, + 0.0050994944758713245, + 0.010591386817395687, + 0.0110046761110425, + 0.009116608649492264, + 0.0061693559400737286, + 0.011838246136903763, + 0.007364187389612198, + 0.00692999130114913, + 0.03300255164504051, + 0.031164582818746567, + 0.026215750724077225, + 0.023054620251059532, + 0.060670673847198486, + 0.028828157112002373, + 0.04086819663643837, + 0.08867403119802475, + 0.026429960504174232, + 0.13015161454677582, + 0.015069014392793179, + 0.011221502907574177, + 0.010926592163741589, + 0.01195928268134594, + 0.016089124605059624, + 0.01354958675801754, + 0.01144902128726244, + 0.012619378976523876, + 0.004446959588676691, + 0.004649466369301081, + 0.002825614297762513, + 0.0040953438729047775, + 0.003338458016514778, + 0.0033562479075044394, + 0.0038336790166795254, + 0.0027207036036998034, + 0.0074726007878780365, + 0.004617436323314905, + 0.004306190647184849, + 0.0040837121196091175, + 0.005356151610612869, + 0.002754915738478303, + 0.002975048730149865, + 0.003855922492220998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.028596071526408195 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023310085758566856 + ], + "opposite_logprob_tokens": [ + -3.7733101844787598 + ], + "logprob_reward": -0.023310085758566856, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023310085758566856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1953568458557129, + 0.03301398456096649, + 0.08553571254014969, + 0.17300041019916534, + 0.0854889377951622, + 0.052571602165699005, + 0.11424719542264938, + 0.1335037499666214, + 0.06336154788732529, + 0.06391998380422592 + ], + "credit_map": [ + 0.016217507421970367, + 0.0351480133831501, + 0.020819423720240593, + 0.02197544276714325, + 0.02581963874399662, + 0.03361578285694122, + 0.021209152415394783, + 0.020551880821585655, + 0.005216669291257858, + 0.003901992691680789, + 0.004358798265457153, + 0.003187852678820491, + 0.0026322423946112394, + 0.005568646360188723, + 0.0030480443965643644, + 0.0050997380167245865, + 0.009789966978132725, + 0.016197212040424347, + 0.009068984538316727, + 0.008335145190358162, + 0.008404849097132683, + 0.01252697128802538, + 0.005319320596754551, + 0.01589326560497284, + 0.0158537607640028, + 0.04078112542629242, + 0.021208299323916435, + 0.02857944369316101, + 0.025059884414076805, + 0.019204603508114815, + 0.011035144329071045, + 0.011278126388788223, + 0.011400408111512661, + 0.011530821211636066, + 0.011639369651675224, + 0.01091862004250288, + 0.009987439960241318, + 0.014728876762092113, + 0.0072416625916957855, + 0.008041744120419025, + 0.0048377132043242455, + 0.007981859147548676, + 0.005517724435776472, + 0.007569756358861923, + 0.006176593713462353, + 0.008195911534130573, + 0.006455489434301853, + 0.0058365557342767715, + 0.010191911831498146, + 0.0091154919937253, + 0.008913944475352764, + 0.008342361077666283, + 0.01276738103479147, + 0.03067377209663391, + 0.008607247844338417, + 0.02563508413732052, + 0.033202484250068665, + 0.013689840212464333, + 0.01619759015738964, + 0.010034164413809776, + 0.013650608249008656, + 0.015270417556166649, + 0.01420123502612114, + 0.017257407307624817, + 0.011135758832097054, + 0.008996774442493916, + 0.007144239731132984, + 0.007132669445127249, + 0.00784117728471756, + 0.006958238314837217, + 0.008302567526698112, + 0.005850121844559908, + 0.009787231683731079, + 0.011311261914670467, + 0.009148847311735153, + 0.009165617637336254, + 0.005399610847234726, + 0.005812084302306175, + 0.005901953671127558, + 0.0073933787643909454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.023310085758566856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636489786207676 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.004636489786207676, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636489786207676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07766486704349518, + 0.03603226691484451, + 0.09240379929542542, + 0.03796878829598427, + 0.08139294385910034, + 0.10882972180843353, + 0.35917606949806213, + 0.09100863337516785, + 0.04567193612456322, + 0.06985089182853699 + ], + "credit_map": [ + 0.008422324433922768, + 0.009316856972873211, + 0.010379482991993427, + 0.006095650140196085, + 0.008543199859559536, + 0.01609746553003788, + 0.007604712154716253, + 0.011205167509615421, + 0.006186683662235737, + 0.003579620970413089, + 0.004509258083999157, + 0.0033383197151124477, + 0.002821614034473896, + 0.006624509114772081, + 0.0030714955646544695, + 0.00590076670050621, + 0.012770525179803371, + 0.009651396423578262, + 0.012771451845765114, + 0.0050246682949364185, + 0.007582854945212603, + 0.01795083098113537, + 0.004898281767964363, + 0.021753789857029915, + 0.003919945564121008, + 0.005453246645629406, + 0.004907007794827223, + 0.00580270541831851, + 0.004733829293400049, + 0.005391652230173349, + 0.0029223014134913683, + 0.004838102497160435, + 0.009487349539995193, + 0.009414652362465858, + 0.01156165823340416, + 0.010764027014374733, + 0.00877328310161829, + 0.015801476314663887, + 0.005966589320451021, + 0.009623910300433636, + 0.007424912415444851, + 0.010009855031967163, + 0.006939292419701815, + 0.013791266828775406, + 0.007584135048091412, + 0.007390410639345646, + 0.027874942868947983, + 0.02781490422785282, + 0.02641664259135723, + 0.027891011908650398, + 0.03718114271759987, + 0.037692826241254807, + 0.0344458632171154, + 0.07384693622589111, + 0.02917221002280712, + 0.09252945333719254, + 0.010200947523117065, + 0.01269339770078659, + 0.009729627519845963, + 0.012731168419122696, + 0.012347511015832424, + 0.01178639568388462, + 0.011024107225239277, + 0.010495482012629509, + 0.006823928561061621, + 0.007922818884253502, + 0.004570715129375458, + 0.005575814284384251, + 0.0062556518241763115, + 0.00662663159891963, + 0.004050662741065025, + 0.003845711937174201, + 0.010406418703496456, + 0.011335535906255245, + 0.008809465914964676, + 0.006930042523890734, + 0.008966708555817604, + 0.007001216057687998, + 0.008381986059248447, + 0.008019519038498402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.004636489786207676 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002484927186742425 + ], + "opposite_logprob_tokens": [ + -6.0024847984313965 + ], + "logprob_reward": -0.002484927186742425, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002484927186742425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10920733958482742, + 0.032194625586271286, + 0.06505273282527924, + 0.047769945114851, + 0.17120331525802612, + 0.06402428448200226, + 0.36250001192092896, + 0.07217838615179062, + 0.02820272371172905, + 0.047666605561971664 + ], + "credit_map": [ + 0.01359873078763485, + 0.009314846247434616, + 0.016803232952952385, + 0.00803627260029316, + 0.009944016113877296, + 0.023950738832354546, + 0.008935733698308468, + 0.01862376369535923, + 0.0038153838831931353, + 0.0027963374741375446, + 0.004159172531217337, + 0.0021139681339263916, + 0.002542549977079034, + 0.007920238189399242, + 0.0030777244828641415, + 0.005769250448793173, + 0.009605309925973415, + 0.006572986952960491, + 0.009157134220004082, + 0.003766456386074424, + 0.005291950423270464, + 0.013375895097851753, + 0.0036901840940117836, + 0.013592814095318317, + 0.0047578648664057255, + 0.009374679066240788, + 0.00698061054572463, + 0.006293070502579212, + 0.006078926380723715, + 0.00652567483484745, + 0.0033422033302485943, + 0.004416915122419596, + 0.019073793664574623, + 0.009973296895623207, + 0.023714056238532066, + 0.012528685852885246, + 0.020613407716155052, + 0.032520294189453125, + 0.019480913877487183, + 0.03329887613654137, + 0.009703263640403748, + 0.005677226930856705, + 0.005406954325735569, + 0.003246998181566596, + 0.006874073296785355, + 0.013663490302860737, + 0.009876708500087261, + 0.009575567208230495, + 0.027580367401242256, + 0.017191460356116295, + 0.057204511016607285, + 0.0185511726886034, + 0.030542301014065742, + 0.07801380753517151, + 0.01896779239177704, + 0.11444859951734543, + 0.012667113915085793, + 0.00844957958906889, + 0.0077523174695670605, + 0.007524332497268915, + 0.01147459726780653, + 0.00924557913094759, + 0.007473777513951063, + 0.007591087371110916, + 0.004648106172680855, + 0.005465133581310511, + 0.003036479465663433, + 0.003547284984961152, + 0.002130423206835985, + 0.004622945096343756, + 0.00262980698607862, + 0.002122544450685382, + 0.00890260934829712, + 0.009653451852500439, + 0.005149479955434799, + 0.004896274767816067, + 0.00901130959391594, + 0.003973227459937334, + 0.003216566750779748, + 0.0028636865317821503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.002484927186742425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621733006089926 + ], + "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": [ + 0.11695742607116699, + 0.04925970733165741, + 0.07425779849290848, + 0.13628169894218445, + 0.10756684839725494, + 0.18373282253742218, + 0.23218496143817902, + 0.030012940987944603, + 0.02186005748808384, + 0.047885749489068985 + ], + "credit_map": [ + 0.014017283916473389, + 0.014310220256447792, + 0.012809967622160912, + 0.0075557115487754345, + 0.0140009056776762, + 0.022930452600121498, + 0.013868891634047031, + 0.017463993281126022, + 0.004549933597445488, + 0.006005838047713041, + 0.006828465033322573, + 0.004932719748467207, + 0.004517682828009129, + 0.007047880440950394, + 0.006723380647599697, + 0.008653809316456318, + 0.006971554830670357, + 0.005423642694950104, + 0.009484577924013138, + 0.007177925668656826, + 0.007386793382465839, + 0.013559757731854916, + 0.0056504905223846436, + 0.018603051081299782, + 0.019388273358345032, + 0.020222224295139313, + 0.017735300585627556, + 0.01643737405538559, + 0.015943046659231186, + 0.02130816876888275, + 0.009197991341352463, + 0.01604931801557541, + 0.010294366627931595, + 0.006342117674648762, + 0.0088322339579463, + 0.004742405377328396, + 0.018582697957754135, + 0.022095004096627235, + 0.013187042437493801, + 0.02349098213016987, + 0.023240135982632637, + 0.008797417394816875, + 0.010594301857054234, + 0.006921081803739071, + 0.014230413362383842, + 0.018880488350987434, + 0.05336838588118553, + 0.04770060256123543, + 0.013188734650611877, + 0.012514874339103699, + 0.02837951108813286, + 0.01948520541191101, + 0.02043040655553341, + 0.05225444585084915, + 0.017478862777352333, + 0.06845290213823318, + 0.0034512921702116728, + 0.0031810076907277107, + 0.0029180292040109634, + 0.00444328598678112, + 0.004433623980730772, + 0.0038365796208381653, + 0.0043040853925049305, + 0.003445038106292486, + 0.004342324566096067, + 0.0035106074064970016, + 0.0029452252201735973, + 0.0024412900675088167, + 0.0020023242104798555, + 0.0031307260505855083, + 0.0019318668637424707, + 0.001555693568661809, + 0.006765462923794985, + 0.010583463124930859, + 0.008419797755777836, + 0.006831725127995014, + 0.005430654156953096, + 0.00396755849942565, + 0.0031488933600485325, + 0.0027381947729736567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.007621733006089926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene opens with a high-angle, wide shot, the camera slowly tilting downward to follow a young, happy Asian child as he expertly rides a BMX bike through a park. The child, about seven years old, has an expression of pure, focused joy, thick, dark curls peeking out from under a vibrant orange beanie. They're dressed in a layered, comfortable style—a slightly oversized gray zip-up hoodie paired with a plain white T-shirt and dark, durable jeans. Their small frame rests on a compact, matte black BMX bike, which they expertly maneuver. By deliberately shifting their weight, the child leans back, arms extended, and lifts the front wheel off the ground in a slow, steady manual effort. The bike glides gracefully and steadily along the smooth concrete path. The entire park is bathed in the magical glow of dusk's golden hour. A low, warm sun slants through a thin, ethereal layer of mist, casting long, soft shadows that make the air itself seem to glow. The background reveals the soft outlines of tall trees and a park bench in the distance, all softened by a light mist, creating a serene, dreamy and deeply peaceful atmosphere that celebrates a moment of youthful freedom and skill.", + "seed": 2434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_5/index_8292_seed_2434_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": -0.05926054334267974, + "vlm_logprob": -0.05926054334267974, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.05926054334267974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24420104920864105 + ], + "opposite_logprob_tokens": [ + -1.6192010641098022 + ], + "logprob_reward": -0.24420104920864105, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24420104920864105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08674361556768417, + 0.049104444682598114, + 0.07038604468107224, + 0.07952270656824112, + 0.04570595920085907, + 0.08996117115020752, + 0.10033684968948364, + 0.09424346685409546, + 0.18908166885375977, + 0.1949141025543213 + ], + "credit_map": [ + 0.011580000631511211, + 0.011651510372757912, + 0.014096113853156567, + 0.01165924221277237, + 0.014515726827085018, + 0.009411381557583809, + 0.006974338553845882, + 0.00685530249029398, + 0.010312872938811779, + 0.0076241446658968925, + 0.006475492380559444, + 0.007331737317144871, + 0.005210438277572393, + 0.004507791716605425, + 0.0038361591286957264, + 0.003805809887126088, + 0.020825201645493507, + 0.008505754172801971, + 0.009028462693095207, + 0.006646040827035904, + 0.008819547481834888, + 0.006213012151420116, + 0.005503131542354822, + 0.0048448932357132435, + 0.02475390024483204, + 0.013737591914832592, + 0.011809607036411762, + 0.008091412484645844, + 0.006826227996498346, + 0.00463430630043149, + 0.0049470216035842896, + 0.004722632933408022, + 0.009074208326637745, + 0.008946276269853115, + 0.005944916047155857, + 0.00596956443041563, + 0.005729700904339552, + 0.0038799571339040995, + 0.003233524039387703, + 0.0029278136789798737, + 0.020720867440104485, + 0.012445119209587574, + 0.012052210979163647, + 0.011252843774855137, + 0.009131825529038906, + 0.00886670220643282, + 0.008033950813114643, + 0.007457645609974861, + 0.030110592022538185, + 0.0133949164301157, + 0.013353888876736164, + 0.012502714060246944, + 0.007280745077878237, + 0.009642657823860645, + 0.006866934709250927, + 0.007184400223195553, + 0.011256099678575993, + 0.009406760334968567, + 0.011107227765023708, + 0.014346788637340069, + 0.015787003561854362, + 0.006277775391936302, + 0.012205000035464764, + 0.013856813311576843, + 0.01910414919257164, + 0.03281549736857414, + 0.023448845371603966, + 0.018606221303343773, + 0.02711375616490841, + 0.015521635301411152, + 0.0279062669724226, + 0.024565314874053, + 0.013459860347211361, + 0.01174922939389944, + 0.05766666308045387, + 0.009957650676369667, + 0.010931110009551048, + 0.013827113434672356, + 0.013089965097606182, + 0.06423251330852509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.24420104920864105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023310551419854164 + ], + "opposite_logprob_tokens": [ + -3.773310661315918 + ], + "logprob_reward": -0.023310551419854164, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023310551419854164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14072465896606445, + 0.07312889397144318, + 0.11857295036315918, + 0.06739600747823715, + 0.05938202142715454, + 0.047094471752643585, + 0.10784924030303955, + 0.07559305429458618, + 0.06903725862503052, + 0.24122148752212524 + ], + "credit_map": [ + 0.01432157214730978, + 0.013383260928094387, + 0.01908966898918152, + 0.01957196742296219, + 0.021031297743320465, + 0.02838834747672081, + 0.012255006469786167, + 0.012683536857366562, + 0.009236310608685017, + 0.005293260328471661, + 0.008491307497024536, + 0.007281860802322626, + 0.01295672357082367, + 0.012262659147381783, + 0.00867542251944542, + 0.008931350894272327, + 0.03554832190275192, + 0.013186279684305191, + 0.014120679348707199, + 0.010740049183368683, + 0.01182437501847744, + 0.017803028225898743, + 0.00965309888124466, + 0.005697120446711779, + 0.01676981896162033, + 0.010045409202575684, + 0.010033359751105309, + 0.006389377638697624, + 0.00843240786343813, + 0.006853764411062002, + 0.005378293339163065, + 0.0034935707226395607, + 0.0070993029512465, + 0.008153469301760197, + 0.007585932034999132, + 0.007613217458128929, + 0.009899439290165901, + 0.009030384011566639, + 0.004936160519719124, + 0.005064112599939108, + 0.007064166944473982, + 0.00896535161882639, + 0.0061454796232283115, + 0.005244701635092497, + 0.0042827134020626545, + 0.004884731024503708, + 0.005414139479398727, + 0.005093186628073454, + 0.017101481556892395, + 0.013704742304980755, + 0.013676244765520096, + 0.012168807908892632, + 0.013166561722755432, + 0.017682382836937904, + 0.007991655729711056, + 0.012357356958091259, + 0.008137066848576069, + 0.008070669136941433, + 0.007787705399096012, + 0.008649894036352634, + 0.02043825574219227, + 0.006206877063959837, + 0.009636923670768738, + 0.006665657740086317, + 0.007193966303020716, + 0.012644407339394093, + 0.008943788707256317, + 0.00601010350510478, + 0.009812694042921066, + 0.00982561893761158, + 0.006085191387683153, + 0.008521484211087227, + 0.021152779459953308, + 0.018823176622390747, + 0.053774695843458176, + 0.01544278021901846, + 0.026275038719177246, + 0.02202705480158329, + 0.014742220751941204, + 0.06898371875286102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.023310551419854164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636031299829483 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.008636031299829483, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636031299829483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1632651388645172, + 0.07378005236387253, + 0.06157976761460304, + 0.047193486243486404, + 0.08658159524202347, + 0.07690080255270004, + 0.11738819628953934, + 0.10955431312322617, + 0.06607857346534729, + 0.1976781189441681 + ], + "credit_map": [ + 0.018434273079037666, + 0.019000962376594543, + 0.029171360656619072, + 0.019702652469277382, + 0.020869728177785873, + 0.026678143069148064, + 0.014597348868846893, + 0.014810657128691673, + 0.011749524623155594, + 0.008631664328277111, + 0.008382885716855526, + 0.006872933357954025, + 0.008593912236392498, + 0.010282576084136963, + 0.012085160240530968, + 0.007181387394666672, + 0.018000533804297447, + 0.007896779105067253, + 0.00670141214504838, + 0.005741646979004145, + 0.004901045933365822, + 0.0069135152734816074, + 0.007536881137639284, + 0.00388794532045722, + 0.009494172409176826, + 0.0071679516695439816, + 0.007732469588518143, + 0.00531263742595911, + 0.005907035898417234, + 0.0048634931445121765, + 0.004248157609254122, + 0.0024675659369677305, + 0.01559572946280241, + 0.014689695090055466, + 0.011867456138134003, + 0.011815576814115047, + 0.0070806099101901054, + 0.011489616706967354, + 0.005773328710347414, + 0.008269578218460083, + 0.012429836206138134, + 0.010699753649532795, + 0.0094003239646554, + 0.008389859460294247, + 0.008206688798964024, + 0.010768801905214787, + 0.008316733874380589, + 0.008688800036907196, + 0.018457569181919098, + 0.012413364835083485, + 0.014092491008341312, + 0.011551952920854092, + 0.011808904819190502, + 0.029054656624794006, + 0.008151199668645859, + 0.011858044192194939, + 0.010602153837680817, + 0.009620952419936657, + 0.013435861095786095, + 0.017160886898636818, + 0.018861902877688408, + 0.011306578293442726, + 0.013768812641501427, + 0.014797159470617771, + 0.006757495459169149, + 0.008367979899048805, + 0.009039127267897129, + 0.006425626575946808, + 0.008965895511209965, + 0.008209118619561195, + 0.009664271958172321, + 0.008649054914712906, + 0.014131370931863785, + 0.012623917311429977, + 0.054701972752809525, + 0.011894982308149338, + 0.015022292733192444, + 0.01333156879991293, + 0.013708219863474369, + 0.06226375699043274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.008636031299829483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533470056951046 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533470056951046, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533470056951046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2985825538635254, + 0.06579715758562088, + 0.21165597438812256, + 0.057170893996953964, + 0.11378898471593857, + 0.04486345499753952, + 0.029627973213791847, + 0.05629311874508858, + 0.061042796820402145, + 0.061177097260951996 + ], + "credit_map": [ + 0.04108024016022682, + 0.03496619686484337, + 0.05506886914372444, + 0.04871112480759621, + 0.056438151746988297, + 0.02171141281723976, + 0.024365540593862534, + 0.016241008415818214, + 0.013032419607043266, + 0.008916550315916538, + 0.007892998866736889, + 0.008007800206542015, + 0.006823008880019188, + 0.006495044566690922, + 0.008744974620640278, + 0.005884358659386635, + 0.06501991301774979, + 0.025302395224571228, + 0.02473217248916626, + 0.017391832545399666, + 0.012337572872638702, + 0.040919553488492966, + 0.01576191745698452, + 0.010190614499151707, + 0.010051668621599674, + 0.012736574746668339, + 0.009855852462351322, + 0.007402053568512201, + 0.005135233514010906, + 0.005776974838227034, + 0.003304430516436696, + 0.0029081094544380903, + 0.01714540831744671, + 0.02260500378906727, + 0.016674377024173737, + 0.015601211227476597, + 0.00938078947365284, + 0.015418099239468575, + 0.006809016223996878, + 0.010155086405575275, + 0.0053245266899466515, + 0.007094549015164375, + 0.006907198112457991, + 0.006371120922267437, + 0.003019751748070121, + 0.006244468968361616, + 0.004668849986046553, + 0.005232990253716707, + 0.004173909779638052, + 0.003173037664964795, + 0.003609646111726761, + 0.0027590100653469563, + 0.0026006167754530907, + 0.008017162792384624, + 0.001921775401569903, + 0.003372815204784274, + 0.006430055946111679, + 0.004718167241662741, + 0.0074276733212172985, + 0.009099144488573074, + 0.007387329358607531, + 0.005323540419340134, + 0.0068775503896176815, + 0.009029658511281013, + 0.005813822150230408, + 0.006668195594102144, + 0.008507455699145794, + 0.00536910118535161, + 0.008221952244639397, + 0.008983501233160496, + 0.009172524325549603, + 0.008306250907480717, + 0.004639880731701851, + 0.004886109847575426, + 0.012370827607810497, + 0.005255431402474642, + 0.006610859651118517, + 0.006878415588289499, + 0.004744344856590033, + 0.01579122431576252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.012533470056951046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762161472812295 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.00762161472812295, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762161472812295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33917367458343506, + 0.15221188962459564, + 0.023635899648070335, + 0.03917417302727699, + 0.045166149735450745, + 0.03496424853801727, + 0.05229031294584274, + 0.08593055605888367, + 0.13171443343162537, + 0.09573869407176971 + ], + "credit_map": [ + 0.053085725754499435, + 0.046542104333639145, + 0.0554213784635067, + 0.04630221799015999, + 0.05737766996026039, + 0.030611500144004822, + 0.025315728038549423, + 0.02451734058558941, + 0.031220583245158195, + 0.02676384709775448, + 0.0242491215467453, + 0.026268992573022842, + 0.012504132464528084, + 0.01208345964550972, + 0.011541416868567467, + 0.007580322679132223, + 0.003516684053465724, + 0.0024241022765636444, + 0.0029071818571537733, + 0.0026528772432357073, + 0.0034875806886702776, + 0.002966830041259527, + 0.0031454197596758604, + 0.0025352241937071085, + 0.0039145671762526035, + 0.008031442761421204, + 0.006181287579238415, + 0.006029273848980665, + 0.005507939960807562, + 0.004922389518469572, + 0.0022431546822190285, + 0.0023441174998879433, + 0.006448636297136545, + 0.00800260528922081, + 0.005948417820036411, + 0.006894259247928858, + 0.005097455810755491, + 0.005843102466315031, + 0.0029989085160195827, + 0.003932761959731579, + 0.003355534980073571, + 0.005419077351689339, + 0.005379890091717243, + 0.005921554286032915, + 0.0031638219952583313, + 0.004032248165458441, + 0.004075066652148962, + 0.003617053385823965, + 0.00881271529942751, + 0.005287675186991692, + 0.006810255814343691, + 0.005876593757420778, + 0.005448745563626289, + 0.010012323036789894, + 0.003996791318058968, + 0.006045213900506496, + 0.008818389847874641, + 0.00786303635686636, + 0.010093784891068935, + 0.015985574573278427, + 0.016187695786356926, + 0.006765890866518021, + 0.010415583848953247, + 0.009800598956644535, + 0.015475058928132057, + 0.015926728025078773, + 0.01858682744204998, + 0.013143942691385746, + 0.016651209443807602, + 0.01893165335059166, + 0.01331422571092844, + 0.019684791564941406, + 0.006677583325654268, + 0.007235032971948385, + 0.026960792019963264, + 0.005452641285955906, + 0.0061083585023880005, + 0.007314356509596109, + 0.007106046192348003, + 0.028883881866931915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.00762161472812295 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene opens with a high-angle, wide shot, the camera slowly tilting downward to follow a young, happy Asian child as he expertly rides a BMX bike through a park. The child, about seven years old, has an expression of pure, focused joy, thick, dark curls peeking out from under a vibrant orange beanie. They're dressed in a layered, comfortable style—a slightly oversized gray zip-up hoodie paired with a plain white T-shirt and dark, durable jeans. Their small frame rests on a compact, matte black BMX bike, which they expertly maneuver. By deliberately shifting their weight, the child leans back, arms extended, and lifts the front wheel off the ground in a slow, steady manual effort. The bike glides gracefully and steadily along the smooth concrete path. The entire park is bathed in the magical glow of dusk's golden hour. A low, warm sun slants through a thin, ethereal layer of mist, casting long, soft shadows that make the air itself seem to glow. The background reveals the soft outlines of tall trees and a park bench in the distance, all softened by a light mist, creating a serene, dreamy and deeply peaceful atmosphere that celebrates a moment of youthful freedom and skill.", + "seed": 2435, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_5/index_8292_seed_2435_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": -0.0207466883584857, + "vlm_logprob": -0.0207466883584857, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.0207466883584857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04046105965971947 + ], + "opposite_logprob_tokens": [ + -3.6654610633850098 + ], + "logprob_reward": -0.04046105965971947, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04046105965971947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3676507771015167, + 0.05183909088373184, + 0.06747029721736908, + 0.07376734912395477, + 0.07568572461605072, + 0.16363629698753357, + 0.0724877119064331, + 0.040136080235242844, + 0.027483897283673286, + 0.05984284356236458 + ], + "credit_map": [ + 0.02850554697215557, + 0.05071908235549927, + 0.06889539211988449, + 0.08523149788379669, + 0.027509668841958046, + 0.030027220025658607, + 0.026949865743517876, + 0.04981248453259468, + 0.004619533661752939, + 0.009716395288705826, + 0.006474450696259737, + 0.008434227667748928, + 0.007297401316463947, + 0.005956454202532768, + 0.003714790567755699, + 0.0056258318945765495, + 0.009659020230174065, + 0.010872979648411274, + 0.014324729330837727, + 0.013013888150453568, + 0.007125641219317913, + 0.0045812358148396015, + 0.003405617084354162, + 0.004487178288400173, + 0.008131925947964191, + 0.01418430544435978, + 0.012155645526945591, + 0.012476041913032532, + 0.009343879297375679, + 0.008112816140055656, + 0.004510639701038599, + 0.00485208909958601, + 0.010669860988855362, + 0.01475380826741457, + 0.00840186607092619, + 0.008710111491382122, + 0.010031596757471561, + 0.01087455078959465, + 0.005623535718768835, + 0.006620384752750397, + 0.023718662559986115, + 0.018577653914690018, + 0.020508544519543648, + 0.023734990507364273, + 0.016697118058800697, + 0.015107105486094952, + 0.020597489550709724, + 0.02469470538198948, + 0.014280669391155243, + 0.00827392004430294, + 0.010487320832908154, + 0.011637957766652107, + 0.0057942308485507965, + 0.007554424926638603, + 0.0058431243523955345, + 0.008616059087216854, + 0.007437253836542368, + 0.004779218230396509, + 0.007594745140522718, + 0.005766650661826134, + 0.0020845686085522175, + 0.004753059707581997, + 0.0029304001946002245, + 0.004790178034454584, + 0.0036405008286237717, + 0.0024270436260849237, + 0.0038917530328035355, + 0.004526556469500065, + 0.002915631979703903, + 0.0034134737215936184, + 0.003124300157651305, + 0.003544633975252509, + 0.008173391222953796, + 0.010470953769981861, + 0.009850346483290195, + 0.008006677962839603, + 0.007267187349498272, + 0.0039469473995268345, + 0.005588439758867025, + 0.006538889836519957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.04046105965971947 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982505038380623 + ], + "opposite_logprob_tokens": [ + -3.52982497215271 + ], + "logprob_reward": -0.02982505038380623, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982505038380623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2854872941970825, + 0.09730302542448044, + 0.1664123833179474, + 0.0742768868803978, + 0.09503301233053207, + 0.07593086361885071, + 0.04696548730134964, + 0.032876454293727875, + 0.07470452040433884, + 0.05101010203361511 + ], + "credit_map": [ + 0.02141631208360195, + 0.023665959015488625, + 0.08027590811252594, + 0.030033618211746216, + 0.028540991246700287, + 0.016701973974704742, + 0.01962321810424328, + 0.06522931158542633, + 0.010008696466684341, + 0.01265677809715271, + 0.012352917343378067, + 0.01328431535512209, + 0.015298445709049702, + 0.01412767730653286, + 0.01005074568092823, + 0.009523453190922737, + 0.013119976967573166, + 0.01873171143233776, + 0.01736297458410263, + 0.023308243602514267, + 0.035378601402044296, + 0.018537895753979683, + 0.021421421319246292, + 0.01855156198143959, + 0.006217439193278551, + 0.00932676624506712, + 0.015769699588418007, + 0.013078397139906883, + 0.011096592992544174, + 0.007443058770149946, + 0.0057817003689706326, + 0.005563234444707632, + 0.012159858830273151, + 0.014579687267541885, + 0.010582858696579933, + 0.009258441627025604, + 0.015950683504343033, + 0.011531015858054161, + 0.011923343874514103, + 0.009047120809555054, + 0.00975048914551735, + 0.009002100676298141, + 0.008967220783233643, + 0.01147829182446003, + 0.00922170840203762, + 0.007409993093460798, + 0.010778882540762424, + 0.00932217389345169, + 0.008752262219786644, + 0.005683435592800379, + 0.007302913814783096, + 0.008050491102039814, + 0.003970660734921694, + 0.004519999958574772, + 0.004046655725687742, + 0.004639065358787775, + 0.00581973185762763, + 0.0032613619696348906, + 0.004852123558521271, + 0.004493396263569593, + 0.003918256610631943, + 0.003908498678356409, + 0.0034690832253545523, + 0.003154000034555793, + 0.008147486485540867, + 0.005470044445246458, + 0.01005154475569725, + 0.008888106793165207, + 0.007448486518114805, + 0.008677133359014988, + 0.010594918392598629, + 0.015426797792315483, + 0.0071990275755524635, + 0.006779419258236885, + 0.007762366905808449, + 0.005428314208984375, + 0.0073612648993730545, + 0.003582792356610298, + 0.004608848597854376, + 0.008288068696856499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.02982505038380623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25308123230934143, + 0.03773503005504608, + 0.11970280110836029, + 0.07256580889225006, + 0.09525371342897415, + 0.09675975888967514, + 0.10037215799093246, + 0.10746076703071594, + 0.07287459075450897, + 0.04419417679309845 + ], + "credit_map": [ + 0.025045767426490784, + 0.05179017782211304, + 0.04080068692564964, + 0.03301233425736427, + 0.019587911665439606, + 0.03077668882906437, + 0.01849454641342163, + 0.03357309103012085, + 0.0034693556372076273, + 0.0053995694033801556, + 0.005544651765376329, + 0.006239113863557577, + 0.005419985856860876, + 0.004091951530426741, + 0.003990091383457184, + 0.003580305492505431, + 0.015741530805826187, + 0.022217068821191788, + 0.02067536860704422, + 0.024108944460749626, + 0.011977470479905605, + 0.009502718225121498, + 0.00948749203234911, + 0.005992196965962648, + 0.006906550377607346, + 0.011892245151102543, + 0.013575367629528046, + 0.013063034042716026, + 0.010387305170297623, + 0.007378675509244204, + 0.0049837506376206875, + 0.004378870595246553, + 0.01721644401550293, + 0.01611769013106823, + 0.009607694111764431, + 0.01047916617244482, + 0.014085073955357075, + 0.011509021744132042, + 0.007840482518076897, + 0.008398129604756832, + 0.01067786943167448, + 0.012267176061868668, + 0.012382506392896175, + 0.017650101333856583, + 0.010034749284386635, + 0.009425465948879719, + 0.011792493984103203, + 0.012529389932751656, + 0.014343616552650928, + 0.011999071575701237, + 0.01384945958852768, + 0.010182661935687065, + 0.013226285576820374, + 0.009251614101231098, + 0.012160533107817173, + 0.015358904376626015, + 0.016535747796297073, + 0.010534631088376045, + 0.01845657080411911, + 0.014315227046608925, + 0.010435480624437332, + 0.014319658279418945, + 0.009067874401807785, + 0.013795576058328152, + 0.010661888867616653, + 0.007581161800771952, + 0.010709092952311039, + 0.008000222966074944, + 0.005671292543411255, + 0.007397030014544725, + 0.01032013539224863, + 0.012533755972981453, + 0.006786794401705265, + 0.006845055613666773, + 0.00672889081761241, + 0.005074738524854183, + 0.006373761221766472, + 0.003451109165325761, + 0.004220741335302591, + 0.00471308222040534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0052520157769322395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624335587024689 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.007624335587024689, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624335587024689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3356298506259918, + 0.024041082710027695, + 0.2012910395860672, + 0.0572713240981102, + 0.12050044536590576, + 0.1053805872797966, + 0.03869485855102539, + 0.033465102314949036, + 0.024160273373126984, + 0.059565383940935135 + ], + "credit_map": [ + 0.026028845459222794, + 0.06311790645122528, + 0.04421169310808182, + 0.07108171284198761, + 0.03462257608771324, + 0.03436438739299774, + 0.026784982532262802, + 0.03541775420308113, + 0.002937354613095522, + 0.0041806017979979515, + 0.0032893717288970947, + 0.003488746238872409, + 0.0026732911355793476, + 0.0032512787729501724, + 0.001979410881176591, + 0.002241028007119894, + 0.026872172951698303, + 0.03926263749599457, + 0.0308059211820364, + 0.04997444525361061, + 0.01528154592961073, + 0.012541987001895905, + 0.011008081957697868, + 0.015544259920716286, + 0.007685042917728424, + 0.01007462851703167, + 0.012657960876822472, + 0.012531680054962635, + 0.005894738249480724, + 0.0035207991022616625, + 0.0024558648001402617, + 0.002450609114021063, + 0.026978252455592155, + 0.021929506212472916, + 0.014447614550590515, + 0.012697257101535797, + 0.011505465023219585, + 0.010057888925075531, + 0.01280234195291996, + 0.01008212473243475, + 0.017451515421271324, + 0.01354036945849657, + 0.012793687172234058, + 0.015197786502540112, + 0.009493368677794933, + 0.0143604576587677, + 0.011344576254487038, + 0.01119882334023714, + 0.004241038579493761, + 0.0047258660197257996, + 0.005670918617397547, + 0.008108695037662983, + 0.004781121853739023, + 0.003942538984119892, + 0.004020866937935352, + 0.0032038127537816763, + 0.0050739711150527, + 0.003593271831050515, + 0.005293527618050575, + 0.004608137533068657, + 0.003366560908034444, + 0.004052802454680204, + 0.003805771004408598, + 0.0036710621789097786, + 0.003573503578081727, + 0.003133303252980113, + 0.003218779806047678, + 0.002888207323849201, + 0.0021882837172597647, + 0.003473640652373433, + 0.002778226276859641, + 0.002906328998506069, + 0.010147927328944206, + 0.011232913471758366, + 0.009114883840084076, + 0.007893599569797516, + 0.005777049344033003, + 0.004748879000544548, + 0.004098041914403439, + 0.0065520890057086945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.007624335587024689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057098038494587 + ], + "opposite_logprob_tokens": [ + -3.895570993423462 + ], + "logprob_reward": -0.02057098038494587, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057098038494587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5232247114181519, + 0.06934819370508194, + 0.06309272348880768, + 0.02427055686712265, + 0.06546641141176224, + 0.11507876217365265, + 0.027415364980697632, + 0.027732189744710922, + 0.048418767750263214, + 0.0359523706138134 + ], + "credit_map": [ + 0.05635346099734306, + 0.048545654863119125, + 0.1473941206932068, + 0.0474315769970417, + 0.034856703132390976, + 0.04910462349653244, + 0.039112601429224014, + 0.10042592883110046, + 0.00546635314822197, + 0.012242886237800121, + 0.009191210381686687, + 0.017125651240348816, + 0.006495641078799963, + 0.007782320026308298, + 0.004723872058093548, + 0.0063202595338225365, + 0.008118517696857452, + 0.010799473151564598, + 0.010272093117237091, + 0.015064310282468796, + 0.004763155709952116, + 0.00416056253015995, + 0.004444373305886984, + 0.005470242351293564, + 0.0034023793414235115, + 0.005317851435393095, + 0.0031010955572128296, + 0.0034236619248986244, + 0.0030219657346606255, + 0.002356983022764325, + 0.0014983195578679442, + 0.0021483011078089476, + 0.01383200567215681, + 0.013131787069141865, + 0.00942948367446661, + 0.0074836029671132565, + 0.00578730134293437, + 0.007288225460797548, + 0.004394774325191975, + 0.004119234625250101, + 0.01649256981909275, + 0.013727307319641113, + 0.01475217193365097, + 0.015947895124554634, + 0.011007663793861866, + 0.011600963771343231, + 0.01071123592555523, + 0.02083895355463028, + 0.00455743633210659, + 0.0031752479262650013, + 0.004473516251891851, + 0.004135889001190662, + 0.0015335001517087221, + 0.002379131969064474, + 0.0018707786221057177, + 0.0052898651920259, + 0.004744392819702625, + 0.00259158737026155, + 0.0038864784874022007, + 0.0038011150900274515, + 0.002939260797575116, + 0.0034556922037154436, + 0.003270074026659131, + 0.0030435887165367603, + 0.008122566156089306, + 0.00635076267644763, + 0.006842327304184437, + 0.0053344122134149075, + 0.004688221029937267, + 0.004922823049128056, + 0.007554787211120129, + 0.004602864384651184, + 0.007019281852990389, + 0.009221742860972881, + 0.006158244330435991, + 0.0029198171105235815, + 0.0030135242268443108, + 0.002652292139828205, + 0.0020301195327192545, + 0.0029373455327004194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02057098038494587 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene opens with a high-angle, wide shot, the camera slowly tilting downward to follow a young, happy Asian child as he expertly rides a BMX bike through a park. The child, about seven years old, has an expression of pure, focused joy, thick, dark curls peeking out from under a vibrant orange beanie. They're dressed in a layered, comfortable style—a slightly oversized gray zip-up hoodie paired with a plain white T-shirt and dark, durable jeans. Their small frame rests on a compact, matte black BMX bike, which they expertly maneuver. By deliberately shifting their weight, the child leans back, arms extended, and lifts the front wheel off the ground in a slow, steady manual effort. The bike glides gracefully and steadily along the smooth concrete path. The entire park is bathed in the magical glow of dusk's golden hour. A low, warm sun slants through a thin, ethereal layer of mist, casting long, soft shadows that make the air itself seem to glow. The background reveals the soft outlines of tall trees and a park bench in the distance, all softened by a light mist, creating a serene, dreamy and deeply peaceful atmosphere that celebrates a moment of youthful freedom and skill.", + "seed": 2436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_5/index_8292_seed_2436_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": -0.04242324726656079, + "vlm_logprob": -0.04242324726656079, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.04242324726656079, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14666801691055298 + ], + "opposite_logprob_tokens": [ + -2.146667957305908 + ], + "logprob_reward": -0.14666801691055298, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14666801691055298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12808065116405487, + 0.046687014400959015, + 0.05672826990485191, + 0.06423480063676834, + 0.13514117896556854, + 0.3449274003505707, + 0.03268162161111832, + 0.03040388412773609, + 0.04313166067004204, + 0.11798352748155594 + ], + "credit_map": [ + 0.0068745058961212635, + 0.005820701364427805, + 0.005154518876224756, + 0.005757768638432026, + 0.013032898306846619, + 0.04303375631570816, + 0.007653359789401293, + 0.040753137320280075, + 0.004746131598949432, + 0.006508580408990383, + 0.004962513688951731, + 0.004838532768189907, + 0.008294820785522461, + 0.006517548579722643, + 0.005033691879361868, + 0.005785197485238314, + 0.008906987495720387, + 0.005225752480328083, + 0.01186919491738081, + 0.003877529874444008, + 0.0061539551243186, + 0.008888373151421547, + 0.004675374366343021, + 0.007131101563572884, + 0.009922066703438759, + 0.005899772979319096, + 0.01270416658371687, + 0.004862817469984293, + 0.006267516873776913, + 0.010169189423322678, + 0.005409799981862307, + 0.008999472483992577, + 0.020987408235669136, + 0.02439901791512966, + 0.0244589950889349, + 0.008133322931826115, + 0.011044073849916458, + 0.016895318403840065, + 0.010450208559632301, + 0.01877283863723278, + 0.04723673313856125, + 0.0800420269370079, + 0.04150814935564995, + 0.027892377227544785, + 0.06405029445886612, + 0.03391089662909508, + 0.028133098036050797, + 0.02215379849076271, + 0.005838245619088411, + 0.0048508429899811745, + 0.0048002623952925205, + 0.0038338101003319025, + 0.0029329494573175907, + 0.0038225529715418816, + 0.002983351005241275, + 0.003619611728936434, + 0.005081772338598967, + 0.003540136618539691, + 0.004387230146676302, + 0.0040700482204556465, + 0.004911992233246565, + 0.003289208747446537, + 0.002079598605632782, + 0.0030438953544944525, + 0.004405319225043058, + 0.007435136940330267, + 0.006184497382491827, + 0.005686587654054165, + 0.004846874624490738, + 0.004861242137849331, + 0.004386625252664089, + 0.005325379315763712, + 0.009135488420724869, + 0.007776797749102116, + 0.018643852323293686, + 0.013591053895652294, + 0.014446821995079517, + 0.022571789100766182, + 0.017447886988520622, + 0.014369833283126354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.14666801691055298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812266141176224 + ], + "opposite_logprob_tokens": [ + -3.2881226539611816 + ], + "logprob_reward": -0.03812266141176224, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812266141176224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09484253078699112, + 0.09958065301179886, + 0.03094855323433876, + 0.031574200838804245, + 0.06915022432804108, + 0.4569610357284546, + 0.04809552803635597, + 0.05528726056218147, + 0.07807399332523346, + 0.035486042499542236 + ], + "credit_map": [ + 0.008502176962792873, + 0.011604752391576767, + 0.005030896980315447, + 0.00763896107673645, + 0.014280201867222786, + 0.023386964574456215, + 0.008012155070900917, + 0.01638641394674778, + 0.011301133781671524, + 0.017996588721871376, + 0.009633444249629974, + 0.01052005309611559, + 0.015450351871550083, + 0.013381754979491234, + 0.008701502345502377, + 0.01259582955390215, + 0.004581871908158064, + 0.0037391765508800745, + 0.004511340521275997, + 0.0029726685024797916, + 0.0029335985891520977, + 0.004174185451120138, + 0.0036803181283175945, + 0.004355394281446934, + 0.004068049136549234, + 0.0038977412041276693, + 0.005108352284878492, + 0.0028372181113809347, + 0.00397211778908968, + 0.004964530933648348, + 0.002857168670743704, + 0.0038690215442329645, + 0.008216618560254574, + 0.011496249586343765, + 0.010548187419772148, + 0.005061442963778973, + 0.005387758370488882, + 0.011650440283119678, + 0.007502885069698095, + 0.009286638349294662, + 0.06561194360256195, + 0.10655062645673752, + 0.05428103730082512, + 0.03608598932623863, + 0.08450545370578766, + 0.04493483528494835, + 0.036525942385196686, + 0.028465230017900467, + 0.0077021424658596516, + 0.004925244487822056, + 0.005920094437897205, + 0.005246364511549473, + 0.0034503580536693335, + 0.00899682380259037, + 0.003741456661373377, + 0.008113044314086437, + 0.005027061328291893, + 0.004915360826998949, + 0.007164355833083391, + 0.008103291504085064, + 0.011062796227633953, + 0.007649320177733898, + 0.005248622503131628, + 0.006116453558206558, + 0.008723175153136253, + 0.009795472025871277, + 0.011956177651882172, + 0.010037010535597801, + 0.011511191725730896, + 0.007834821939468384, + 0.011115670204162598, + 0.007100472692400217, + 0.004586539231240749, + 0.0029197155963629484, + 0.0065414332784712315, + 0.0038133671041578054, + 0.003955118358135223, + 0.006733219139277935, + 0.003223082982003689, + 0.0037135668098926544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.03812266141176224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762729300186038 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.00762729300186038, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762729300186038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1283217817544937, + 0.03748389333486557, + 0.03622591495513916, + 0.03512844443321228, + 0.09829755127429962, + 0.5752666592597961, + 0.01824934221804142, + 0.016127225011587143, + 0.022487686946988106, + 0.032411474734544754 + ], + "credit_map": [ + 0.0072606587782502174, + 0.007486060727387667, + 0.004226002376526594, + 0.0048943134024739265, + 0.011439288035035133, + 0.04363337531685829, + 0.006835519336163998, + 0.04254657030105591, + 0.0036151912063360214, + 0.00610749376937747, + 0.0037571764551103115, + 0.0035856678150594234, + 0.005663858726620674, + 0.004639870952814817, + 0.004853059537708759, + 0.005261576734483242, + 0.006301005836576223, + 0.002496402245014906, + 0.007656530011445284, + 0.0017951433546841145, + 0.0035305407363921404, + 0.007919334806501865, + 0.0018320729723200202, + 0.00469488650560379, + 0.005132805090397596, + 0.004056529141962528, + 0.006049760151654482, + 0.0031788235064595938, + 0.0036271968856453896, + 0.0054042888805270195, + 0.0030695628374814987, + 0.004609480034559965, + 0.01447928138077259, + 0.01821885071694851, + 0.016663668677210808, + 0.005637860391288996, + 0.007863981649279594, + 0.014550329186022282, + 0.007673891261219978, + 0.013209686614573002, + 0.08408965170383453, + 0.1354919672012329, + 0.06725823134183884, + 0.04502040892839432, + 0.10522297769784927, + 0.057923365384340286, + 0.04459523409605026, + 0.03566485643386841, + 0.0030849333852529526, + 0.0020788998808711767, + 0.0028048953972756863, + 0.00159823230933398, + 0.0018555988790467381, + 0.0028345035389065742, + 0.001477248384617269, + 0.0025150321889668703, + 0.0024895702954381704, + 0.002991311950609088, + 0.002241977257654071, + 0.0018773346673697233, + 0.0021587046794593334, + 0.001290482236072421, + 0.0012501649325713515, + 0.001827679225243628, + 0.0026000728830695152, + 0.0027784595731645823, + 0.002682974562048912, + 0.0025704815052449703, + 0.0034693824127316475, + 0.0026008544955402613, + 0.003366881748661399, + 0.002418583258986473, + 0.0039038776885718107, + 0.0024390299804508686, + 0.00807151012122631, + 0.0026000612415373325, + 0.0033785777632147074, + 0.00613470608368516, + 0.0037142906803637743, + 0.002169423270970583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.00762729300186038 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631540462374687 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.008631540462374687, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631540462374687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06000473350286484, + 0.02861432544887066, + 0.02663188800215721, + 0.04018586501479149, + 0.13445886969566345, + 0.6640849709510803, + 0.011480121873319149, + 0.007956406101584435, + 0.017577538266777992, + 0.00900530256330967 + ], + "credit_map": [ + 0.005695302039384842, + 0.005431637167930603, + 0.004007542505860329, + 0.003867704886943102, + 0.011338273994624615, + 0.011105760931968689, + 0.005812540650367737, + 0.012745970860123634, + 0.002312503522261977, + 0.003123764880001545, + 0.003526951652020216, + 0.0026523759588599205, + 0.005092577077448368, + 0.004094078671187162, + 0.00370951765216887, + 0.0041025555692613125, + 0.002987440675497055, + 0.0027032492216676474, + 0.005700023379176855, + 0.0019246443407610059, + 0.002397479023784399, + 0.004841030575335026, + 0.002359941601753235, + 0.0037180755753070116, + 0.0051424321718513966, + 0.004805182572454214, + 0.006295585073530674, + 0.0033813787158578634, + 0.0046894545666873455, + 0.00555600319057703, + 0.00442164484411478, + 0.005894179455935955, + 0.018556546419858932, + 0.020047150552272797, + 0.02292725257575512, + 0.008579126559197903, + 0.010055222548544407, + 0.023417117074131966, + 0.012224551290273666, + 0.01865188032388687, + 0.09253399819135666, + 0.15400050580501556, + 0.07943537086248398, + 0.05329485610127449, + 0.1238258108496666, + 0.06568977981805801, + 0.05328274890780449, + 0.04202185571193695, + 0.0025989357382059097, + 0.0016842001350596547, + 0.0014421169180423021, + 0.0018533668480813503, + 0.0007073090528137982, + 0.0019005308859050274, + 0.0005828127032145858, + 0.0007108482532203197, + 0.0010075580794364214, + 0.0009160413173958659, + 0.0010684749577194452, + 0.0007880296325311065, + 0.0013881416525691748, + 0.0009808654431253672, + 0.0006496987771242857, + 0.0011575964745134115, + 0.002435202244669199, + 0.0019420274766162038, + 0.0022543061058968306, + 0.002379276091232896, + 0.0024677205365151167, + 0.002123631536960602, + 0.00207105022855103, + 0.0019043228821828961, + 0.0014341974165290594, + 0.000617689685896039, + 0.0010772666428238153, + 0.0012945026392117143, + 0.001138487015850842, + 0.0011353642912581563, + 0.0012843483127653599, + 0.0010234457440674305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.008631540462374687 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle, wide shot that slowly tilts downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young Asian child (about seven years old) wearing a vibrant orange beanie and a gray zip-up hoodie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the child riding a compact, matte black BMX bike and performing a slow, steady manual (lifting the front wheel)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a park with a smooth concrete path, tall trees, and a park bench in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low, warm sun at dusk with a thin layer of mist casting long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066724546253681 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011066724546253681, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066724546253681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1501884162425995, + 0.11319717019796371, + 0.1217636913061142, + 0.027094891294836998, + 0.11012914031744003, + 0.33694133162498474, + 0.0368824303150177, + 0.02030552551150322, + 0.020491745322942734, + 0.06300575286149979 + ], + "credit_map": [ + 0.009609706699848175, + 0.009059624746441841, + 0.005952389445155859, + 0.005746694281697273, + 0.01989947073161602, + 0.04854382947087288, + 0.00943753495812416, + 0.041939157992601395, + 0.011462748982012272, + 0.013659275136888027, + 0.012532140128314495, + 0.00859749410301447, + 0.019561633467674255, + 0.01707066036760807, + 0.015107020735740662, + 0.015206187963485718, + 0.015360121615231037, + 0.008047427982091904, + 0.02740110084414482, + 0.006361941806972027, + 0.012229814194142818, + 0.02726682275533676, + 0.007735153194516897, + 0.017361296340823174, + 0.0041716271080076694, + 0.002743742661550641, + 0.004975767340511084, + 0.0017938718665391207, + 0.00252859340980649, + 0.004135941155254841, + 0.0027453287038952112, + 0.004000017885118723, + 0.016904085874557495, + 0.017017526552081108, + 0.020509695634245872, + 0.006090078037232161, + 0.008250231854617596, + 0.016892028972506523, + 0.008173123002052307, + 0.016292357817292213, + 0.046517591923475266, + 0.07827448844909668, + 0.040513698011636734, + 0.027100812643766403, + 0.06261850148439407, + 0.03314590081572533, + 0.027354983612895012, + 0.0214153490960598, + 0.008813625201582909, + 0.0060274736024439335, + 0.005123504437506199, + 0.0035583474673330784, + 0.002601941814646125, + 0.006377521436661482, + 0.002114413306117058, + 0.0022655969951301813, + 0.003886862425133586, + 0.0026243356987833977, + 0.002844809787347913, + 0.001423030043952167, + 0.003545533400028944, + 0.0022843724582344294, + 0.001661674934439361, + 0.002034904435276985, + 0.002545478055253625, + 0.003653261112049222, + 0.0029084929265081882, + 0.0022682324051856995, + 0.0023270181845873594, + 0.002583520021289587, + 0.0018293331377208233, + 0.002376406453549862, + 0.009879600256681442, + 0.004355307202786207, + 0.015484756790101528, + 0.006780742201954126, + 0.0056389993987977505, + 0.009673778899013996, + 0.006694348528981209, + 0.004498211666941643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.011066724546253681 + } + ], + "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/0000052/group_6/index_tensor([5733, 5733])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_6/index_tensor([5733, 5733])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e1e4cbddddd69422eae27c44163d4f21efc1092c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_6/index_tensor([5733, 5733])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium close-up captures a moment of quiet discovery in a dense bamboo forest. Dappled sunlight filters through the tall canopy, creating soft, shifting patterns of light and shadow. The furry head of a giant panda cautiously emerges from behind a thick screen of vibrant green bamboo stems and leaves. Its fur is a dense, perfect white, each hair catching the soft light and contrasting sharply with the signature, soft-rimmed black patch surrounding its visible eye. From within that field, a single dark, gleaming, inquisitive eye peers out, its gaze gentle and inquisitive. The panda's head tilts slightly, slowly and deliberately advancing, causing several delicate bamboo clumps to part, its black nose twitching subtly. The entire action is executed with extreme caution, creating an atmosphere of tranquility and intimacy, as if the viewer has stumbled upon a secret, private moment deep within the heart of nature.", + "seed": 30408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_6/index_5733_seed_30408_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": -0.059532348823267967, + "vlm_logprob": -0.059532348823267967, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.059532348823267967, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08577068895101547 + ], + "opposite_logprob_tokens": [ + -3.210770606994629 + ], + "logprob_reward": -0.08577068895101547, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.08577068895101547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05455492436885834, + 0.09526830911636353, + 0.0716644674539566, + 0.05327202379703522, + 0.08769067376852036, + 0.09575305134057999, + 0.0458119660615921, + 0.27173423767089844, + 0.13243535161018372, + 0.0918150320649147 + ], + "credit_map": [ + 0.006885995622724295, + 0.004827678203582764, + 0.013988110236823559, + 0.009769056923687458, + 0.004781175870448351, + 0.004670410417020321, + 0.005613786168396473, + 0.004018709529191256, + 0.017971284687519073, + 0.009220181964337826, + 0.009676988236606121, + 0.008351104333996773, + 0.009926235303282738, + 0.019815314561128616, + 0.008192924782633781, + 0.012114287354052067, + 0.00895584188401699, + 0.005734608508646488, + 0.011201323941349983, + 0.009363043121993542, + 0.009806721471250057, + 0.008578689768910408, + 0.008333498612046242, + 0.009690734557807446, + 0.005267767235636711, + 0.004668086767196655, + 0.006663170643150806, + 0.010192920453846455, + 0.007314844988286495, + 0.005701234098523855, + 0.006684443447738886, + 0.006779559887945652, + 0.008777095936238766, + 0.009747116826474667, + 0.010456054471433163, + 0.01220745500177145, + 0.008595292456448078, + 0.009500442072749138, + 0.013263052329421043, + 0.0151441665366292, + 0.010550204664468765, + 0.005244414787739515, + 0.025134583935141563, + 0.01682756096124649, + 0.004715552553534508, + 0.007574147079139948, + 0.013927518390119076, + 0.01177907083183527, + 0.005554468370974064, + 0.003854178823530674, + 0.004697268828749657, + 0.004157521761953831, + 0.0032996307127177715, + 0.005444616079330444, + 0.005536784417927265, + 0.013267495669424534, + 0.033051785081624985, + 0.0418354794383049, + 0.032991115003824234, + 0.041760459542274475, + 0.02772240899503231, + 0.0366302914917469, + 0.029245764017105103, + 0.02849689871072769, + 0.013568920083343983, + 0.011685232631862164, + 0.021142465993762016, + 0.02490292489528656, + 0.010335053317248821, + 0.021888839080929756, + 0.01256770920008421, + 0.016344206407666206, + 0.008264821954071522, + 0.02585500292479992, + 0.010514428839087486, + 0.010298941284418106, + 0.007034026551991701, + 0.008448262698948383, + 0.006774734705686569, + 0.014624809846282005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.08577068895101547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042770445346832 + ], + "opposite_logprob_tokens": [ + -1.9104276895523071 + ], + "logprob_reward": -0.16042770445346832, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042770445346832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13120037317276, + 0.06712193787097931, + 0.15079981088638306, + 0.050379808992147446, + 0.03190598264336586, + 0.49252238869667053, + 0.020338715985417366, + 0.018274663016200066, + 0.023044634610414505, + 0.014411707408726215 + ], + "credit_map": [ + 0.019456539303064346, + 0.03003544919192791, + 0.00925004854798317, + 0.012616847641766071, + 0.01755739189684391, + 0.018387239426374435, + 0.012815441004931927, + 0.011081424541771412, + 0.007703568786382675, + 0.008930113166570663, + 0.010048608295619488, + 0.00804429966956377, + 0.00880781002342701, + 0.006822316441684961, + 0.008295819163322449, + 0.008469406515359879, + 0.012277554720640182, + 0.006430909503251314, + 0.05275343358516693, + 0.019085347652435303, + 0.012093229219317436, + 0.01001770980656147, + 0.013601981103420258, + 0.024539662525057793, + 0.003976150415837765, + 0.005864896811544895, + 0.008267827332019806, + 0.009071184322237968, + 0.006911483593285084, + 0.004027566406875849, + 0.005579985212534666, + 0.006680720020085573, + 0.005303698126226664, + 0.004201792646199465, + 0.004139884375035763, + 0.0037368955090641975, + 0.003731528762727976, + 0.0026957527734339237, + 0.0037499063182622194, + 0.0043465266935527325, + 0.06379376351833344, + 0.026753269135951996, + 0.19175493717193604, + 0.04065268486738205, + 0.019698020070791245, + 0.018702827394008636, + 0.08549190312623978, + 0.04567499831318855, + 0.00228897575289011, + 0.0017728408565744758, + 0.0024805546272546053, + 0.0024935698602348566, + 0.0015254839090630412, + 0.002619127742946148, + 0.003310393076390028, + 0.0038477699272334576, + 0.002445744350552559, + 0.0020718122832477093, + 0.003027810715138912, + 0.002833425533026457, + 0.0015609599649906158, + 0.00184550229460001, + 0.0022679055109620094, + 0.0022215035278350115, + 0.0017347416142001748, + 0.002627883804962039, + 0.003767507616430521, + 0.0038499527145177126, + 0.0036847807932645082, + 0.003271978348493576, + 0.0024396225344389677, + 0.001668170210905373, + 0.0016995874466374516, + 0.001188321621157229, + 0.0017456458881497383, + 0.0019477840978652239, + 0.0013429500395432115, + 0.0017235770355910063, + 0.0023018554784357548, + 0.0024619861505925655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.16042770445346832 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008220868767239153 + ], + "opposite_logprob_tokens": [ + -7.125822067260742 + ], + "logprob_reward": -0.0008220868767239153, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008220868767239153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05829285457730293, + 0.0707741379737854, + 0.16333448886871338, + 0.05493883788585663, + 0.17800863087177277, + 0.20260177552700043, + 0.06975455582141876, + 0.10062896460294724, + 0.0662698894739151, + 0.03539583832025528 + ], + "credit_map": [ + 0.008774763904511929, + 0.005579011049121618, + 0.009039757773280144, + 0.011611765250563622, + 0.005696878768503666, + 0.005884767975658178, + 0.007110936101526022, + 0.004594970494508743, + 0.012936788611114025, + 0.012911050580441952, + 0.007610725704580545, + 0.006455859635025263, + 0.007638461422175169, + 0.01131722703576088, + 0.005551624111831188, + 0.0063524022698402405, + 0.010926833376288414, + 0.007189110387116671, + 0.06424858421087265, + 0.02056116797029972, + 0.010632549412548542, + 0.006978544406592846, + 0.008169450797140598, + 0.034628257155418396, + 0.004982370417565107, + 0.004787121899425983, + 0.00792520772665739, + 0.011000392958521843, + 0.00866033136844635, + 0.005393501371145248, + 0.00583622558042407, + 0.0063536870293319225, + 0.024750588461756706, + 0.026985744014382362, + 0.02448098734021187, + 0.028210796415805817, + 0.01671425811946392, + 0.018656328320503235, + 0.016943782567977905, + 0.02126612700521946, + 0.02494678646326065, + 0.008061928674578667, + 0.08268611878156662, + 0.017933785915374756, + 0.006194500718265772, + 0.006861360277980566, + 0.03789331763982773, + 0.018023988232016563, + 0.0075836279429495335, + 0.0066069685854017735, + 0.009846026077866554, + 0.009579150006175041, + 0.006783611606806517, + 0.008598772808909416, + 0.010039689019322395, + 0.010716712102293968, + 0.012589724734425545, + 0.010406604036688805, + 0.015270543284714222, + 0.015469346195459366, + 0.008611583150923252, + 0.011471876874566078, + 0.013839471153914928, + 0.012969817966222763, + 0.006117760669440031, + 0.0069722943007946014, + 0.0109027698636055, + 0.011530361138284206, + 0.009391860105097294, + 0.00667530857026577, + 0.008616268634796143, + 0.0060632661916315556, + 0.004689349327236414, + 0.004233699291944504, + 0.005366974510252476, + 0.006745642516762018, + 0.0028496216982603073, + 0.00507546728476882, + 0.0035777855664491653, + 0.0028572974260896444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.0008220868767239153 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544653378427029 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544653378427029, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544653378427029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02147934027016163, + 0.019754918292164803, + 0.01754157245159149, + 0.017452791333198547, + 0.03948018699884415, + 0.6962015628814697, + 0.040592510253190994, + 0.11532679945230484, + 0.015724502503871918, + 0.0164458230137825 + ], + "credit_map": [ + 0.002826896496117115, + 0.004595372825860977, + 0.002413837006315589, + 0.0022148822899907827, + 0.0029004551470279694, + 0.0022892686538398266, + 0.0024406742304563522, + 0.0017979555996134877, + 0.0025712689384818077, + 0.002462874399498105, + 0.0029834068845957518, + 0.002785119693726301, + 0.0022070317063480616, + 0.0019222189439460635, + 0.00240976526401937, + 0.0024132325779646635, + 0.0025356195401400328, + 0.0015907667111605406, + 0.003575877984985709, + 0.0027515413239598274, + 0.0019099435303360224, + 0.0013119849609211087, + 0.002088617766276002, + 0.0017772214487195015, + 0.0015168398385867476, + 0.0018641799688339233, + 0.003227019915357232, + 0.003866784041747451, + 0.002061951206997037, + 0.0017574750818312168, + 0.0013300401624292135, + 0.0018285042606294155, + 0.004035061690956354, + 0.004362518899142742, + 0.004392902832478285, + 0.004625893197953701, + 0.0033821966499090195, + 0.005345093552023172, + 0.006819855887442827, + 0.006516666151583195, + 0.09112474322319031, + 0.030077442526817322, + 0.2826240360736847, + 0.05549253895878792, + 0.02271828055381775, + 0.02355489507317543, + 0.1267015039920807, + 0.0639081820845604, + 0.0032774368301033974, + 0.003153142984956503, + 0.004765888210386038, + 0.007456360850483179, + 0.002259695902466774, + 0.006672794930636883, + 0.005337433889508247, + 0.007669760379940271, + 0.011830857954919338, + 0.011941291391849518, + 0.019440101459622383, + 0.019740717485547066, + 0.013879009522497654, + 0.009707904420793056, + 0.012439963407814503, + 0.01634696125984192, + 0.0013784191105514765, + 0.0020495823118835688, + 0.004067082889378071, + 0.0016193463234230876, + 0.0011750312987715006, + 0.002586045768111944, + 0.0014035353669896722, + 0.0014454605989158154, + 0.0020886585116386414, + 0.00172588846180588, + 0.002741864649578929, + 0.00270223175175488, + 0.0014176917029544711, + 0.0021546767093241215, + 0.0015700515359640121, + 0.0020447613205760717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.012544653378427029 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.042287543416023254, + 0.08945021033287048, + 0.06107836589217186, + 0.07532091438770294, + 0.3355555236339569, + 0.24744215607643127, + 0.023876875638961792, + 0.08746182173490524, + 0.027721351012587547, + 0.009805265814065933 + ], + "credit_map": [ + 0.00576880993321538, + 0.005829555913805962, + 0.006940817926079035, + 0.007636637426912785, + 0.0036030856426805258, + 0.004637365695089102, + 0.004666486289352179, + 0.003204782959073782, + 0.018582642078399658, + 0.014978300780057907, + 0.009079627692699432, + 0.007580226752907038, + 0.007052636705338955, + 0.017834806814789772, + 0.006916452199220657, + 0.0074255201034247875, + 0.004208868369460106, + 0.003443004796281457, + 0.024512307718396187, + 0.0076438868418335915, + 0.0035501534584909678, + 0.0022242802660912275, + 0.0029172205831855536, + 0.01257864385843277, + 0.007279445417225361, + 0.007418438326567411, + 0.013562969863414764, + 0.017694851383566856, + 0.007384254597127438, + 0.004506052937358618, + 0.006065509282052517, + 0.011409389786422253, + 0.05193183198571205, + 0.05141549929976463, + 0.0531080886721611, + 0.05721636116504669, + 0.0333850160241127, + 0.030876215547323227, + 0.026550890877842903, + 0.03107164427638054, + 0.02765929140150547, + 0.014027681201696396, + 0.0899864137172699, + 0.028114328160881996, + 0.012610084377229214, + 0.01290274690836668, + 0.04143980145454407, + 0.020701808854937553, + 0.0019181694369763136, + 0.0015481975860893726, + 0.002657648641616106, + 0.0030810933094471693, + 0.0014720162143930793, + 0.0038993393536657095, + 0.004426819272339344, + 0.004873592872172594, + 0.008480369113385677, + 0.007970510050654411, + 0.0168139711022377, + 0.0146786505356431, + 0.010549886152148247, + 0.006554966792464256, + 0.009413469582796097, + 0.012999992817640305, + 0.0025808128993958235, + 0.002457820111885667, + 0.0038909483700990677, + 0.006292130332440138, + 0.002253513550385833, + 0.003410870674997568, + 0.0025178389623761177, + 0.00431741401553154, + 0.001533546019345522, + 0.0015071152010932565, + 0.001310165855102241, + 0.0010479558259248734, + 0.0008466299041174352, + 0.0010406920919194818, + 0.0011525454465299845, + 0.001366614829748869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.03809661045670509 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a moment of quiet discovery in a dense bamboo forest. Dappled sunlight filters through the tall canopy, creating soft, shifting patterns of light and shadow. The furry head of a giant panda cautiously emerges from behind a thick screen of vibrant green bamboo stems and leaves. Its fur is a dense, perfect white, each hair catching the soft light and contrasting sharply with the signature, soft-rimmed black patch surrounding its visible eye. From within that field, a single dark, gleaming, inquisitive eye peers out, its gaze gentle and inquisitive. The panda's head tilts slightly, slowly and deliberately advancing, causing several delicate bamboo clumps to part, its black nose twitching subtly. The entire action is executed with extreme caution, creating an atmosphere of tranquility and intimacy, as if the viewer has stumbled upon a secret, private moment deep within the heart of nature.", + "seed": 30409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_6/index_5733_seed_30409_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": -0.14661241308785974, + "vlm_logprob": -0.14661241308785974, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.14661241308785974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13585272431373596 + ], + "opposite_logprob_tokens": [ + -2.510852813720703 + ], + "logprob_reward": -0.13585272431373596, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.13585272431373596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03797805681824684, + 0.025523951277136803, + 0.09847024828195572, + 0.3811458647251129, + 0.03189070522785187, + 0.09444475173950195, + 0.2819884419441223, + 0.012150616385042667, + 0.014421840198338032, + 0.021985458210110664 + ], + "credit_map": [ + 0.002968845423310995, + 0.0046291048638522625, + 0.003618853632360697, + 0.003605496371164918, + 0.005059398245066404, + 0.005259839352220297, + 0.007097112946212292, + 0.0057394071482121944, + 0.0030262607615441084, + 0.0029605375602841377, + 0.0035679021384567022, + 0.0029922376852482557, + 0.002249212935566902, + 0.003930133301764727, + 0.002848970703780651, + 0.003948700148612261, + 0.011450624093413353, + 0.00813339464366436, + 0.012087840586900711, + 0.0057267132215201855, + 0.028599610552191734, + 0.008026499301195145, + 0.0068117533810436726, + 0.017633823677897453, + 0.03561599180102348, + 0.07069256156682968, + 0.012677364982664585, + 0.011856204830110073, + 0.1368045210838318, + 0.031603481620550156, + 0.015876928344368935, + 0.06601886451244354, + 0.0031846906058490276, + 0.0038469901774078608, + 0.003158785402774811, + 0.00288579473271966, + 0.0033650915138423443, + 0.0061338855884969234, + 0.0030014999210834503, + 0.006313970312476158, + 0.009176480583846569, + 0.02418603003025055, + 0.0072130016051232815, + 0.00872343685477972, + 0.015189691446721554, + 0.008886299096047878, + 0.008063201792538166, + 0.013006624765694141, + 0.0316021554172039, + 0.026046888902783394, + 0.04488373175263405, + 0.024479612708091736, + 0.02229168266057968, + 0.07147206366062164, + 0.01836041547358036, + 0.04285191372036934, + 0.0012849555350840092, + 0.001388882752507925, + 0.0010691320057958364, + 0.0014417414786294103, + 0.0016710731433704495, + 0.002140627708286047, + 0.0015447235200554132, + 0.0016094817547127604, + 0.0012284780386835337, + 0.0011916436487808824, + 0.0013608421431854367, + 0.002696485025808215, + 0.0013908600667491555, + 0.0024288969580084085, + 0.0016311958897858858, + 0.0024934406392276287, + 0.0017220270819962025, + 0.0026697285939007998, + 0.002221201080828905, + 0.00400106655433774, + 0.002167345490306616, + 0.0026553638745099306, + 0.0038655551616102457, + 0.0026831738650798798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.13585272431373596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233625173568726 + ], + "opposite_logprob_tokens": [ + -0.8983625173568726 + ], + "logprob_reward": -0.5233625173568726, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233625173568726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05901828780770302, + 0.0679507777094841, + 0.04577125608921051, + 0.2500392198562622, + 0.06083547696471214, + 0.24889133870601654, + 0.21419090032577515, + 0.017999956384301186, + 0.0058250040747225285, + 0.02947780303657055 + ], + "credit_map": [ + 0.00825467612594366, + 0.006110865157097578, + 0.010377413593232632, + 0.007262632250785828, + 0.006226355209946632, + 0.007119615562260151, + 0.007251490838825703, + 0.006415235809981823, + 0.01183850783854723, + 0.007032901048660278, + 0.01074607577174902, + 0.0049441177397966385, + 0.006895630620419979, + 0.006869015283882618, + 0.0037757179234176874, + 0.015848815441131592, + 0.005801918916404247, + 0.0033574490807950497, + 0.008980022743344307, + 0.003240083111450076, + 0.009735732339322567, + 0.003619593335315585, + 0.0017951307818293571, + 0.009241328574717045, + 0.02313576452434063, + 0.04434709995985031, + 0.008896778337657452, + 0.008646749891340733, + 0.08810711652040482, + 0.021320456638932228, + 0.010440231300890446, + 0.04514504224061966, + 0.007734255865216255, + 0.006521818693727255, + 0.00702318362891674, + 0.007022927515208721, + 0.007040403317660093, + 0.00945302564650774, + 0.005234420299530029, + 0.010805441997945309, + 0.028614353388547897, + 0.06818270683288574, + 0.007886836305260658, + 0.009358351118862629, + 0.03311111778020859, + 0.02056927978992462, + 0.009907609783113003, + 0.07126107811927795, + 0.025312960147857666, + 0.01901000551879406, + 0.035679515451192856, + 0.017219284549355507, + 0.015332159586250782, + 0.05604683980345726, + 0.012026551179587841, + 0.03356358781456947, + 0.0012371896300464869, + 0.0012460836442187428, + 0.0036439807154238224, + 0.0036594246048480272, + 0.001494507770985365, + 0.0023305772338062525, + 0.0014723633648827672, + 0.0029158303514122963, + 0.0004328105424065143, + 0.0005279533215798438, + 0.0005755405290983617, + 0.0009762101108208299, + 0.00048201647587120533, + 0.001103461836464703, + 0.0006819969275966287, + 0.0010450143599882722, + 0.0024090015795081854, + 0.002736482536420226, + 0.0031713820062577724, + 0.004300748463720083, + 0.0022652943152934313, + 0.004383760970085859, + 0.005412749946117401, + 0.004798382055014372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.5233625173568726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028276958037167788 + ], + "opposite_logprob_tokens": [ + -5.8778276443481445 + ], + "logprob_reward": -0.0028276958037167788, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028276958037167788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07370566576719284, + 0.05354198068380356, + 0.12749303877353668, + 0.41059207916259766, + 0.05180088430643082, + 0.08604317158460617, + 0.05230928212404251, + 0.043533869087696075, + 0.06042056903243065, + 0.04055946320295334 + ], + "credit_map": [ + 0.005767673719674349, + 0.007971037179231644, + 0.006025554146617651, + 0.008310106582939625, + 0.010019512847065926, + 0.008846110664308071, + 0.015327605418860912, + 0.011438069865107536, + 0.009795624762773514, + 0.008088654838502407, + 0.005340991076081991, + 0.0040724570862948895, + 0.004335148259997368, + 0.006327364128082991, + 0.004640392493456602, + 0.010941347107291222, + 0.015417175367474556, + 0.014777859672904015, + 0.013790169730782509, + 0.009960255585610867, + 0.029949292540550232, + 0.015972964465618134, + 0.015321057289838791, + 0.01230426225811243, + 0.03849763795733452, + 0.07107919454574585, + 0.015614047646522522, + 0.016537023708224297, + 0.14185070991516113, + 0.03492974489927292, + 0.01737390086054802, + 0.07470982521772385, + 0.005840224679559469, + 0.005124378949403763, + 0.006146873347461224, + 0.0052336142398417, + 0.005636830348521471, + 0.008406111970543861, + 0.005972142796963453, + 0.009440707974135876, + 0.008439704775810242, + 0.01886141672730446, + 0.0064673773013055325, + 0.011235497891902924, + 0.012411104515194893, + 0.008257700130343437, + 0.01015222817659378, + 0.010218147188425064, + 0.006607737857848406, + 0.0035085510462522507, + 0.008967466652393341, + 0.004339961335062981, + 0.0032024714164435863, + 0.014106004498898983, + 0.0026421695947647095, + 0.008934926241636276, + 0.004547083284705877, + 0.004510068334639072, + 0.004327916540205479, + 0.00543855968862772, + 0.005502096377313137, + 0.00709049915894866, + 0.006035955157130957, + 0.006081688217818737, + 0.004278300330042839, + 0.005727359559386969, + 0.008780372329056263, + 0.012618651613593102, + 0.0067511326633393764, + 0.007698557805269957, + 0.0062700253911316395, + 0.008296164683997631, + 0.003339581424370408, + 0.004090892616659403, + 0.004845588468015194, + 0.008126205764710903, + 0.003769457805901766, + 0.004761084448546171, + 0.006065225228667259, + 0.005561430472880602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.0028276958037167788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607329398393631 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607329398393631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607329398393631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0729091688990593, + 0.08544973284006119, + 0.08482175320386887, + 0.16201290488243103, + 0.09190608561038971, + 0.23980741202831268, + 0.051825668662786484, + 0.05020115152001381, + 0.10944671928882599, + 0.05161941796541214 + ], + "credit_map": [ + 0.010048585012555122, + 0.01235106959939003, + 0.008137136697769165, + 0.0066585480235517025, + 0.00783538818359375, + 0.008595931343734264, + 0.008647230453789234, + 0.010635281912982464, + 0.014246907085180283, + 0.010529124177992344, + 0.012133428826928139, + 0.006191728636622429, + 0.012209567241370678, + 0.0113228689879179, + 0.0057121748104691505, + 0.013103934936225414, + 0.011099487543106079, + 0.008753024972975254, + 0.016888834536075592, + 0.007024116348475218, + 0.008028868585824966, + 0.011310743167996407, + 0.00822553038597107, + 0.01349114254117012, + 0.016009962186217308, + 0.025669271126389503, + 0.006913708057254553, + 0.008771567605435848, + 0.051770009100437164, + 0.015141448006033897, + 0.00846265722066164, + 0.029274296015501022, + 0.011612333357334137, + 0.007368726190179586, + 0.010981675237417221, + 0.010741115547716618, + 0.010176103562116623, + 0.014714431017637253, + 0.009159775450825691, + 0.01715192385017872, + 0.019400596618652344, + 0.06701330095529556, + 0.014394535683095455, + 0.020512655377388, + 0.037207815796136856, + 0.021531645208597183, + 0.01892785355448723, + 0.040818993002176285, + 0.006944316904991865, + 0.0032487153075635433, + 0.009017209522426128, + 0.005030878819525242, + 0.0031913116108626127, + 0.012881414964795113, + 0.0025525155942887068, + 0.008959303610026836, + 0.006049204617738724, + 0.004316193051636219, + 0.005527444649487734, + 0.005333507899194956, + 0.0030360675882548094, + 0.01328551210463047, + 0.004805681761354208, + 0.007847540080547333, + 0.007963398471474648, + 0.01083606481552124, + 0.009202448651194572, + 0.018871190026402473, + 0.01003153994679451, + 0.02349911816418171, + 0.00914503913372755, + 0.01989792101085186, + 0.004248116165399551, + 0.005336488131433725, + 0.00431597838178277, + 0.00930609181523323, + 0.004401138052344322, + 0.009185820817947388, + 0.007899945601820946, + 0.006925840862095356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.01607329398393631 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494583398103714 + ], + "opposite_logprob_tokens": [ + -2.929945945739746 + ], + "logprob_reward": -0.05494583398103714, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494583398103714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02068885788321495, + 0.02762557566165924, + 0.060006167739629745, + 0.6043893098831177, + 0.029496999457478523, + 0.041886378079652786, + 0.16308607161045074, + 0.014989805407822132, + 0.0182840246707201, + 0.01954677887260914 + ], + "credit_map": [ + 0.0024392171762883663, + 0.002191765932366252, + 0.0021150035317987204, + 0.002557971980422735, + 0.002714936388656497, + 0.002655573422089219, + 0.003301439806818962, + 0.0027129491791129112, + 0.004013329278677702, + 0.0024650455452501774, + 0.003982781432569027, + 0.0016142121748998761, + 0.004529633559286594, + 0.003864723490551114, + 0.001532726688310504, + 0.0056231203489005566, + 0.007128364406526089, + 0.004717435222119093, + 0.00835532508790493, + 0.003360101254656911, + 0.01721404492855072, + 0.004792376887053251, + 0.004286371171474457, + 0.010152147151529789, + 0.055807434022426605, + 0.10694762319326401, + 0.020804980769753456, + 0.021780923008918762, + 0.21233788132667542, + 0.051814183592796326, + 0.025299105793237686, + 0.10959722101688385, + 0.00408085435628891, + 0.0024986204225569963, + 0.003133573569357395, + 0.0033799156080931425, + 0.0026738420128822327, + 0.005474873352795839, + 0.002453345339745283, + 0.005801977124065161, + 0.0034961358178406954, + 0.009973433800041676, + 0.0033922509755939245, + 0.004719956312328577, + 0.0068732076324522495, + 0.0031642529647797346, + 0.004458515904843807, + 0.005808625370264053, + 0.019402842968702316, + 0.01346393208950758, + 0.02943737991154194, + 0.012890965677797794, + 0.01119457371532917, + 0.042884163558483124, + 0.008459535427391529, + 0.025352677330374718, + 0.0012990084942430258, + 0.0008587224874645472, + 0.0013140089577063918, + 0.0031664015259593725, + 0.0013198660453781486, + 0.003187921131029725, + 0.0014709992101415992, + 0.002372876275330782, + 0.0014820165233686566, + 0.0017358524492010474, + 0.0018551672110334039, + 0.003978649619966745, + 0.0021603761706501245, + 0.002153285313397646, + 0.0022131416480988264, + 0.0027055356185883284, + 0.0017095652874559164, + 0.0019815668929368258, + 0.0015562426997348666, + 0.0030703016091138124, + 0.0014093511272221804, + 0.003218812867999077, + 0.0022998491767793894, + 0.004301088862121105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.05494583398103714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a moment of quiet discovery in a dense bamboo forest. Dappled sunlight filters through the tall canopy, creating soft, shifting patterns of light and shadow. The furry head of a giant panda cautiously emerges from behind a thick screen of vibrant green bamboo stems and leaves. Its fur is a dense, perfect white, each hair catching the soft light and contrasting sharply with the signature, soft-rimmed black patch surrounding its visible eye. From within that field, a single dark, gleaming, inquisitive eye peers out, its gaze gentle and inquisitive. The panda's head tilts slightly, slowly and deliberately advancing, causing several delicate bamboo clumps to part, its black nose twitching subtly. The entire action is executed with extreme caution, creating an atmosphere of tranquility and intimacy, as if the viewer has stumbled upon a secret, private moment deep within the heart of nature.", + "seed": 30410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_6/index_5733_seed_30410_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": -0.0813481301185675, + "vlm_logprob": -0.0813481301185675, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.0813481301185675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0924711599946022 + ], + "opposite_logprob_tokens": [ + -3.342471122741699 + ], + "logprob_reward": -0.0924711599946022, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0924711599946022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055197492241859436, + 0.011605407111346722, + 0.03975384309887886, + 0.03981070965528488, + 0.03536917269229889, + 0.08209250122308731, + 0.0646323710680008, + 0.643462598323822, + 0.01708882302045822, + 0.01098704058676958 + ], + "credit_map": [ + 0.0029038486536592245, + 0.0057126800529658794, + 0.0022704307921230793, + 0.0027774914633482695, + 0.019802948459982872, + 0.005323873367160559, + 0.003282315330579877, + 0.013123907148838043, + 0.001123210764490068, + 0.0008822212112136185, + 0.0019039280014112592, + 0.0015239851782098413, + 0.001483369036577642, + 0.0015263059176504612, + 0.001454991870559752, + 0.0017073954222723842, + 0.007299641612917185, + 0.0011658573057502508, + 0.01464757788926363, + 0.00359923648647964, + 0.00478472700342536, + 0.00310372281819582, + 0.0013084921520203352, + 0.003844591323286295, + 0.005229582544416189, + 0.0048104748129844666, + 0.007061487529426813, + 0.004637915175408125, + 0.0033309964928776026, + 0.00515044666826725, + 0.004803161136806011, + 0.004786646459251642, + 0.008751451037824154, + 0.008525011129677296, + 0.005336655303835869, + 0.0030724508687853813, + 0.0025361005682498217, + 0.0020305842626839876, + 0.0028418120928108692, + 0.0022751097567379475, + 0.009219796396791935, + 0.007376829627901316, + 0.018130416050553322, + 0.017276398837566376, + 0.007207169197499752, + 0.006431554909795523, + 0.010878081433475018, + 0.005572259426116943, + 0.010756420902907848, + 0.00934969075024128, + 0.008282576687633991, + 0.007102572824805975, + 0.005766879767179489, + 0.007462908513844013, + 0.006834025029093027, + 0.009077301248908043, + 0.041251685470342636, + 0.21766448020935059, + 0.24513807892799377, + 0.03512169420719147, + 0.050093647092580795, + 0.015188445337116718, + 0.02384358085691929, + 0.015161002986133099, + 0.0017825711984187365, + 0.0019451980479061604, + 0.0013208199525251985, + 0.002390120876953006, + 0.0014547794125974178, + 0.0018165868241339922, + 0.002836155239492655, + 0.0035425936803221703, + 0.002482346724718809, + 0.0012807773891836405, + 0.0015758981462568045, + 0.0013717305846512318, + 0.0009483235771767795, + 0.0010092342272400856, + 0.0009887657361105084, + 0.001329965190961957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.0924711599946022 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215306878089905 + ], + "opposite_logprob_tokens": [ + -1.5021530389785767 + ], + "logprob_reward": -0.25215306878089905, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215306878089905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1347280591726303, + 0.17580664157867432, + 0.03847662732005119, + 0.06401005387306213, + 0.04629361256957054, + 0.05204489454627037, + 0.12565331161022186, + 0.329685777425766, + 0.02224281243979931, + 0.011058148927986622 + ], + "credit_map": [ + 0.008317524567246437, + 0.007836051285266876, + 0.005673477426171303, + 0.0064889187924563885, + 0.04606824368238449, + 0.007537556812167168, + 0.009631662629544735, + 0.04317464306950569, + 0.012411183677613735, + 0.01329726167023182, + 0.0316605269908905, + 0.028483884409070015, + 0.02341925911605358, + 0.020139295607805252, + 0.020160827785730362, + 0.026234418153762817, + 0.003917199559509754, + 0.0016176800709217787, + 0.00882050208747387, + 0.003964994568377733, + 0.0093262093141675, + 0.0046577733010053635, + 0.0027694543823599815, + 0.0034028186928480864, + 0.00870743952691555, + 0.004342097323387861, + 0.011647526174783707, + 0.007009940221905708, + 0.004942742176353931, + 0.00866701826453209, + 0.00719613628461957, + 0.011497166939079762, + 0.009346717037260532, + 0.008964398875832558, + 0.006421097554266453, + 0.005806674715131521, + 0.0035261246375739574, + 0.004021631553769112, + 0.004456354305148125, + 0.003750619012862444, + 0.006405578926205635, + 0.01237552985548973, + 0.018096087500452995, + 0.004223570693284273, + 0.0037048663944005966, + 0.0027019083499908447, + 0.0024340758100152016, + 0.002103285863995552, + 0.022499093785881996, + 0.020522110164165497, + 0.014985260553658009, + 0.014372842386364937, + 0.014910426922142506, + 0.015512569807469845, + 0.010578783228993416, + 0.01227224338799715, + 0.021181153133511543, + 0.1115623265504837, + 0.1258043795824051, + 0.018275732174515724, + 0.025765052065253258, + 0.007504073437303305, + 0.011985652148723602, + 0.007607456762343645, + 0.0026171773206442595, + 0.0027243108488619328, + 0.0021971690002828836, + 0.003955444321036339, + 0.002098564524203539, + 0.0014265900244936347, + 0.002920962404459715, + 0.004302597139030695, + 0.0017762851202860475, + 0.0014827315462753177, + 0.002366321627050638, + 0.0012728534638881683, + 0.0007989960140548646, + 0.0008461432298645377, + 0.0010021268390119076, + 0.0015126920770853758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.25215306878089905 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009284476400353014 + ], + "opposite_logprob_tokens": [ + -7.0009284019470215 + ], + "logprob_reward": -0.0009284476400353014, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009284476400353014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010846599005162716, + 0.036581650376319885, + 0.011265624314546585, + 0.01166734378784895, + 0.020413914695382118, + 0.1335078775882721, + 0.04110688716173172, + 0.7169626355171204, + 0.0102645643055439, + 0.007382852956652641 + ], + "credit_map": [ + 0.0010121039813384414, + 0.0010367088252678514, + 0.0009694088366813958, + 0.0010201578261330724, + 0.0025625405833125114, + 0.0014390331925824285, + 0.0010092598386108875, + 0.001797387609258294, + 0.003197684418410063, + 0.003517182543873787, + 0.004377174656838179, + 0.005354591645300388, + 0.004367150366306305, + 0.006462897639721632, + 0.004191347863525152, + 0.005113624036312103, + 0.0013867623638361692, + 0.000482298550195992, + 0.0016966366674751043, + 0.0011686582583934069, + 0.002812655409798026, + 0.0016442913329228759, + 0.0008408973808400333, + 0.0012334248749539256, + 0.0018713896861299872, + 0.0007931163418106735, + 0.0018791643669828773, + 0.0015822664136067033, + 0.0011283895000815392, + 0.001522826380096376, + 0.0015674177557229996, + 0.0013227744493633509, + 0.005432453006505966, + 0.005244328640401363, + 0.0030084780883044004, + 0.0017451654421165586, + 0.0015488005010411143, + 0.0012403444852679968, + 0.001327225472778082, + 0.0008671219111420214, + 0.018273433670401573, + 0.0229618139564991, + 0.04931177943944931, + 0.012501897290349007, + 0.00904812291264534, + 0.007563529536128044, + 0.00838929321616888, + 0.005458020139485598, + 0.00761889573186636, + 0.007219342514872551, + 0.005934105254709721, + 0.004965519532561302, + 0.003909476567059755, + 0.004730285611003637, + 0.0037590593565255404, + 0.0029702060855925083, + 0.04617682471871376, + 0.24306735396385193, + 0.27432647347450256, + 0.03953586146235466, + 0.055767785757780075, + 0.016801336780190468, + 0.02571832202374935, + 0.01556877326220274, + 0.0011870552552863955, + 0.00108919118065387, + 0.0010621801484376192, + 0.002303621033206582, + 0.0009833135409280658, + 0.0007901987410150468, + 0.0018033665837720037, + 0.0010456390446051955, + 0.0006477212882600725, + 0.0011467095464468002, + 0.0005921084084548056, + 0.001338317058980465, + 0.0007131485617719591, + 0.0007039716001600027, + 0.0009934711270034313, + 0.0012474057730287313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.0009284476400353014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541710399091244 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541710399091244, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541710399091244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09368575364351273, + 0.04965386912226677, + 0.012219131924211979, + 0.0316951684653759, + 0.04888082668185234, + 0.09431912750005722, + 0.13275186717510223, + 0.5091267824172974, + 0.017187006771564484, + 0.010480441153049469 + ], + "credit_map": [ + 0.004742259159684181, + 0.005211548879742622, + 0.003435059217736125, + 0.005907508544623852, + 0.033569443970918655, + 0.0064769042655825615, + 0.006522699259221554, + 0.02782033383846283, + 0.0044210972264409065, + 0.004873008467257023, + 0.006790005601942539, + 0.006483472418040037, + 0.005219978280365467, + 0.007320556789636612, + 0.006633011158555746, + 0.007912745699286461, + 0.0022498194593936205, + 0.0005786754772998393, + 0.0012683813693001866, + 0.0012628774857148528, + 0.003052940359339118, + 0.0016428485978394747, + 0.0010132560273632407, + 0.0011503330897539854, + 0.005320478230714798, + 0.001443701796233654, + 0.0046425615437328815, + 0.00305706518702209, + 0.002027245471253991, + 0.0061980658210814, + 0.002509078476577997, + 0.006496972404420376, + 0.01205650344491005, + 0.01178863923996687, + 0.007976917549967766, + 0.003994648810476065, + 0.00374646601267159, + 0.0034287574235349894, + 0.0031383284367620945, + 0.0027505706530064344, + 0.012730042450129986, + 0.01917443238198757, + 0.03461572527885437, + 0.007714631501585245, + 0.006592699326574802, + 0.005560867488384247, + 0.004116223659366369, + 0.003814514959231019, + 0.022802863270044327, + 0.023327672854065895, + 0.016746018081903458, + 0.017071392387151718, + 0.015916401520371437, + 0.015117586590349674, + 0.01093159057199955, + 0.01083835493773222, + 0.03281564638018608, + 0.17210721969604492, + 0.19377034902572632, + 0.028251631185412407, + 0.03958319127559662, + 0.011806134134531021, + 0.01901312731206417, + 0.01177951879799366, + 0.00304790074005723, + 0.0018743504770100117, + 0.0017742190975695848, + 0.0027146528009325266, + 0.001258000498637557, + 0.00121261493768543, + 0.0019001034088432789, + 0.0034051663242280483, + 0.0014586604665964842, + 0.0011451252503320575, + 0.0015195159940049052, + 0.0014307487290352583, + 0.0011303747305646539, + 0.0010960582876577973, + 0.0011344049125909805, + 0.001565554179251194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.012541710399091244 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646263778209686 + ], + "opposite_logprob_tokens": [ + -3.0486462116241455 + ], + "logprob_reward": -0.048646263778209686, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646263778209686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.00656643882393837, + 0.010596560314297676, + 0.0074292561039328575, + 0.008230968378484249, + 0.009199719876050949, + 0.040977030992507935, + 0.067782461643219, + 0.8304071426391602, + 0.009990115649998188, + 0.008820295333862305 + ], + "credit_map": [ + 0.0005317851318977773, + 0.0008671483374200761, + 0.000792359234765172, + 0.000764230266213417, + 0.001070889295078814, + 0.0009324687998741865, + 0.00063178880373016, + 0.0009757686639204621, + 0.0007912252331152558, + 0.0009718305082060397, + 0.0014693897683173418, + 0.001664613839238882, + 0.0012563925702124834, + 0.0017953633796423674, + 0.0011784143280237913, + 0.0014693309785798192, + 0.0016372088575735688, + 0.00026513979537412524, + 0.0019985053222626448, + 0.0007544538821093738, + 0.0006582910427823663, + 0.0008136849501170218, + 0.0005457816878333688, + 0.0007561907405033708, + 0.0013401882024481893, + 0.0006077675498090684, + 0.001016119378618896, + 0.0013610543683171272, + 0.0006931836833246052, + 0.0014673278201371431, + 0.0008008881704881787, + 0.0009444400784559548, + 0.0021034267265349627, + 0.001992936013266444, + 0.0013315019896253943, + 0.0010049109114333987, + 0.0006012977100908756, + 0.0008255508146248758, + 0.0008030034368857741, + 0.0005370923900045455, + 0.006001511123031378, + 0.006567684002220631, + 0.015917422249913216, + 0.003331520827487111, + 0.0024398898240178823, + 0.0028243272099643946, + 0.0020097000524401665, + 0.001884977100417018, + 0.010810772888362408, + 0.01262135524302721, + 0.009367699734866619, + 0.008046764880418777, + 0.007397532928735018, + 0.008174651302397251, + 0.005930580198764801, + 0.00543310446664691, + 0.053112562745809555, + 0.28163525462150574, + 0.3178880512714386, + 0.04544604569673538, + 0.06459992378950119, + 0.019902316853404045, + 0.0297409500926733, + 0.018082095310091972, + 0.00109073705971241, + 0.001169326831586659, + 0.000912586459890008, + 0.002090664580464363, + 0.0009951293468475342, + 0.0006333638448268175, + 0.001689238240942359, + 0.00140906882006675, + 0.0009571143891662359, + 0.001292409491725266, + 0.0008391322335228324, + 0.0014714299468323588, + 0.0009610329871065915, + 0.0007920754724182189, + 0.0013014836004003882, + 0.0012056170962750912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.048646263778209686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a moment of quiet discovery in a dense bamboo forest. Dappled sunlight filters through the tall canopy, creating soft, shifting patterns of light and shadow. The furry head of a giant panda cautiously emerges from behind a thick screen of vibrant green bamboo stems and leaves. Its fur is a dense, perfect white, each hair catching the soft light and contrasting sharply with the signature, soft-rimmed black patch surrounding its visible eye. From within that field, a single dark, gleaming, inquisitive eye peers out, its gaze gentle and inquisitive. The panda's head tilts slightly, slowly and deliberately advancing, causing several delicate bamboo clumps to part, its black nose twitching subtly. The entire action is executed with extreme caution, creating an atmosphere of tranquility and intimacy, as if the viewer has stumbled upon a secret, private moment deep within the heart of nature.", + "seed": 30411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_6/index_5733_seed_30411_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": -0.02840598749462515, + "vlm_logprob": -0.02840598749462515, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.02840598749462515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06426741182804108 + ], + "opposite_logprob_tokens": [ + -3.31426739692688 + ], + "logprob_reward": -0.06426741182804108, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06426741182804108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22254236042499542, + 0.038431260734796524, + 0.028180653229355812, + 0.14300736784934998, + 0.03091917186975479, + 0.10794433206319809, + 0.03372269496321678, + 0.04166344553232193, + 0.024648673832416534, + 0.3289400041103363 + ], + "credit_map": [ + 0.02257687598466873, + 0.029607847332954407, + 0.02138923667371273, + 0.018153401091694832, + 0.025534547865390778, + 0.04035386070609093, + 0.030990218743681908, + 0.03393637388944626, + 0.004403618164360523, + 0.005235760938376188, + 0.00432882085442543, + 0.004726969171315432, + 0.005113593302667141, + 0.005414864514023066, + 0.005327539052814245, + 0.0038800945039838552, + 0.004348601680248976, + 0.004296999890357256, + 0.005033890716731548, + 0.0034982559736818075, + 0.002640988677740097, + 0.002384863793849945, + 0.003080085851252079, + 0.002896965481340885, + 0.011122694239020348, + 0.04795677214860916, + 0.00924516562372446, + 0.01200802717357874, + 0.011577548459172249, + 0.0062139080837368965, + 0.010909685865044594, + 0.0339735671877861, + 0.0041242195293307304, + 0.0018372624181210995, + 0.005196135491132736, + 0.0038107354193925858, + 0.002232802100479603, + 0.0040704174898564816, + 0.004035680089145899, + 0.005611918866634369, + 0.016265423968434334, + 0.009778426960110664, + 0.021854186430573463, + 0.01016484946012497, + 0.01048403512686491, + 0.011560879647731781, + 0.011550668627023697, + 0.016285866498947144, + 0.005016989540308714, + 0.002975749084725976, + 0.006327340845018625, + 0.0038354666903615, + 0.0029449528083205223, + 0.0033253636211156845, + 0.005256341770291328, + 0.004040492232888937, + 0.00648991484194994, + 0.0023926745634526014, + 0.012399896048009396, + 0.004419329110532999, + 0.00294469203799963, + 0.003729314776137471, + 0.004794818349182606, + 0.004492803476750851, + 0.003210261696949601, + 0.002115236595273018, + 0.005438224412500858, + 0.003046430181711912, + 0.0018826051382347941, + 0.0024146465584635735, + 0.0031407487113028765, + 0.003400521818548441, + 0.060764748603105545, + 0.020833434537053108, + 0.08327391743659973, + 0.04413890466094017, + 0.02143266052007675, + 0.017331665381789207, + 0.04010195657610893, + 0.04106268659234047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.06426741182804108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337129682302475 + ], + "opposite_logprob_tokens": [ + -3.408712863922119 + ], + "logprob_reward": -0.0337129682302475, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337129682302475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38438406586647034, + 0.06324977427721024, + 0.05719032883644104, + 0.1524566262960434, + 0.014191756024956703, + 0.02223333716392517, + 0.05431170016527176, + 0.07998106628656387, + 0.054103951901197433, + 0.1178974062204361 + ], + "credit_map": [ + 0.04422292858362198, + 0.05499211326241493, + 0.039877504110336304, + 0.03186400979757309, + 0.039243634790182114, + 0.06700266897678375, + 0.04248999431729317, + 0.0646912083029747, + 0.010839703492820263, + 0.011257878504693508, + 0.009837396442890167, + 0.006073596887290478, + 0.0054441094398498535, + 0.005298590753227472, + 0.005654269363731146, + 0.008844230324029922, + 0.005926790181547403, + 0.02053164131939411, + 0.010610321536660194, + 0.005300750490278006, + 0.004355372861027718, + 0.0035718856379389763, + 0.0032492612954229116, + 0.00364431063644588, + 0.019460005685687065, + 0.026438239961862564, + 0.03753812238574028, + 0.0093874866142869, + 0.011455584317445755, + 0.00790814682841301, + 0.012516758404672146, + 0.027752269059419632, + 0.0019865469075739384, + 0.0011163700837641954, + 0.0021518429275602102, + 0.0021933596581220627, + 0.0013136041816323996, + 0.0015953117981553078, + 0.002057422185316682, + 0.00177729781717062, + 0.003840845311060548, + 0.001860939315520227, + 0.006044796667993069, + 0.002014391589909792, + 0.0019798274151980877, + 0.0019533392041921616, + 0.0025320255663245916, + 0.002007171045988798, + 0.00835177768021822, + 0.005340344272553921, + 0.00949955452233553, + 0.006432457827031612, + 0.004857515450567007, + 0.006496536545455456, + 0.00675181532278657, + 0.006581698544323444, + 0.015616402961313725, + 0.0033512867521494627, + 0.034217339009046555, + 0.005861198995262384, + 0.002418752759695053, + 0.004693995229899883, + 0.005662267096340656, + 0.008159819059073925, + 0.00950417947024107, + 0.0045110490173101425, + 0.008336062543094158, + 0.005751060321927071, + 0.006017128471285105, + 0.00397578626871109, + 0.00816925335675478, + 0.007839435711503029, + 0.018748944625258446, + 0.006586508825421333, + 0.0311457309871912, + 0.012125186622142792, + 0.008412769064307213, + 0.005609537940472364, + 0.019124876707792282, + 0.016143841668963432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0337129682302475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001718236249871552 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001718236249871552, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001718236249871552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18481943011283875, + 0.021798087283968925, + 0.04682368412613869, + 0.11210381239652634, + 0.02753383480012417, + 0.017890844494104385, + 0.01800602488219738, + 0.05306277424097061, + 0.030248230323195457, + 0.48771336674690247 + ], + "credit_map": [ + 0.015414436347782612, + 0.024056129157543182, + 0.01665031537413597, + 0.0155131621286273, + 0.02256263792514801, + 0.03649377450346947, + 0.024664444848895073, + 0.029464509338140488, + 0.0040938532911241055, + 0.00248720939271152, + 0.003771263873204589, + 0.0025038972962647676, + 0.002087065251544118, + 0.002297068713232875, + 0.0026495461352169514, + 0.0019081808859482408, + 0.00456949882209301, + 0.016870230436325073, + 0.010851922444999218, + 0.003351782914251089, + 0.003823277074843645, + 0.002760733012109995, + 0.0018959339940920472, + 0.002700300421565771, + 0.01474239956587553, + 0.005787733010947704, + 0.029530320316553116, + 0.007026623468846083, + 0.006557782180607319, + 0.006000167224556208, + 0.011188059113919735, + 0.03127071261405945, + 0.004128531087189913, + 0.0026951327454298735, + 0.00325378798879683, + 0.005686038639396429, + 0.0030299911741167307, + 0.002491022227331996, + 0.003951880615204573, + 0.0022974470630288124, + 0.0027320070657879114, + 0.0016486338572576642, + 0.002034897683188319, + 0.0018472227966412902, + 0.0015938583528622985, + 0.0023923986591398716, + 0.0026882928796112537, + 0.0029535305220633745, + 0.002890209900215268, + 0.0015385299921035767, + 0.002473822794854641, + 0.002286000642925501, + 0.0018176144221797585, + 0.0018397956155240536, + 0.002570193726569414, + 0.002589855808764696, + 0.009975116699934006, + 0.002750340150669217, + 0.01936899498105049, + 0.0040632011368870735, + 0.002928931498900056, + 0.0040336898528039455, + 0.004237735643982887, + 0.005704755894839764, + 0.00473268236964941, + 0.0029115055222064257, + 0.004870726261287928, + 0.0048826830461621284, + 0.0025398791767656803, + 0.00235665962100029, + 0.003120291978120804, + 0.004833798389881849, + 0.06664256751537323, + 0.02826102077960968, + 0.11358343809843063, + 0.04721371456980705, + 0.03668953850865364, + 0.022623497992753983, + 0.10786090046167374, + 0.06483865529298782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.001718236249871552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638868108391762 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638868108391762, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638868108391762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12268561869859695, + 0.037686578929424286, + 0.052569493651390076, + 0.3638778030872345, + 0.04716238006949425, + 0.06784889847040176, + 0.020101988688111305, + 0.12683233618736267, + 0.058273933827877045, + 0.10296093672513962 + ], + "credit_map": [ + 0.017867300659418106, + 0.017397601157426834, + 0.012845730409026146, + 0.010653099045157433, + 0.014627140015363693, + 0.0186420027166605, + 0.01779811643064022, + 0.012854632921516895, + 0.004385202191770077, + 0.005065515637397766, + 0.004468526691198349, + 0.005564478691667318, + 0.005498024169355631, + 0.0035369526594877243, + 0.00456698052585125, + 0.004600898362696171, + 0.004517013672739267, + 0.016024228185415268, + 0.010448984801769257, + 0.0036210310645401478, + 0.005313095171004534, + 0.0033090070355683565, + 0.003895985661074519, + 0.00544014573097229, + 0.04416390135884285, + 0.019916731864213943, + 0.08365333080291748, + 0.020922644063830376, + 0.01781410165131092, + 0.019652925431728363, + 0.0384085550904274, + 0.11934563517570496, + 0.008900594897568226, + 0.0028057238087058067, + 0.005286982748657465, + 0.004158610012382269, + 0.0032168724574148655, + 0.006445962470024824, + 0.006879307795315981, + 0.0094683226197958, + 0.010102005675435066, + 0.005733712576329708, + 0.015257498249411583, + 0.005654448177665472, + 0.006233950611203909, + 0.006537254899740219, + 0.006760464981198311, + 0.011569562368094921, + 0.003460309933871031, + 0.001241876743733883, + 0.004615061450749636, + 0.0022874732967466116, + 0.0013620340032503009, + 0.0026687306817620993, + 0.002289270516484976, + 0.0021772326435893774, + 0.024561259895563126, + 0.005312055349349976, + 0.0564543753862381, + 0.009308724664151669, + 0.004014290869235992, + 0.007286552339792252, + 0.006724881008267403, + 0.013170209713280201, + 0.009746002033352852, + 0.004536429885774851, + 0.010341626591980457, + 0.009689847938716412, + 0.004574801307171583, + 0.004267028998583555, + 0.0065500554628670216, + 0.008568144403398037, + 0.01402407418936491, + 0.0059837764129042625, + 0.022685209289193153, + 0.011828918009996414, + 0.007360741961747408, + 0.004465405363589525, + 0.024410242214798927, + 0.012202569283545017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.008638868108391762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dense bamboo forest with dappled sunlight filtering through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visible subject the furry head of a giant panda with white fur and a black patch around its eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the panda's head tilt slightly and advance slowly while parting bamboo clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a single dark, gleaming eye visible peering out from behind the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369245305657387 + ], + "opposite_logprob_tokens": [ + -3.4086923599243164 + ], + "logprob_reward": -0.03369245305657387, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369245305657387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1474752575159073, + 0.031592968851327896, + 0.025612400844693184, + 0.16921861469745636, + 0.018283290788531303, + 0.056602220982313156, + 0.031106896698474884, + 0.07180081307888031, + 0.040381744503974915, + 0.4079258143901825 + ], + "credit_map": [ + 0.01623867265880108, + 0.019081654027104378, + 0.014934125356376171, + 0.013767940923571587, + 0.020166000351309776, + 0.02407407946884632, + 0.02168133296072483, + 0.01753145083785057, + 0.0028190428856760263, + 0.004726705607026815, + 0.002551419660449028, + 0.0029931056778877974, + 0.004953986499458551, + 0.0038199026603251696, + 0.005158442072570324, + 0.004570363089442253, + 0.0031093896832317114, + 0.006123596802353859, + 0.002016881015151739, + 0.004587563686072826, + 0.0024776458740234375, + 0.00173176487442106, + 0.0032333123963326216, + 0.0023322429042309523, + 0.021994614973664284, + 0.007148320320993662, + 0.041626304388046265, + 0.009731565602123737, + 0.009793814271688461, + 0.009394347667694092, + 0.01671745255589485, + 0.05281217023730278, + 0.002883664797991514, + 0.0016860002651810646, + 0.002662907587364316, + 0.0025678868405520916, + 0.001900963019579649, + 0.002191064180806279, + 0.00265955226495862, + 0.0017312528798356652, + 0.008195276372134686, + 0.005617261398583651, + 0.01051287166774273, + 0.005565676372498274, + 0.005856624338775873, + 0.005641716532409191, + 0.006064774002879858, + 0.009148018434643745, + 0.005227130372077227, + 0.0018771751783788204, + 0.007423213217407465, + 0.002646686742082238, + 0.0020284669008105993, + 0.004838148597627878, + 0.0037664950359612703, + 0.0032995790243148804, + 0.013978955335915089, + 0.0029903240501880646, + 0.03250822424888611, + 0.005061815492808819, + 0.0020802500657737255, + 0.0036069764755666256, + 0.0039741275832057, + 0.0076001365669071674, + 0.005140326917171478, + 0.003778900718316436, + 0.0073330155573785305, + 0.004869269207119942, + 0.003471903968602419, + 0.004425039980560541, + 0.006760639604181051, + 0.0046026441268622875, + 0.05582890287041664, + 0.02478792890906334, + 0.09566012769937515, + 0.04032184183597565, + 0.0302903912961483, + 0.019460005685687065, + 0.09170378744602203, + 0.049872804433107376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.03369245305657387 + } + ], + "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/0000052/group_7/index_tensor([885, 885])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_tensor([885, 885])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e87fd3c3084e71fb3056b62788b9cc332938b2f4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_tensor([885, 885])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "glossy, eye-level camera relentlessly follows a large group of marathon runners, capturing the scene in stunning, surreal detail. The camera moves at their pace, placing a few distinct runners in the foreground with razor-sharp focus. A woman with a high, swinging ponytail, rosy-faced and focused, wears a bright yellow technical vest drenched in sweat. Beside her, a man with a shaved head and a focused expression, every tense muscle fiber in his calves and thighs visible with each powerful stride. Their outfits are a vibrant, chaotic mosaic of neon green, electric blue, and hot pink, all rendered with realistic textures. Their collective action is a thunderous, rhythmic thud against the asphalt of the broad city avenue, a unified forward wave, thousands of individual bodies moving as one undulating mass. Crisp, low-angle morning sunlight filters through the scene, gleaming on their foreheads and casting long, dynamic shadows as they run. In the background, the blurred forms of cheering spectators and city buildings streak by, enhancing the profound sense of speed and unstoppable momentum, creating an epic atmosphere rich in raw determination and human endurance.", + "seed": 21556, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_885_seed_21556_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": -0.21637061722576617, + "vlm_logprob": -0.21637061722576617, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.21637061722576617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.2523123025894165 + ], + "opposite_logprob_tokens": [ + -1.8773123025894165 + ], + "logprob_reward": -0.2523123025894165, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.2523123025894165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2548491060733795, + 0.12036414444446564, + 0.10949493199586868, + 0.0381256639957428, + 0.1794285774230957, + 0.04988061636686325, + 0.06340155005455017, + 0.048733022063970566, + 0.10259978473186493, + 0.03312259912490845 + ], + "credit_map": [ + 0.021531768143177032, + 0.11059222370386124, + 0.008393528871238232, + 0.020111074671149254, + 0.032835911959409714, + 0.011457828804850578, + 0.04265588894486427, + 0.007270880974829197, + 0.0126273138448596, + 0.04563964158296585, + 0.005229227244853973, + 0.007193738128989935, + 0.011747167445719242, + 0.006067199166864157, + 0.02774789184331894, + 0.004111974034458399, + 0.011144678108394146, + 0.01750756986439228, + 0.01683429442346096, + 0.009260767139494419, + 0.023623140528798103, + 0.0113974092528224, + 0.01311960257589817, + 0.006607477553188801, + 0.0038942371029406786, + 0.00748624000698328, + 0.004621419589966536, + 0.0031988814007490873, + 0.0049504125490784645, + 0.0041627599857747555, + 0.004186322446912527, + 0.005625393707305193, + 0.0175846628844738, + 0.01680232398211956, + 0.01425995770841837, + 0.0187198705971241, + 0.021241948008537292, + 0.012516354210674763, + 0.04075505957007408, + 0.03754841163754463, + 0.004920311737805605, + 0.0072259316220879555, + 0.007150530349463224, + 0.007006700616329908, + 0.005756196100264788, + 0.003656101180240512, + 0.004959520418196917, + 0.009205327369272709, + 0.007489966228604317, + 0.0050322264432907104, + 0.005616509355604649, + 0.007331485860049725, + 0.006936783902347088, + 0.00795907061547041, + 0.011165942996740341, + 0.01186957024037838, + 0.005317391362041235, + 0.005856451578438282, + 0.004921470303088427, + 0.008480822667479515, + 0.008265111595392227, + 0.004697768483310938, + 0.006049553398042917, + 0.005144457332789898, + 0.009901704266667366, + 0.009830079041421413, + 0.008276671171188354, + 0.0346326045691967, + 0.01054154708981514, + 0.007272902876138687, + 0.013375706039369106, + 0.008768577128648758, + 0.002968656364828348, + 0.004255420062690973, + 0.002272704616189003, + 0.007140998728573322, + 0.0026501675602048635, + 0.003343063872307539, + 0.008541136980056763, + 0.0019504564115777612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.2523123025894165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.02638973481953144 + ], + "opposite_logprob_tokens": [ + -3.6513898372650146 + ], + "logprob_reward": -0.02638973481953144, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638973481953144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29422691464424133, + 0.09469973295927048, + 0.0763765275478363, + 0.06194261834025383, + 0.20619027316570282, + 0.05519350618124008, + 0.060073766857385635, + 0.0402354970574379, + 0.0743943452835083, + 0.03666684031486511 + ], + "credit_map": [ + 0.02761939913034439, + 0.02754337154328823, + 0.019815271720290184, + 0.03517383709549904, + 0.05283238738775253, + 0.022732984274625778, + 0.08895063400268555, + 0.019559014588594437, + 0.009095815941691399, + 0.029528886079788208, + 0.006629097741097212, + 0.009147821925580502, + 0.009634474292397499, + 0.00784924067556858, + 0.01768842153251171, + 0.005125976167619228, + 0.007896477356553078, + 0.010853090323507786, + 0.009855513460934162, + 0.005952634382992983, + 0.015386142767965794, + 0.008145179599523544, + 0.01277527492493391, + 0.005512219853699207, + 0.00423763133585453, + 0.007738240994513035, + 0.004638415761291981, + 0.007450606673955917, + 0.011623458005487919, + 0.007931129075586796, + 0.010235950350761414, + 0.008087188005447388, + 0.019851263612508774, + 0.02041189931333065, + 0.013946166262030602, + 0.02198108099400997, + 0.0217853095382452, + 0.01318396721035242, + 0.04961182549595833, + 0.04541876167058945, + 0.00560782290995121, + 0.008145257830619812, + 0.0060655660927295685, + 0.008394144475460052, + 0.005920335650444031, + 0.005707225762307644, + 0.006474899593740702, + 0.008878253400325775, + 0.005793096963316202, + 0.008774424903094769, + 0.004562045447528362, + 0.008182302117347717, + 0.013878473080694675, + 0.007454758509993553, + 0.005630855914205313, + 0.0057978094555437565, + 0.004446710459887981, + 0.005034693516790867, + 0.0037858858704566956, + 0.0052547762170434, + 0.007908795028924942, + 0.005319016519933939, + 0.0043791052885353565, + 0.0041065155528485775, + 0.008555514737963676, + 0.006361182779073715, + 0.007304705213755369, + 0.020490722730755806, + 0.010292277671396732, + 0.008085167035460472, + 0.00761705357581377, + 0.0056877294555306435, + 0.0024432677309960127, + 0.005232477560639381, + 0.0017665467457845807, + 0.00798081886023283, + 0.0035038485657423735, + 0.0037042032927274704, + 0.008908014744520187, + 0.0031276613008230925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02638973481953144 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.1003272533416748 + ], + "opposite_logprob_tokens": [ + -2.350327253341675 + ], + "logprob_reward": -0.1003272533416748, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003272533416748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36579787731170654, + 0.045405760407447815, + 0.06541203707456589, + 0.02042439393699169, + 0.053250495344400406, + 0.0892825648188591, + 0.06953693926334381, + 0.08147559314966202, + 0.1613355129957199, + 0.04807886481285095 + ], + "credit_map": [ + 0.04011177644133568, + 0.13630467653274536, + 0.016773132607340813, + 0.034002888947725296, + 0.08138296008110046, + 0.017705881968140602, + 0.02881472185254097, + 0.010701836086809635, + 0.004294389858841896, + 0.010438352823257446, + 0.003256096038967371, + 0.0058637396432459354, + 0.007962260395288467, + 0.004298948682844639, + 0.006620114203542471, + 0.0026718585286289454, + 0.006941480096429586, + 0.007813719101250172, + 0.008524387143552303, + 0.006494061090052128, + 0.013626987114548683, + 0.005007029976695776, + 0.012149360962212086, + 0.004855011124163866, + 0.0022525000385940075, + 0.0037339471746236086, + 0.0027187790255993605, + 0.0027134770061820745, + 0.002728768391534686, + 0.0023003884125500917, + 0.002129714237526059, + 0.0018468188354745507, + 0.007000783924013376, + 0.007149890065193176, + 0.0047356500290334225, + 0.006301417015492916, + 0.007411211729049683, + 0.004538260865956545, + 0.009478467516601086, + 0.006634813733398914, + 0.006801602430641651, + 0.011556519195437431, + 0.008065182715654373, + 0.024545270949602127, + 0.010047449730336666, + 0.006984727457165718, + 0.011626877821981907, + 0.009654935449361801, + 0.006357274949550629, + 0.00825798511505127, + 0.0052807386964559555, + 0.011757098138332367, + 0.01415807195007801, + 0.00590459955856204, + 0.008918046951293945, + 0.008903124369680882, + 0.009021677076816559, + 0.010526991449296474, + 0.0065792882815003395, + 0.014606528915464878, + 0.01683732122182846, + 0.00877340603619814, + 0.007982623763382435, + 0.007147754542529583, + 0.019017212092876434, + 0.015084460377693176, + 0.009381440468132496, + 0.0465097650885582, + 0.02838602475821972, + 0.015235550701618195, + 0.017738420516252518, + 0.009982647374272346, + 0.0038574908394366503, + 0.011854567565023899, + 0.003165914211422205, + 0.0060080187395215034, + 0.005294217728078365, + 0.005284943617880344, + 0.00866337213665247, + 0.00395033648237586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.1003272533416748 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.07002804428339005 + ], + "opposite_logprob_tokens": [ + -2.695028066635132 + ], + "logprob_reward": -0.07002804428339005, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002804428339005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1588735580444336, + 0.09050621092319489, + 0.11502066999673843, + 0.030831733718514442, + 0.2433631867170334, + 0.05404605343937874, + 0.03349470719695091, + 0.02833828516304493, + 0.09548772126436234, + 0.15003788471221924 + ], + "credit_map": [ + 0.01934089884161949, + 0.028562750667333603, + 0.02208997868001461, + 0.019742274656891823, + 0.025270609185099602, + 0.013322942890226841, + 0.015933234244585037, + 0.014610856771469116, + 0.006110406015068293, + 0.03286970406770706, + 0.005149614065885544, + 0.004962858743965626, + 0.007798631675541401, + 0.004440218675881624, + 0.02459905482828617, + 0.004575722850859165, + 0.011334022507071495, + 0.016236962750554085, + 0.017940636724233627, + 0.014367993921041489, + 0.019698649644851685, + 0.012535693123936653, + 0.013058435171842575, + 0.009848286397755146, + 0.0037587322294712067, + 0.00549115426838398, + 0.005188371054828167, + 0.0031827609054744244, + 0.004063823726028204, + 0.0030057772528380156, + 0.003269820474088192, + 0.002871295204386115, + 0.02657606452703476, + 0.027824096381664276, + 0.02096969448029995, + 0.026894526556134224, + 0.024746425449848175, + 0.01720387302339077, + 0.05232753977179527, + 0.04682095721364021, + 0.006036822218447924, + 0.009075971320271492, + 0.006814961321651936, + 0.006182834506034851, + 0.008420204743742943, + 0.004674904048442841, + 0.006556137930601835, + 0.006284219212830067, + 0.004074258264154196, + 0.004805153701454401, + 0.006071631330996752, + 0.0053408523090183735, + 0.0030863622669130564, + 0.00422480795532465, + 0.0030830148607492447, + 0.0028086293023079634, + 0.0031580536160618067, + 0.003186767688021064, + 0.0026892991736531258, + 0.006089572329074144, + 0.005183547269552946, + 0.00268750311806798, + 0.0030717519111931324, + 0.0022717893589287996, + 0.013163499534130096, + 0.010224045254290104, + 0.013966187834739685, + 0.015997599810361862, + 0.012288357131183147, + 0.01184121984988451, + 0.00945174228399992, + 0.008555075153708458, + 0.01014174334704876, + 0.013931652531027794, + 0.008776921778917313, + 0.02802053466439247, + 0.01614239811897278, + 0.011805783957242966, + 0.05043444037437439, + 0.010784411802887917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.07002804428339005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.6327957510948181 + ], + "opposite_logprob_tokens": [ + -0.7577957510948181 + ], + "logprob_reward": -0.6327957510948181, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327957510948181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5018558502197266, + 0.11528825014829636, + 0.0460648313164711, + 0.019281893968582153, + 0.0906749963760376, + 0.031065350398421288, + 0.03144115209579468, + 0.015621368773281574, + 0.030827797949314117, + 0.11787852644920349 + ], + "credit_map": [ + 0.04396090283989906, + 0.2190459817647934, + 0.018292229622602463, + 0.02733328752219677, + 0.08346740901470184, + 0.021538160741329193, + 0.07593081146478653, + 0.012287057004868984, + 0.011365511454641819, + 0.029343200847506523, + 0.007829858921468258, + 0.01424532663077116, + 0.016552217304706573, + 0.012109886854887009, + 0.015264753252267838, + 0.008577493019402027, + 0.0035423741210252047, + 0.008024352602660656, + 0.0035094679333269596, + 0.008067948743700981, + 0.00799388624727726, + 0.004803174175322056, + 0.005777837708592415, + 0.0043457867577672005, + 0.0018051941879093647, + 0.002860746346414089, + 0.002871323609724641, + 0.002841664245352149, + 0.002186766592785716, + 0.0019083003280684352, + 0.0024454831145703793, + 0.0023624151945114136, + 0.00815808866173029, + 0.008560345508158207, + 0.006776818074285984, + 0.007257618010044098, + 0.008873678743839264, + 0.006888727191835642, + 0.023903025314211845, + 0.02025669999420643, + 0.0028935777954757214, + 0.005479507613927126, + 0.0027129650115966797, + 0.0056779272854328156, + 0.0038763522170484066, + 0.0024766544811427593, + 0.0039011978078633547, + 0.004047167021781206, + 0.0039010681211948395, + 0.00437117600813508, + 0.0036014269571751356, + 0.004489455372095108, + 0.0037864253390580416, + 0.00398529227823019, + 0.0031466917134821415, + 0.004159615840762854, + 0.001642212737351656, + 0.0018536311108618975, + 0.0020792153663933277, + 0.0027384147979319096, + 0.0017749723047018051, + 0.0019265520386397839, + 0.0017795070307329297, + 0.001826863968744874, + 0.0029257964342832565, + 0.0034250174649059772, + 0.002878765342757106, + 0.008904348127543926, + 0.003561619436368346, + 0.002962610451504588, + 0.0035976606886833906, + 0.0025719807017594576, + 0.0077690002508461475, + 0.013417313806712627, + 0.007600588258355856, + 0.020338155329227448, + 0.010977525264024734, + 0.012353957630693913, + 0.03733343631029129, + 0.008088551461696625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.6327957510948181 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "glossy, eye-level camera relentlessly follows a large group of marathon runners, capturing the scene in stunning, surreal detail. The camera moves at their pace, placing a few distinct runners in the foreground with razor-sharp focus. A woman with a high, swinging ponytail, rosy-faced and focused, wears a bright yellow technical vest drenched in sweat. Beside her, a man with a shaved head and a focused expression, every tense muscle fiber in his calves and thighs visible with each powerful stride. Their outfits are a vibrant, chaotic mosaic of neon green, electric blue, and hot pink, all rendered with realistic textures. Their collective action is a thunderous, rhythmic thud against the asphalt of the broad city avenue, a unified forward wave, thousands of individual bodies moving as one undulating mass. Crisp, low-angle morning sunlight filters through the scene, gleaming on their foreheads and casting long, dynamic shadows as they run. In the background, the blurred forms of cheering spectators and city buildings streak by, enhancing the profound sense of speed and unstoppable momentum, creating an epic atmosphere rich in raw determination and human endurance.", + "seed": 21557, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_885_seed_21557_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": -0.21497448850423098, + "vlm_logprob": -0.21497448850423098, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.21497448850423098, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.2844991683959961 + ], + "opposite_logprob_tokens": [ + -1.784499168395996 + ], + "logprob_reward": -0.2844991683959961, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2844991683959961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07004725188016891, + 0.06857958436012268, + 0.08217168599367142, + 0.045260533690452576, + 0.09648345410823822, + 0.27763065695762634, + 0.12447365373373032, + 0.1055564358830452, + 0.07364612072706223, + 0.05615070089697838 + ], + "credit_map": [ + 0.00885461550205946, + 0.010125108994543552, + 0.007141874637454748, + 0.00525806937366724, + 0.009192501194775105, + 0.013137012720108032, + 0.008458809927105904, + 0.007879240438342094, + 0.00995196495205164, + 0.010132546536624432, + 0.0068014622665941715, + 0.00568727171048522, + 0.008274918422102928, + 0.008539635688066483, + 0.010529866442084312, + 0.008661898784339428, + 0.007050713524222374, + 0.007219369988888502, + 0.00712406262755394, + 0.012131229043006897, + 0.01881590113043785, + 0.008919835090637207, + 0.008722392842173576, + 0.012188165448606014, + 0.005276463460177183, + 0.00850517489016056, + 0.002918638987466693, + 0.0051364777609705925, + 0.0087345065549016, + 0.003527306718751788, + 0.006164706777781248, + 0.0049972496926784515, + 0.013021416030824184, + 0.014230839908123016, + 0.013848032802343369, + 0.011516788974404335, + 0.01574242301285267, + 0.013336964882910252, + 0.006458452437072992, + 0.008328515104949474, + 0.022285711020231247, + 0.015683136880397797, + 0.012973200529813766, + 0.028947768732905388, + 0.10890331119298935, + 0.017069216817617416, + 0.011588050052523613, + 0.060180213302373886, + 0.010768968611955643, + 0.03883402794599533, + 0.006890634074807167, + 0.004512988496571779, + 0.04485579952597618, + 0.0058004898019135, + 0.0069798072800040245, + 0.005830902606248856, + 0.009029686450958252, + 0.019243737682700157, + 0.007392466068267822, + 0.009290611371397972, + 0.02551615610718727, + 0.01410206314176321, + 0.012844680808484554, + 0.008137013763189316, + 0.005405711941421032, + 0.004640826024115086, + 0.004509355407208204, + 0.004642777610570192, + 0.018918151035904884, + 0.006271186750382185, + 0.005702140741050243, + 0.02355595864355564, + 0.005948795937001705, + 0.005491924937814474, + 0.009667695499956608, + 0.0032800098415464163, + 0.004417170770466328, + 0.012912226840853691, + 0.0036193884443491697, + 0.010813474655151367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.2844991683959961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.01422515232115984 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.01422515232115984, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422515232115984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09046994149684906, + 0.05032141134142876, + 0.030209582298994064, + 0.033598750829696655, + 0.04146350547671318, + 0.4306703209877014, + 0.0911860316991806, + 0.03180559352040291, + 0.13100096583366394, + 0.06927385926246643 + ], + "credit_map": [ + 0.009810133837163448, + 0.012137474492192268, + 0.011632706969976425, + 0.010169445537030697, + 0.01134051289409399, + 0.013371924869716167, + 0.011383012868463993, + 0.010624729096889496, + 0.005675229709595442, + 0.008475575596094131, + 0.006364106200635433, + 0.005277838092297316, + 0.006734970957040787, + 0.007421204820275307, + 0.006142123602330685, + 0.004230362828820944, + 0.0037907275836914778, + 0.003950031008571386, + 0.003187181893736124, + 0.0033017604146152735, + 0.003437747247517109, + 0.0036253705620765686, + 0.004894541576504707, + 0.004022225737571716, + 0.002712264424189925, + 0.0060642557218670845, + 0.0033764932304620743, + 0.005623307079076767, + 0.004862746689468622, + 0.003101211739704013, + 0.004595673643052578, + 0.0032628008630126715, + 0.004319518338888884, + 0.007075176108628511, + 0.002341488841921091, + 0.0028502903878688812, + 0.008347607217729092, + 0.009779109619557858, + 0.0028996476903557777, + 0.0038506684359163046, + 0.03204295411705971, + 0.02378326654434204, + 0.01839986816048622, + 0.040840260684490204, + 0.17124249041080475, + 0.028935452923178673, + 0.017361387610435486, + 0.09806463867425919, + 0.007070585153996944, + 0.024490835145115852, + 0.006021889392286539, + 0.00505949929356575, + 0.029950499534606934, + 0.004854814615100622, + 0.009758266620337963, + 0.003979647532105446, + 0.004756183363497257, + 0.004651049617677927, + 0.0021281286608427763, + 0.001927490928210318, + 0.007082952652126551, + 0.005321997683495283, + 0.003459915518760681, + 0.002477877074852586, + 0.01099632028490305, + 0.026061134412884712, + 0.012151515111327171, + 0.008414765819907188, + 0.03424123302102089, + 0.012310092337429523, + 0.013170371763408184, + 0.013655547052621841, + 0.0072343009524047375, + 0.0054867141880095005, + 0.004979490768164396, + 0.0046457890421152115, + 0.006697864271700382, + 0.01947450079023838, + 0.003532114438712597, + 0.017223091796040535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.01422515232115984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.04306305572390556 + ], + "opposite_logprob_tokens": [ + -3.168063163757324 + ], + "logprob_reward": -0.04306305572390556, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306305572390556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11418890953063965, + 0.0519646480679512, + 0.09079723805189133, + 0.08316370844841003, + 0.0781099945306778, + 0.271584689617157, + 0.04995004087686539, + 0.0714757964015007, + 0.13307633996009827, + 0.05568869039416313 + ], + "credit_map": [ + 0.01541684940457344, + 0.013090666383504868, + 0.012845370918512344, + 0.011682871729135513, + 0.016109133139252663, + 0.02020501159131527, + 0.01304200105369091, + 0.01179699320346117, + 0.005913649685680866, + 0.010672369971871376, + 0.004841673653572798, + 0.004273339174687862, + 0.006261083297431469, + 0.008172813802957535, + 0.006921347230672836, + 0.004908365663141012, + 0.007206883281469345, + 0.010617091320455074, + 0.004908329341560602, + 0.007176948711276054, + 0.018633464351296425, + 0.019817067310214043, + 0.011558405123651028, + 0.010879035107791424, + 0.0082868542522192, + 0.010014967061579227, + 0.004331523086875677, + 0.005228244233876467, + 0.028948292136192322, + 0.009424936957657337, + 0.011122461408376694, + 0.0058064223267138, + 0.010283241048455238, + 0.010124744847416878, + 0.010204707272350788, + 0.007739697583019733, + 0.01373380608856678, + 0.01670207269489765, + 0.0039605009369552135, + 0.005361219402402639, + 0.020440809428691864, + 0.015400190837681293, + 0.010474016889929771, + 0.027921350672841072, + 0.11025833338499069, + 0.01715031825006008, + 0.009967638179659843, + 0.05997199937701225, + 0.005581228993833065, + 0.005780939944088459, + 0.003705972572788596, + 0.002733356785029173, + 0.009570654481649399, + 0.007681029848754406, + 0.008829517289996147, + 0.006067340262234211, + 0.009518571197986603, + 0.00957372784614563, + 0.007131274789571762, + 0.006269271019846201, + 0.013170317746698856, + 0.007505974266678095, + 0.012900463305413723, + 0.005406185984611511, + 0.006533393636345863, + 0.008332795463502407, + 0.0029159646946936846, + 0.004541113507002592, + 0.04480670765042305, + 0.010150283575057983, + 0.006622468587011099, + 0.04917360469698906, + 0.004841267596930265, + 0.00413252180442214, + 0.0036450850311666727, + 0.003454075660556555, + 0.0064090546220541, + 0.01755261793732643, + 0.004075285512953997, + 0.011578775011003017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.04306305572390556 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.10030320286750793 + ], + "opposite_logprob_tokens": [ + -2.3503031730651855 + ], + "logprob_reward": -0.10030320286750793, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030320286750793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06089840456843376, + 0.044499319046735764, + 0.03686133027076721, + 0.0326341949403286, + 0.039211831986904144, + 0.6223511695861816, + 0.046989426016807556, + 0.04293443262577057, + 0.03509865701198578, + 0.038521308451890945 + ], + "credit_map": [ + 0.005636689253151417, + 0.00709339277818799, + 0.006196219939738512, + 0.005101253744214773, + 0.01042395643889904, + 0.011585316620767117, + 0.008358011953532696, + 0.006503558252006769, + 0.007696908432990313, + 0.005585836712270975, + 0.004400820005685091, + 0.0032400989439338446, + 0.0051563032902777195, + 0.004790856968611479, + 0.005211373325437307, + 0.008417121134698391, + 0.003395666601136327, + 0.0037236353382468224, + 0.002632349031046033, + 0.00477741239592433, + 0.004709584638476372, + 0.007252328563481569, + 0.005313267931342125, + 0.005057082511484623, + 0.0029834704473614693, + 0.0035425431560724974, + 0.0016312220832332969, + 0.0026720664463937283, + 0.009483863599598408, + 0.0034062624908983707, + 0.00568371219560504, + 0.0032310504466295242, + 0.003284983802586794, + 0.004674915689975023, + 0.002407717751339078, + 0.0029432198498398066, + 0.008216061629354954, + 0.01364661380648613, + 0.0018992050318047404, + 0.0021391138434410095, + 0.04421692714095116, + 0.03550596907734871, + 0.02367621287703514, + 0.06484504044055939, + 0.2382248491048813, + 0.058418553322553635, + 0.025823112577199936, + 0.13164043426513672, + 0.0034793901722878218, + 0.013728369027376175, + 0.0025249391328543425, + 0.0021600937470793724, + 0.01760449819266796, + 0.0026095202192664146, + 0.002730529522523284, + 0.002152081811800599, + 0.0039056451059877872, + 0.010986924171447754, + 0.0028989191632717848, + 0.0036941063590347767, + 0.008393102325499058, + 0.005151356104761362, + 0.003930558916181326, + 0.003973817452788353, + 0.0022592786699533463, + 0.0027910356875509024, + 0.0018585644429549575, + 0.0019455344881862402, + 0.009686766192317009, + 0.0035537693183869123, + 0.002542555797845125, + 0.010461151599884033, + 0.00455106794834137, + 0.004250200930982828, + 0.0054014502093195915, + 0.0023123505525290966, + 0.006790364161133766, + 0.005477858241647482, + 0.0029566006269305944, + 0.0067814127542078495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.10030320286750793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.6327818632125854 + ], + "opposite_logprob_tokens": [ + -0.7577818632125854 + ], + "logprob_reward": -0.6327818632125854, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327818632125854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15018847584724426, + 0.05019360035657883, + 0.08376327157020569, + 0.06271445751190186, + 0.0330386720597744, + 0.16641825437545776, + 0.04921214282512665, + 0.09723571687936783, + 0.08714047819375992, + 0.22009491920471191 + ], + "credit_map": [ + 0.018879691138863564, + 0.014305046759545803, + 0.01572468876838684, + 0.018994789570569992, + 0.027835484594106674, + 0.022387173026800156, + 0.019327374175190926, + 0.012734225019812584, + 0.007302839774638414, + 0.006288113538175821, + 0.003942175768315792, + 0.003786630230024457, + 0.006467654835432768, + 0.006348852068185806, + 0.006527936551719904, + 0.009529398754239082, + 0.005986829753965139, + 0.007435676176100969, + 0.0037595571484416723, + 0.003934560809284449, + 0.02905992791056633, + 0.011054493486881256, + 0.005967547185719013, + 0.01656467467546463, + 0.0059191263280808926, + 0.0075654881075024605, + 0.00422503799200058, + 0.004534205887466669, + 0.015279967337846756, + 0.006073514465242624, + 0.008805150166153908, + 0.010311969555914402, + 0.003217237303033471, + 0.005327434744685888, + 0.0036221477203071117, + 0.0042600492015480995, + 0.004000760614871979, + 0.0074473037384450436, + 0.002229422563686967, + 0.0029343185015022755, + 0.01177741028368473, + 0.011831446550786495, + 0.00996153149753809, + 0.023406878113746643, + 0.049084536731243134, + 0.01963592693209648, + 0.015145357698202133, + 0.025575166568160057, + 0.005647765006870031, + 0.010087213478982449, + 0.0047968789003789425, + 0.0038471135776489973, + 0.008025356568396091, + 0.0048359171487390995, + 0.007738966029137373, + 0.0042329332791268826, + 0.012838136404752731, + 0.010908261872828007, + 0.008129413239657879, + 0.00801533181220293, + 0.018738890066742897, + 0.012692323885858059, + 0.01849174126982689, + 0.0074216146022081375, + 0.006229528225958347, + 0.011211935430765152, + 0.006596530321985483, + 0.007876736111938953, + 0.012665006332099438, + 0.00918327085673809, + 0.008197328075766563, + 0.025180140510201454, + 0.02082510106265545, + 0.015962209552526474, + 0.012059367261826992, + 0.010656451806426048, + 0.01899879425764084, + 0.07580415904521942, + 0.013636041432619095, + 0.05215279012918472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.6327818632125854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "glossy, eye-level camera relentlessly follows a large group of marathon runners, capturing the scene in stunning, surreal detail. The camera moves at their pace, placing a few distinct runners in the foreground with razor-sharp focus. A woman with a high, swinging ponytail, rosy-faced and focused, wears a bright yellow technical vest drenched in sweat. Beside her, a man with a shaved head and a focused expression, every tense muscle fiber in his calves and thighs visible with each powerful stride. Their outfits are a vibrant, chaotic mosaic of neon green, electric blue, and hot pink, all rendered with realistic textures. Their collective action is a thunderous, rhythmic thud against the asphalt of the broad city avenue, a unified forward wave, thousands of individual bodies moving as one undulating mass. Crisp, low-angle morning sunlight filters through the scene, gleaming on their foreheads and casting long, dynamic shadows as they run. In the background, the blurred forms of cheering spectators and city buildings streak by, enhancing the profound sense of speed and unstoppable momentum, creating an epic atmosphere rich in raw determination and human endurance.", + "seed": 21558, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_885_seed_21558_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": -0.15781547948718072, + "vlm_logprob": -0.15781547948718072, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.15781547948718072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.30542701482772827 + ], + "opposite_logprob_tokens": [ + -1.805427074432373 + ], + "logprob_reward": -0.30542701482772827, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.30542701482772827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1720186024904251, + 0.15043102204799652, + 0.11561393737792969, + 0.06154152750968933, + 0.21542826294898987, + 0.025048846378922462, + 0.10008535534143448, + 0.0925254300236702, + 0.045381754636764526, + 0.021925266832113266 + ], + "credit_map": [ + 0.018754273653030396, + 0.01941251941025257, + 0.011971100233495235, + 0.021453164517879486, + 0.05449327453970909, + 0.026550590991973877, + 0.008273352868855, + 0.011110327206552029, + 0.01839483343064785, + 0.006495742592960596, + 0.00767282210290432, + 0.05335038900375366, + 0.03892609477043152, + 0.011632299050688744, + 0.00798215251415968, + 0.005976679269224405, + 0.01683613285422325, + 0.010623835027217865, + 0.008374171331524849, + 0.03354382887482643, + 0.011683209799230099, + 0.011690042912960052, + 0.011094268411397934, + 0.011768446303904057, + 0.00452639814466238, + 0.004421003628522158, + 0.020931320264935493, + 0.012891033664345741, + 0.007728200871497393, + 0.0027470712084323168, + 0.005947783589363098, + 0.0023487145081162453, + 0.010702828876674175, + 0.010965758934617043, + 0.011086717247962952, + 0.0649302750825882, + 0.014170144684612751, + 0.07416597753763199, + 0.0156630277633667, + 0.013743539340794086, + 0.0032092484179884195, + 0.0036602907348424196, + 0.0035677491687238216, + 0.0026733451522886753, + 0.0037772522773593664, + 0.003495532087981701, + 0.0029309645760804415, + 0.0017344631487503648, + 0.013523991219699383, + 0.01122756116092205, + 0.009512034244835377, + 0.01035168394446373, + 0.01652003452181816, + 0.017335200682282448, + 0.011992216110229492, + 0.009622630663216114, + 0.009469476528465748, + 0.021286070346832275, + 0.011988661251962185, + 0.014375175349414349, + 0.009363562799990177, + 0.009972108528017998, + 0.009391670115292072, + 0.006678708363324404, + 0.00592564232647419, + 0.008262604475021362, + 0.01180583331733942, + 0.0041618929244577885, + 0.004758407827466726, + 0.002936693374067545, + 0.004693240858614445, + 0.0028374402318149805, + 0.0024971747770905495, + 0.0033168038353323936, + 0.002464348217472434, + 0.0017413367750123143, + 0.0030561047606170177, + 0.0036967149935662746, + 0.0018616759916767478, + 0.0032911072485148907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.30542701482772827 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.01422879472374916 + ], + "opposite_logprob_tokens": [ + -4.264228820800781 + ], + "logprob_reward": -0.01422879472374916, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422879472374916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15962152183055878, + 0.10714615881443024, + 0.03732791170477867, + 0.1289062798023224, + 0.12402601540088654, + 0.05712379515171051, + 0.13456091284751892, + 0.1800580471754074, + 0.03410162404179573, + 0.03712772950530052 + ], + "credit_map": [ + 0.017413122579455376, + 0.01633826084434986, + 0.01070974487811327, + 0.021282965317368507, + 0.05407087877392769, + 0.023541290313005447, + 0.008205931633710861, + 0.00805931817740202, + 0.00953197106719017, + 0.005572700407356024, + 0.004327662754803896, + 0.02371133491396904, + 0.04181210324168205, + 0.011288930661976337, + 0.006506792735308409, + 0.004394663963466883, + 0.004812149330973625, + 0.004077425692230463, + 0.0027804975397884846, + 0.0035032902378588915, + 0.0045706345699727535, + 0.010898042470216751, + 0.002585017355158925, + 0.0041008563712239265, + 0.008294071070849895, + 0.007968584075570107, + 0.04490945115685463, + 0.03568318486213684, + 0.011674990877509117, + 0.0044203451834619045, + 0.011010762304067612, + 0.004944892600178719, + 0.006616136524826288, + 0.016906393691897392, + 0.006482238415628672, + 0.02506931498646736, + 0.008596044033765793, + 0.0441303551197052, + 0.00807224866002798, + 0.008153270930051804, + 0.006068872753530741, + 0.010392154566943645, + 0.0061165629886090755, + 0.008116372860968113, + 0.006857318338006735, + 0.007741850800812244, + 0.0083077447488904, + 0.003522914368659258, + 0.018667107447981834, + 0.013626000843942165, + 0.012548597529530525, + 0.012029072269797325, + 0.02426895685493946, + 0.025650188326835632, + 0.014398031868040562, + 0.013372971676290035, + 0.016949350014328957, + 0.047655217349529266, + 0.02176695503294468, + 0.018483966588974, + 0.023205431178212166, + 0.02347792498767376, + 0.015729308128356934, + 0.012789907865226269, + 0.004415625706315041, + 0.005917966365814209, + 0.007476376369595528, + 0.004266577307134867, + 0.004001501947641373, + 0.0022083078511059284, + 0.0037747162859886885, + 0.0020405533723533154, + 0.0038798381574451923, + 0.004212345462292433, + 0.003288342384621501, + 0.0023086804430931807, + 0.0050046551041305065, + 0.006213819142431021, + 0.003530916990712285, + 0.008689132519066334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01422879472374916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.04305963218212128 + ], + "opposite_logprob_tokens": [ + -3.1680595874786377 + ], + "logprob_reward": -0.04305963218212128, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305963218212128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09491614997386932, + 0.1281740665435791, + 0.054758887737989426, + 0.331756591796875, + 0.10677676647901535, + 0.03695724532008171, + 0.05895354598760605, + 0.10004612803459167, + 0.040922533720731735, + 0.04673809930682182 + ], + "credit_map": [ + 0.010730231180787086, + 0.0133091751486063, + 0.008429603651165962, + 0.011611366644501686, + 0.0181732177734375, + 0.01978475973010063, + 0.005596112925559282, + 0.007281680125743151, + 0.010547409765422344, + 0.005284433253109455, + 0.0047226459719240665, + 0.038664672523736954, + 0.04439134523272514, + 0.01136060431599617, + 0.006796292494982481, + 0.006406660191714764, + 0.0060270437970757484, + 0.004232492763549089, + 0.005311548709869385, + 0.011595587246119976, + 0.007654519286006689, + 0.010793871246278286, + 0.004703975282609463, + 0.004439849406480789, + 0.020760778337717056, + 0.021759072318673134, + 0.12573803961277008, + 0.08098546415567398, + 0.030947962775826454, + 0.011712168343365192, + 0.028824973851442337, + 0.011028112843632698, + 0.005184829700738192, + 0.009958184324204922, + 0.0054746996611356735, + 0.03236035630106926, + 0.007022183854132891, + 0.03273089975118637, + 0.007569128647446632, + 0.006476479582488537, + 0.0041608139872550964, + 0.0046929907985031605, + 0.003976952750235796, + 0.003694037441164255, + 0.0048842653632164, + 0.008130167610943317, + 0.004374977201223373, + 0.0030430390033870935, + 0.0049494486302137375, + 0.005938048008829355, + 0.005005559418350458, + 0.006389264483004808, + 0.012635307386517525, + 0.011290586553514004, + 0.008710856549441814, + 0.004034474957734346, + 0.0140276700258255, + 0.013345946557819843, + 0.01383274607360363, + 0.012955793179571629, + 0.012136788107454777, + 0.013022222556173801, + 0.013529264368116856, + 0.007195694372057915, + 0.0049926224164664745, + 0.007882080040872097, + 0.010143622756004333, + 0.003942172043025494, + 0.0047985888086259365, + 0.0025674104690551758, + 0.0036486804019659758, + 0.002947357017546892, + 0.004795803688466549, + 0.00518159382045269, + 0.0038735384587198496, + 0.0026933117769658566, + 0.00588639173656702, + 0.009512407705187798, + 0.005546716973185539, + 0.009248337708413601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.04305963218212128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.11293596774339676 + ], + "opposite_logprob_tokens": [ + -2.237936019897461 + ], + "logprob_reward": -0.11293596774339676, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293596774339676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10422123968601227, + 0.3896085321903229, + 0.060827091336250305, + 0.12326078116893768, + 0.08539468050003052, + 0.04969952628016472, + 0.04258100688457489, + 0.08258887380361557, + 0.015341079793870449, + 0.04647720232605934 + ], + "credit_map": [ + 0.01303111296147108, + 0.012353342026472092, + 0.009658736176788807, + 0.01045821513980627, + 0.0258138719946146, + 0.016279062256217003, + 0.007101030088961124, + 0.009525880217552185, + 0.04899771139025688, + 0.02022419311106205, + 0.022175490856170654, + 0.05461050942540169, + 0.14236615598201752, + 0.05262944474816322, + 0.028197521343827248, + 0.02040751837193966, + 0.007919056341052055, + 0.007755190599709749, + 0.0053307912312448025, + 0.005963927134871483, + 0.011997886002063751, + 0.009137162938714027, + 0.004984766244888306, + 0.007738310378044844, + 0.008076120167970657, + 0.007427959702908993, + 0.046654634177684784, + 0.031723737716674805, + 0.012003379873931408, + 0.003619101131334901, + 0.009784156456589699, + 0.0039716982282698154, + 0.004594150464981794, + 0.013075658120214939, + 0.0053210267797112465, + 0.012612393125891685, + 0.007356676738709211, + 0.03188161551952362, + 0.004764004144817591, + 0.0057891542091965675, + 0.005144249182194471, + 0.006168903782963753, + 0.006878561340272427, + 0.004829443991184235, + 0.006297868676483631, + 0.011227442882955074, + 0.005239563062787056, + 0.00391349196434021, + 0.007515526842325926, + 0.004548961762338877, + 0.0028786826878786087, + 0.004156844224780798, + 0.008580198511481285, + 0.008529377169907093, + 0.003582834033295512, + 0.00278858351521194, + 0.011825419962406158, + 0.010812901891767979, + 0.011762646958231926, + 0.010825658217072487, + 0.009785555303096771, + 0.011730441823601723, + 0.008699584752321243, + 0.007146666757762432, + 0.0020567788742482662, + 0.001991159049794078, + 0.0024633051361888647, + 0.002050304552540183, + 0.001982829300686717, + 0.0018011594656854868, + 0.0019803026225417852, + 0.0010152406757697463, + 0.004040607716888189, + 0.005811070092022419, + 0.0041997372172772884, + 0.0027245578821748495, + 0.005760232452303171, + 0.008519020862877369, + 0.008947672322392464, + 0.006474303547292948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.11293596774339676 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.3134259879589081 + ], + "opposite_logprob_tokens": [ + -1.3134260177612305 + ], + "logprob_reward": -0.3134259879589081, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134259879589081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18381191790103912, + 0.11029790341854095, + 0.08003323525190353, + 0.18035997450351715, + 0.11175200343132019, + 0.02596919797360897, + 0.11408647894859314, + 0.04413565248250961, + 0.03129872307181358, + 0.11825492978096008 + ], + "credit_map": [ + 0.0261448435485363, + 0.01779242604970932, + 0.01217959076166153, + 0.02133176475763321, + 0.04816789552569389, + 0.030237829312682152, + 0.01027384027838707, + 0.01768372394144535, + 0.021656999364495277, + 0.005482802167534828, + 0.007543686777353287, + 0.03842960298061371, + 0.01820293813943863, + 0.00829901173710823, + 0.005046844482421875, + 0.005636010784655809, + 0.011684727855026722, + 0.008726164698600769, + 0.006353670731186867, + 0.01870930753648281, + 0.010330946184694767, + 0.011108463630080223, + 0.007169829681515694, + 0.0059501295909285545, + 0.011541088111698627, + 0.010307565331459045, + 0.07076285779476166, + 0.04383337125182152, + 0.017804019153118134, + 0.0049987537786364555, + 0.015398448333144188, + 0.005713872145861387, + 0.006396667566150427, + 0.02431168034672737, + 0.005670969374477863, + 0.03455580398440361, + 0.008246179670095444, + 0.020225390791893005, + 0.007315703667700291, + 0.005029604770243168, + 0.002834648359566927, + 0.0035916294436901808, + 0.0026105279102921486, + 0.0021248653065413237, + 0.004913911689072847, + 0.0036229973193258047, + 0.002513049403205514, + 0.0037575657479465008, + 0.025772007182240486, + 0.014746137894690037, + 0.011341639794409275, + 0.01304025948047638, + 0.014396904967725277, + 0.012701389379799366, + 0.009075279347598553, + 0.013012843206524849, + 0.007609752472490072, + 0.006659465841948986, + 0.004661084618419409, + 0.005438825581222773, + 0.0047215684317052364, + 0.0067841713316738605, + 0.004790681414306164, + 0.0034701006952673197, + 0.0038148474413901567, + 0.005430676508694887, + 0.008919278159737587, + 0.003389674471691251, + 0.0030239014886319637, + 0.0024934825487434864, + 0.0024243046063929796, + 0.0018025540048256516, + 0.011297459714114666, + 0.012450176291167736, + 0.008840165100991726, + 0.006764441262930632, + 0.01817830093204975, + 0.0260397307574749, + 0.011570040136575699, + 0.023114603012800217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.3134259879589081 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "glossy, eye-level camera relentlessly follows a large group of marathon runners, capturing the scene in stunning, surreal detail. The camera moves at their pace, placing a few distinct runners in the foreground with razor-sharp focus. A woman with a high, swinging ponytail, rosy-faced and focused, wears a bright yellow technical vest drenched in sweat. Beside her, a man with a shaved head and a focused expression, every tense muscle fiber in his calves and thighs visible with each powerful stride. Their outfits are a vibrant, chaotic mosaic of neon green, electric blue, and hot pink, all rendered with realistic textures. Their collective action is a thunderous, rhythmic thud against the asphalt of the broad city avenue, a unified forward wave, thousands of individual bodies moving as one undulating mass. Crisp, low-angle morning sunlight filters through the scene, gleaming on their foreheads and casting long, dynamic shadows as they run. In the background, the blurred forms of cheering spectators and city buildings streak by, enhancing the profound sense of speed and unstoppable momentum, creating an epic atmosphere rich in raw determination and human endurance.", + "seed": 21559, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_885_seed_21559_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": -0.22438825592398642, + "vlm_logprob": -0.22438825592398642, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.22438825592398642, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.24352100491523743 + ], + "opposite_logprob_tokens": [ + -1.993520975112915 + ], + "logprob_reward": -0.24352100491523743, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.24352100491523743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19493384659290314, + 0.32900553941726685, + 0.05238547921180725, + 0.06285695731639862, + 0.04269378259778023, + 0.07636809349060059, + 0.04001539945602417, + 0.04235837608575821, + 0.04208029434084892, + 0.11730222404003143 + ], + "credit_map": [ + 0.020098896697163582, + 0.027484271675348282, + 0.022621337324380875, + 0.026630399748682976, + 0.027730746194720268, + 0.02552776411175728, + 0.024317562580108643, + 0.020522860810160637, + 0.04349952191114426, + 0.043901730328798294, + 0.045219164341688156, + 0.04801909998059273, + 0.03266385197639465, + 0.04494345560669899, + 0.04340442642569542, + 0.02735433354973793, + 0.01026524230837822, + 0.0057178763672709465, + 0.004304999485611916, + 0.004923594184219837, + 0.005191831383854151, + 0.011859082616865635, + 0.006060679443180561, + 0.00406218133866787, + 0.00769234262406826, + 0.006181418430060148, + 0.008148953318595886, + 0.01073505636304617, + 0.007679932285100222, + 0.006624004803597927, + 0.007895167917013168, + 0.007900083437561989, + 0.007396381814032793, + 0.0053159031085669994, + 0.005446017719805241, + 0.0037880607414990664, + 0.004568007774651051, + 0.008035732433199883, + 0.0035325183998793364, + 0.004611166659742594, + 0.011278815567493439, + 0.007430132478475571, + 0.006375771947205067, + 0.008975502103567123, + 0.009920255281031132, + 0.011127505451440811, + 0.0069416593760252, + 0.014318455941975117, + 0.005653225351125002, + 0.008639448322355747, + 0.0042091612704098225, + 0.0032530715689063072, + 0.003794009331613779, + 0.005752328783273697, + 0.005302814766764641, + 0.003411343088373542, + 0.003594809677451849, + 0.005001841578632593, + 0.002209907630458474, + 0.0031818475108593702, + 0.009091085754334927, + 0.011350938118994236, + 0.0039011177141219378, + 0.004026834852993488, + 0.010292774066329002, + 0.006500708870589733, + 0.006238211411982775, + 0.003739516483619809, + 0.004428846761584282, + 0.0051210601814091206, + 0.0031246866565197706, + 0.0026344931684434414, + 0.014665589667856693, + 0.018128328025341034, + 0.0194473247975111, + 0.013843507505953312, + 0.023036766797304153, + 0.012125702574849129, + 0.008827882818877697, + 0.007227127905935049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.24352100491523743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.023317888379096985 + ], + "opposite_logprob_tokens": [ + -3.773317813873291 + ], + "logprob_reward": -0.023317888379096985, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023317888379096985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10760927945375443, + 0.19954532384872437, + 0.11899067461490631, + 0.06740471720695496, + 0.03560202941298485, + 0.02979595772922039, + 0.04502417519688606, + 0.1556292027235031, + 0.04904235526919365, + 0.1913563311100006 + ], + "credit_map": [ + 0.012096310965716839, + 0.014450916089117527, + 0.008860774338245392, + 0.00891578197479248, + 0.012981120496988297, + 0.020628472790122032, + 0.010546323843300343, + 0.019129566848278046, + 0.02730695717036724, + 0.03221772611141205, + 0.023830709978938103, + 0.022787919268012047, + 0.02229302190244198, + 0.031152907758951187, + 0.022628329694271088, + 0.01732773147523403, + 0.013392477296292782, + 0.011554268188774586, + 0.012580707669258118, + 0.01207402627915144, + 0.012276886031031609, + 0.015837285667657852, + 0.02883915603160858, + 0.012435856275260448, + 0.010690839029848576, + 0.008618046529591084, + 0.0068052844144403934, + 0.00944448634982109, + 0.00729275681078434, + 0.012178477831184864, + 0.00822940468788147, + 0.004145415034145117, + 0.0051371389999985695, + 0.004952786024659872, + 0.0035226133186370134, + 0.005178761202841997, + 0.004114960320293903, + 0.005170010961592197, + 0.0032093720510601997, + 0.004316381644457579, + 0.005000744946300983, + 0.002479823539033532, + 0.0024018948897719383, + 0.004190617240965366, + 0.004415212664753199, + 0.0031928459648042917, + 0.0029031499288976192, + 0.0052116634324193, + 0.004819819238036871, + 0.006829419638961554, + 0.006114486139267683, + 0.007440000772476196, + 0.005213211290538311, + 0.0044205584563314915, + 0.00629765447229147, + 0.0038890179712325335, + 0.009297999553382397, + 0.03198141232132912, + 0.0047452980652451515, + 0.010316099971532822, + 0.012032104656100273, + 0.06679735332727432, + 0.009987070225179195, + 0.010471849702298641, + 0.007512085605412722, + 0.0074803829193115234, + 0.007110218517482281, + 0.004433763679116964, + 0.007573375012725592, + 0.008185399696230888, + 0.004198137670755386, + 0.0025489856489002705, + 0.020015466958284378, + 0.03566848114132881, + 0.021547256037592888, + 0.016066713258624077, + 0.03849925845861435, + 0.022993043065071106, + 0.01899130269885063, + 0.017574792727828026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.023317888379096985 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.04865511879324913 + ], + "opposite_logprob_tokens": [ + -3.0486550331115723 + ], + "logprob_reward": -0.04865511879324913, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865511879324913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11249729990959167, + 0.1684742122888565, + 0.0878143459558487, + 0.06186966970562935, + 0.056503161787986755, + 0.061311691999435425, + 0.06224773824214935, + 0.1217111125588417, + 0.0724259540438652, + 0.19514480233192444 + ], + "credit_map": [ + 0.01453365758061409, + 0.014406571164727211, + 0.009632092900574207, + 0.010314595885574818, + 0.019313480705022812, + 0.013725854456424713, + 0.015316166914999485, + 0.015254870988428593, + 0.03016764111816883, + 0.0412166565656662, + 0.015690220519900322, + 0.011097593232989311, + 0.016438258811831474, + 0.029951827600598335, + 0.013475200161337852, + 0.010436798445880413, + 0.012838099151849747, + 0.01091019343584776, + 0.010073302313685417, + 0.009114984422922134, + 0.010917008854448795, + 0.015295034274458885, + 0.01183722261339426, + 0.006828497629612684, + 0.00835528876632452, + 0.008194180205464363, + 0.007155395578593016, + 0.007927834056317806, + 0.008069696836173534, + 0.00779682956635952, + 0.00795096717774868, + 0.006419474724680185, + 0.006338033825159073, + 0.00897009763866663, + 0.005913459230214357, + 0.008638493716716766, + 0.006676479708403349, + 0.007447593845427036, + 0.005009549204260111, + 0.007509450428187847, + 0.00825060997158289, + 0.006743729114532471, + 0.005438402760773897, + 0.010116789489984512, + 0.008928358554840088, + 0.0068215071223676205, + 0.005287108477205038, + 0.009725182317197323, + 0.0071882158517837524, + 0.008647479116916656, + 0.006180346477776766, + 0.008190963417291641, + 0.00671602925285697, + 0.007105341646820307, + 0.013275302946567535, + 0.004944054409861565, + 0.006840071640908718, + 0.026489825919270515, + 0.003875840688124299, + 0.007542984094470739, + 0.014101607725024223, + 0.04825274646282196, + 0.007460943888872862, + 0.007147080264985561, + 0.009867815300822258, + 0.009593055583536625, + 0.007725601550191641, + 0.006277766544371843, + 0.009381430223584175, + 0.018130643293261528, + 0.005438420455902815, + 0.006011212710291147, + 0.01638028398156166, + 0.03951116278767586, + 0.014612824656069279, + 0.017739640548825264, + 0.043804917484521866, + 0.026535751298069954, + 0.018795665353536606, + 0.017764538526535034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.04865511879324913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": [ + -0.04865682125091553 + ], + "opposite_logprob_tokens": [ + -3.048656940460205 + ], + "logprob_reward": -0.04865682125091553, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865682125091553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0693846195936203, + 0.14303354918956757, + 0.0391470231115818, + 0.05799628794193268, + 0.05766291916370392, + 0.0275848601013422, + 0.09265074878931046, + 0.29366594552993774, + 0.06474986672401428, + 0.1541241854429245 + ], + "credit_map": [ + 0.006772616412490606, + 0.009696043096482754, + 0.007554655894637108, + 0.007806502748280764, + 0.01033758744597435, + 0.01368353608995676, + 0.0075560035184025764, + 0.00597767299041152, + 0.021808942779898643, + 0.03440602496266365, + 0.02903767302632332, + 0.010996157303452492, + 0.010161682032048702, + 0.022090790793299675, + 0.00563255837187171, + 0.008899719454348087, + 0.005655059590935707, + 0.005220689345151186, + 0.004096560180187225, + 0.0033591187093406916, + 0.0040096561424434185, + 0.008722713217139244, + 0.004703924525529146, + 0.0033793060574680567, + 0.009844464249908924, + 0.008487350307404995, + 0.006186383776366711, + 0.006961253006011248, + 0.006219357252120972, + 0.011251147836446762, + 0.0048821414820849895, + 0.004164185840636492, + 0.006300892680883408, + 0.007314994465559721, + 0.006541989278048277, + 0.005598742049187422, + 0.011362588964402676, + 0.008090722374618053, + 0.006922630127519369, + 0.005530362017452717, + 0.003754489356651902, + 0.002938321093097329, + 0.003566259751096368, + 0.0032799921464174986, + 0.004583550151437521, + 0.0030078021809458733, + 0.0036467774771153927, + 0.0028076686430722475, + 0.010019048117101192, + 0.010572764091193676, + 0.01322442851960659, + 0.01597876474261284, + 0.010203424841165543, + 0.008633493445813656, + 0.016222262755036354, + 0.0077965580858290195, + 0.017287611961364746, + 0.06343358010053635, + 0.009927374310791492, + 0.018675772473216057, + 0.026796892285346985, + 0.11739940941333771, + 0.021441750228405, + 0.018703557550907135, + 0.007493395358324051, + 0.007330304477363825, + 0.007803028915077448, + 0.0064445375464856625, + 0.00971653126180172, + 0.013266428373754025, + 0.0070828963071107864, + 0.0056127458810806274, + 0.013978177681565285, + 0.03138034790754318, + 0.01973862573504448, + 0.012666351161897182, + 0.03023335337638855, + 0.016348840668797493, + 0.015437201596796513, + 0.014341291971504688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -0.04865682125091553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 eye-level and moving at the pace of the runners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with a high swinging ponytail wearing a bright yellow technical vest visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man with a shaved head running beside the woman with visible tension in his calf and thigh muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners' outfits rendered in a chaotic mosaic of neon green, electric blue, and hot pink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by crisp low-angle morning sunlight casting long dynamic shadows on the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577904462814331 + ], + "opposite_logprob_tokens": [ + -0.6327904462814331 + ], + "logprob_reward": -0.7577904462814331, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577904462814331, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11003409326076508, + 0.3507859706878662, + 0.08967740088701248, + 0.09796421229839325, + 0.026992393657565117, + 0.05114390328526497, + 0.029191412031650543, + 0.09471248835325241, + 0.042458441108465195, + 0.10703971982002258 + ], + "credit_map": [ + 0.01072978787124157, + 0.012841957621276379, + 0.015344461426138878, + 0.012962118722498417, + 0.014420497231185436, + 0.016810007393360138, + 0.011534316465258598, + 0.015390951186418533, + 0.05634108930826187, + 0.08457188308238983, + 0.05824357271194458, + 0.02729806862771511, + 0.022361433133482933, + 0.05813497304916382, + 0.0245364997535944, + 0.01929844543337822, + 0.01929125376045704, + 0.011112801730632782, + 0.008237134665250778, + 0.0058697364293038845, + 0.005820681806653738, + 0.029136281460523605, + 0.005427335388958454, + 0.004782168660312891, + 0.01445832196623087, + 0.011034301482141018, + 0.018744977191090584, + 0.011389019899070263, + 0.010434422641992569, + 0.014127364382147789, + 0.008067170158028603, + 0.009708637371659279, + 0.0030477456748485565, + 0.0032159883994609118, + 0.003275408176705241, + 0.003960337955504656, + 0.0031733550131320953, + 0.003921242896467447, + 0.0030201757326722145, + 0.0033781398087739944, + 0.008355814032256603, + 0.005380714312195778, + 0.005851590074598789, + 0.005453165620565414, + 0.006541278678923845, + 0.0075959195382893085, + 0.004693255294114351, + 0.007272164337337017, + 0.0036392919719219208, + 0.0038843450602144003, + 0.004598773084580898, + 0.0038246028125286102, + 0.003392484737560153, + 0.004417875315994024, + 0.002458642702549696, + 0.0029753961134701967, + 0.005347527097910643, + 0.017971720546483994, + 0.002920456463471055, + 0.006305227056145668, + 0.00801000650972128, + 0.04049218073487282, + 0.00648361723870039, + 0.007181749679148197, + 0.007018860895186663, + 0.005468674469739199, + 0.0037505177315324545, + 0.004566211719065905, + 0.006849909201264381, + 0.006046648602932692, + 0.0058495537377893925, + 0.002908063121140003, + 0.016052385792136192, + 0.012030789628624916, + 0.02099146507680416, + 0.01911136507987976, + 0.015402517281472683, + 0.007982106879353523, + 0.007643402554094791, + 0.007825683802366257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.7577904462814331 + } + ], + "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/0000052/group_8/index_tensor([8822, 8822])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_8/index_tensor([8822, 8822])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a9cccc5c97fcefa138726c019a05ef1071ca87cb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_8/index_tensor([8822, 8822])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot captures the tender yet momentous exchange of an exquisite antique silver ring. The camera remains perfectly still, focused on two pairs of hands suspended above a dark, out-of-focus wooden surface. The giving hand, showing the fine lines and gentle wrinkles of age, reverently holds the ring between thumb and index finger. The receiving hand, youthful and smooth, is slightly open and upward, fingers poised to receive the precious object. The ring itself is a work of intricate filigree, its silver band bearing a soft, time-worn patina in the crevices, contrasting strikingly with the polished surface as it gleams as it catches a single, warm, directional light source. In slow, deliberate micro-movements, the older hand steadily moves forward, placing the ring into the palm of the younger hand, the metal making a soft, almost imperceptible contact. The entire scene is bathed in warm, low-key lighting, creating deep, soft shadows that accentuate the textures of the skin and the uneven metal, evoking a tranquil and profound atmosphere of inheritance.", + "seed": 24838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_8/index_8822_seed_24838_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": -0.35793848625617103, + "vlm_logprob": -0.35793848625617103, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.35793848625617103, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011915224604308605 + ], + "opposite_logprob_tokens": [ + -5.13691520690918 + ], + "logprob_reward": -0.011915224604308605, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011915224604308605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05219243839383125, + 0.04934142157435417, + 0.16469676792621613, + 0.603895902633667, + 0.02673572674393654, + 0.019230546429753304, + 0.014332409016788006, + 0.022698380053043365, + 0.0311624463647604, + 0.01571396179497242 + ], + "credit_map": [ + 0.005683101713657379, + 0.005724823568016291, + 0.007626781705766916, + 0.008489426225423813, + 0.006425816100090742, + 0.0073059480637311935, + 0.005038755480200052, + 0.005897784139961004, + 0.005898218136280775, + 0.0020624399185180664, + 0.015487445518374443, + 0.003125871764495969, + 0.0025281396228820086, + 0.004345140885561705, + 0.0036844576243311167, + 0.012209709733724594, + 0.013329417444765568, + 0.004514452535659075, + 0.0596139170229435, + 0.005531682167202234, + 0.007197660859674215, + 0.06355028599500656, + 0.003757020691409707, + 0.0072023323737084866, + 0.11935590952634811, + 0.054451972246170044, + 0.05409729480743408, + 0.07085981965065002, + 0.09616110473871231, + 0.12267180532217026, + 0.046653226017951965, + 0.0396447628736496, + 0.0031746893655508757, + 0.0022541156504303217, + 0.004756922367960215, + 0.0036827982403337955, + 0.0016064748633652925, + 0.006433393806219101, + 0.0034294656943529844, + 0.0013978672213852406, + 0.0024911838117986917, + 0.002521938644349575, + 0.0032531947363168, + 0.0026354598812758923, + 0.0014647638890892267, + 0.002253876766189933, + 0.0019466502126306295, + 0.0026634796522557735, + 0.002274914178997278, + 0.0017355934251099825, + 0.0017978852847591043, + 0.0016059806803241372, + 0.0011371533619239926, + 0.002441453281790018, + 0.0017978276591748, + 0.001541602541692555, + 0.0039202445186674595, + 0.0022231419570744038, + 0.0038253264501690865, + 0.003159408224746585, + 0.0016206237487494946, + 0.0028036069124937057, + 0.002547990530729294, + 0.002598038874566555, + 0.004090030677616596, + 0.006393145304173231, + 0.004314659163355827, + 0.0030226614326238632, + 0.0036081206053495407, + 0.003480255603790283, + 0.0034417174756526947, + 0.002811857033520937, + 0.0033120852895081043, + 0.0020065410062670708, + 0.003949332516640425, + 0.001525334664620459, + 0.0009014373645186424, + 0.0017716814763844013, + 0.0009402878349646926, + 0.001307262689806521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.011915224604308605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015365362167358 + ], + "opposite_logprob_tokens": [ + -0.20153652131557465 + ], + "logprob_reward": -1.7015365362167358, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015365362167358, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.02234261855483055, + 0.041473425924777985, + 0.04014483466744423, + 0.7354125380516052, + 0.043447159230709076, + 0.020352190360426903, + 0.027669312432408333, + 0.0203019380569458, + 0.031522706151008606, + 0.017333200201392174 + ], + "credit_map": [ + 0.003499467857182026, + 0.00261171767488122, + 0.002920814324170351, + 0.002757843118160963, + 0.0025435825809836388, + 0.002588255563750863, + 0.002627704758197069, + 0.0027932352386415005, + 0.00482152821496129, + 0.0031954972073435783, + 0.010563593357801437, + 0.0036136603448539972, + 0.0035371470730751753, + 0.005085568875074387, + 0.004388326779007912, + 0.006268106866627932, + 0.007745569106191397, + 0.002872269367799163, + 0.00478714145720005, + 0.002906988374888897, + 0.004835755564272404, + 0.009950535371899605, + 0.002739401999861002, + 0.004307175986468792, + 0.15859109163284302, + 0.053959865123033524, + 0.19021853804588318, + 0.07211872190237045, + 0.09054113924503326, + 0.08263836801052094, + 0.04588215798139572, + 0.04146276414394379, + 0.004315210040658712, + 0.003587346989661455, + 0.012105833739042282, + 0.0042949942871928215, + 0.002480383962392807, + 0.011509671807289124, + 0.0033176441211253405, + 0.0018360824324190617, + 0.00278710899874568, + 0.0021408754400908947, + 0.0037288896273821592, + 0.0024304825346916914, + 0.0017448847647756338, + 0.003933959640562534, + 0.001501630642451346, + 0.002084361854940653, + 0.003609428880736232, + 0.004025743342936039, + 0.0037763991858810186, + 0.0042190770618617535, + 0.0029431015718728304, + 0.004108325112611055, + 0.0027927029877901077, + 0.002194536617025733, + 0.003080102615058422, + 0.002476142020896077, + 0.0033641529735177755, + 0.003118008840829134, + 0.0012872850056737661, + 0.002616123529151082, + 0.0017708441009745002, + 0.002589280717074871, + 0.004373535513877869, + 0.005816150456666946, + 0.00476819509640336, + 0.003157384227961302, + 0.0032878578640520573, + 0.00407594395801425, + 0.0030244109220802784, + 0.0030192327685654163, + 0.0028880632016807795, + 0.003754011122509837, + 0.002510917140170932, + 0.0019088438712060452, + 0.0011416287161409855, + 0.002047209069132805, + 0.0014386101393029094, + 0.0016439197352156043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -1.7015365362167358 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052687362767755985 + ], + "opposite_logprob_tokens": [ + -5.255268573760986 + ], + "logprob_reward": -0.0052687362767755985, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052687362767755985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08154769241809845, + 0.13106875121593475, + 0.06006723642349243, + 0.5337936878204346, + 0.09202037751674652, + 0.015094419941306114, + 0.01615770347416401, + 0.02989773266017437, + 0.022407973185181618, + 0.017944378778338432 + ], + "credit_map": [ + 0.010394901968538761, + 0.007147390861064196, + 0.01548047736287117, + 0.010082827880978584, + 0.00636991485953331, + 0.010644746944308281, + 0.006646741647273302, + 0.01478069368749857, + 0.016062302514910698, + 0.0036116773262619972, + 0.04948115348815918, + 0.00590989924967289, + 0.007267449051141739, + 0.011082970537245274, + 0.0061569432727992535, + 0.03149636462330818, + 0.008374342694878578, + 0.0036709250416606665, + 0.020933454856276512, + 0.0033788522705435753, + 0.004551964811980724, + 0.012913812883198261, + 0.0025999476201832294, + 0.003643938573077321, + 0.10214371234178543, + 0.03281686082482338, + 0.1556224524974823, + 0.046462010592222214, + 0.054346006363630295, + 0.08687296509742737, + 0.029185468330979347, + 0.02634425461292267, + 0.008388514630496502, + 0.004352842923253775, + 0.031292904168367386, + 0.004930869676172733, + 0.0033315643668174744, + 0.032854799181222916, + 0.003343917429447174, + 0.0035249684005975723, + 0.0021099275909364223, + 0.0020218975841999054, + 0.0028634651098400354, + 0.0018576482543721795, + 0.0012851647334173322, + 0.0024749101139605045, + 0.001242665690369904, + 0.0012387412134557962, + 0.0020492970943450928, + 0.0015568885719403625, + 0.002331478288397193, + 0.0019583599641919136, + 0.0011589705245569348, + 0.003100187284871936, + 0.0020487182773649693, + 0.0019538046326488256, + 0.004885400645434856, + 0.0030512281227856874, + 0.005053726024925709, + 0.004063146188855171, + 0.002731068292632699, + 0.004077617544680834, + 0.003190868766978383, + 0.0028446789365261793, + 0.0031241336837410927, + 0.0038652096409350634, + 0.0035492535680532455, + 0.0023809021804481745, + 0.002340466482564807, + 0.0027734471950680017, + 0.002093912335112691, + 0.0022806492634117603, + 0.0036250571720302105, + 0.002568477764725685, + 0.0033984410110861063, + 0.001935355830937624, + 0.0010806625941768289, + 0.002057293662801385, + 0.00162673054728657, + 0.0016523606609553099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.0052687362767755985 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.070048488676548 + ], + "opposite_logprob_tokens": [ + -2.6950485706329346 + ], + "logprob_reward": -0.070048488676548, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.070048488676548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041209109127521515, + 0.0976925790309906, + 0.04338444769382477, + 0.6959121823310852, + 0.04241045191884041, + 0.011767354793846607, + 0.022673428058624268, + 0.005139689892530441, + 0.010321598500013351, + 0.02948918007314205 + ], + "credit_map": [ + 0.004272985737770796, + 0.004217428620904684, + 0.005244475789368153, + 0.007627591956406832, + 0.003990486729890108, + 0.005333006381988525, + 0.004307685419917107, + 0.006215446628630161, + 0.012226484715938568, + 0.0019409821834415197, + 0.03771847113966942, + 0.004122924525290728, + 0.006151992827653885, + 0.008131718263030052, + 0.0036713078152388334, + 0.02372870221734047, + 0.006512348540127277, + 0.003003139980137348, + 0.00599081302061677, + 0.003976327367126942, + 0.0034852938260883093, + 0.013319500721991062, + 0.003646836383268237, + 0.003450188087299466, + 0.15271203219890594, + 0.061442043632268906, + 0.12512914836406708, + 0.08290168642997742, + 0.10695864260196686, + 0.06699281930923462, + 0.05328989028930664, + 0.04648597538471222, + 0.0038419112097471952, + 0.0018410469638183713, + 0.014285538345575333, + 0.0036655922885984182, + 0.0013916639145463705, + 0.01474972628057003, + 0.001438200124539435, + 0.0011967760510742664, + 0.0016518530901521444, + 0.0015327982837334275, + 0.0017612745286896825, + 0.00207275222055614, + 0.0008513187640346587, + 0.0014835642650723457, + 0.0013626425061374903, + 0.00105115061160177, + 0.00282261217944324, + 0.0028091203421354294, + 0.0029454808682203293, + 0.003567917738109827, + 0.001847017090767622, + 0.0036114577669650316, + 0.002596867037937045, + 0.002472955733537674, + 0.0006895553087815642, + 0.0007058395422063768, + 0.0007464592927135527, + 0.0006453670212067664, + 0.0003760316758416593, + 0.000813826743979007, + 0.000628418056294322, + 0.0005341920768842101, + 0.0014425897970795631, + 0.00151224120054394, + 0.0015358225209638476, + 0.0011557876132428646, + 0.0009814929217100143, + 0.001693636178970337, + 0.0009278815123252571, + 0.0010721472790464759, + 0.004851681645959616, + 0.0052254097536206245, + 0.004456616472452879, + 0.002881139051169157, + 0.0018036317778751254, + 0.0033481030259281397, + 0.0038475007750093937, + 0.0030750955920666456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.070048488676548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009234455064870417 + ], + "opposite_logprob_tokens": [ + -7.0009236335754395 + ], + "logprob_reward": -0.0009234455064870417, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009234455064870417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06633125990629196, + 0.04024438560009003, + 0.08253506571054459, + 0.48923206329345703, + 0.08643434196710587, + 0.02990785613656044, + 0.03012095019221306, + 0.04295302927494049, + 0.04133852943778038, + 0.09090249240398407 + ], + "credit_map": [ + 0.011906434781849384, + 0.007875485345721245, + 0.008735813200473785, + 0.0074987770058214664, + 0.004964897409081459, + 0.007632313761860132, + 0.0049110096879303455, + 0.012806526385247707, + 0.006464773789048195, + 0.003322353819385171, + 0.008626622147858143, + 0.004137351177632809, + 0.0033637506421655416, + 0.005368805956095457, + 0.005134099628776312, + 0.0038266258779913187, + 0.007393490057438612, + 0.008408380672335625, + 0.017715297639369965, + 0.006744174752384424, + 0.004821258597075939, + 0.024232210591435432, + 0.007086057215929031, + 0.006134191993623972, + 0.114749014377594, + 0.035174332559108734, + 0.12950994074344635, + 0.0460657998919487, + 0.05677355080842972, + 0.05230637267231941, + 0.029381506145000458, + 0.025271529331803322, + 0.008220481686294079, + 0.003929985221475363, + 0.02900438755750656, + 0.006622646003961563, + 0.0029738766606897116, + 0.030103670433163643, + 0.00276582851074636, + 0.0028134677559137344, + 0.0038251113146543503, + 0.003123234724625945, + 0.005899231880903244, + 0.003705740673467517, + 0.002298773266375065, + 0.006226271390914917, + 0.0023494011256843805, + 0.002480091294273734, + 0.004185713827610016, + 0.0031499231699854136, + 0.00598823931068182, + 0.00321748573333025, + 0.0016628449084237218, + 0.005268163047730923, + 0.002578338608145714, + 0.004070241004228592, + 0.006277271080762148, + 0.004496913868933916, + 0.006923905573785305, + 0.007485099136829376, + 0.0034369879867881536, + 0.005543956533074379, + 0.00401881942525506, + 0.004770078230649233, + 0.005775182042270899, + 0.007649976760149002, + 0.006643658969551325, + 0.004045442678034306, + 0.004545439034700394, + 0.004678610246628523, + 0.004131801892071962, + 0.0038684182800352573, + 0.012777776457369328, + 0.02251638099551201, + 0.011843009851872921, + 0.008701733313500881, + 0.006301030050963163, + 0.010795094072818756, + 0.00799383781850338, + 0.009973632171750069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.0009234455064870417 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the tender yet momentous exchange of an exquisite antique silver ring. The camera remains perfectly still, focused on two pairs of hands suspended above a dark, out-of-focus wooden surface. The giving hand, showing the fine lines and gentle wrinkles of age, reverently holds the ring between thumb and index finger. The receiving hand, youthful and smooth, is slightly open and upward, fingers poised to receive the precious object. The ring itself is a work of intricate filigree, its silver band bearing a soft, time-worn patina in the crevices, contrasting strikingly with the polished surface as it gleams as it catches a single, warm, directional light source. In slow, deliberate micro-movements, the older hand steadily moves forward, placing the ring into the palm of the younger hand, the metal making a soft, almost imperceptible contact. The entire scene is bathed in warm, low-key lighting, creating deep, soft shadows that accentuate the textures of the skin and the uneven metal, evoking a tranquil and profound atmosphere of inheritance.", + "seed": 24839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_8/index_8822_seed_24839_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": -0.3387865411816165, + "vlm_logprob": -0.3387865411816165, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.3387865411816165, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02282746508717537 + ], + "opposite_logprob_tokens": [ + -4.147827625274658 + ], + "logprob_reward": -0.02282746508717537, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02282746508717537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3069533109664917, + 0.10494449734687805, + 0.05054028332233429, + 0.11756906658411026, + 0.04479096457362175, + 0.09103155881166458, + 0.09072906523942947, + 0.046560030430555344, + 0.09773290157318115, + 0.0491483137011528 + ], + "credit_map": [ + 0.04388653486967087, + 0.04855089262127876, + 0.03565540164709091, + 0.027000200003385544, + 0.02873843163251877, + 0.04209042713046074, + 0.044187624007463455, + 0.03684379532933235, + 0.011986488476395607, + 0.017913689836859703, + 0.012514384463429451, + 0.008812076412141323, + 0.015131812542676926, + 0.011807547882199287, + 0.015019225887954235, + 0.011759272776544094, + 0.007901857607066631, + 0.005956158507615328, + 0.006540863309055567, + 0.003623874159529805, + 0.007537344004958868, + 0.006079929880797863, + 0.00705662090331316, + 0.00584363890811801, + 0.019845958799123764, + 0.011605369858443737, + 0.013666951097548008, + 0.012490889988839626, + 0.013390188105404377, + 0.01988069899380207, + 0.013748103752732277, + 0.012940913438796997, + 0.005125223193317652, + 0.004376834258437157, + 0.005616831127554178, + 0.004966098349541426, + 0.00623111380264163, + 0.007389828562736511, + 0.00622754730284214, + 0.0048574889078736305, + 0.008565369062125683, + 0.004225445911288261, + 0.027563637122511864, + 0.006709717679768801, + 0.00775571633130312, + 0.010740409605205059, + 0.005795832723379135, + 0.01967543549835682, + 0.009004582650959492, + 0.00526704266667366, + 0.021970734000205994, + 0.006436709780246019, + 0.011596117168664932, + 0.0082290293648839, + 0.008676422759890556, + 0.01954841986298561, + 0.004154480528086424, + 0.0030799696687608957, + 0.010911060497164726, + 0.003550187451764941, + 0.006643120665103197, + 0.005947657860815525, + 0.0038256170228123665, + 0.008447934873402119, + 0.007416057400405407, + 0.007507034111768007, + 0.02185225486755371, + 0.007431074045598507, + 0.015132453292608261, + 0.010531370528042316, + 0.00820229109376669, + 0.019660353660583496, + 0.004112150985747576, + 0.005059112794697285, + 0.009184475056827068, + 0.00498251523822546, + 0.005949357990175486, + 0.005023411009460688, + 0.0070269666612148285, + 0.007810323033481836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.02282746508717537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5021001100540161 + ], + "opposite_logprob_tokens": [ + -0.2521000802516937 + ], + "logprob_reward": -1.5021001100540161, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5021001100540161, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09212006628513336, + 0.09694235771894455, + 0.036121733486652374, + 0.04831760749220848, + 0.04425426200032234, + 0.19142049551010132, + 0.1277102530002594, + 0.08468267321586609, + 0.13967572152614594, + 0.13875487446784973 + ], + "credit_map": [ + 0.011162051931023598, + 0.012109838426113129, + 0.01106346771121025, + 0.00962758157402277, + 0.014467759989202023, + 0.011587015353143215, + 0.011111029423773289, + 0.01099131628870964, + 0.013076072558760643, + 0.016076436266303062, + 0.011264522559940815, + 0.007204292342066765, + 0.014387775212526321, + 0.009931635111570358, + 0.01634908840060234, + 0.008652529679238796, + 0.00671145785599947, + 0.0041792732663452625, + 0.003931437619030476, + 0.0036084444727748632, + 0.0052486504428088665, + 0.0040415492840111256, + 0.00496203638613224, + 0.003438884625211358, + 0.010099424049258232, + 0.005329234059900045, + 0.006031919736415148, + 0.005892579909414053, + 0.005902568344026804, + 0.005109835881739855, + 0.00502570578828454, + 0.004926339723169804, + 0.005544506944715977, + 0.0039141494780778885, + 0.006860954686999321, + 0.005687615368515253, + 0.004566962365061045, + 0.0063169668428599834, + 0.007048526778817177, + 0.0043145762756466866, + 0.017088159918785095, + 0.0191362164914608, + 0.028815170750021935, + 0.02768305502831936, + 0.033877287060022354, + 0.02394401840865612, + 0.0205785371363163, + 0.02029803767800331, + 0.012532917782664299, + 0.0060860211960971355, + 0.03639737144112587, + 0.0124777527526021, + 0.009827972389757633, + 0.011853408068418503, + 0.009227999486029148, + 0.029306787997484207, + 0.008654733188450336, + 0.005361868068575859, + 0.013344863429665565, + 0.008477794006466866, + 0.010239267721772194, + 0.011590021662414074, + 0.009811547584831715, + 0.017202576622366905, + 0.012025250121951103, + 0.014182266779243946, + 0.02390754036605358, + 0.011609658598899841, + 0.025545738637447357, + 0.016982076689600945, + 0.013066484592854977, + 0.022356700152158737, + 0.014147044159471989, + 0.010729726403951645, + 0.030475648120045662, + 0.011680036783218384, + 0.014510566368699074, + 0.010350530035793781, + 0.018049215897917747, + 0.028812089934945107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -1.5021001100540161 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646813057363033 + ], + "opposite_logprob_tokens": [ + -4.8826470375061035 + ], + "logprob_reward": -0.007646813057363033, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646813057363033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17914748191833496, + 0.10759606957435608, + 0.058928970247507095, + 0.10995635390281677, + 0.036508798599243164, + 0.18674468994140625, + 0.09426762163639069, + 0.09478813409805298, + 0.05331084504723549, + 0.07875099778175354 + ], + "credit_map": [ + 0.028174476698040962, + 0.03117470256984234, + 0.027660571038722992, + 0.016359636560082436, + 0.026664728298783302, + 0.017468364909291267, + 0.014656735584139824, + 0.01698826067149639, + 0.018998760730028152, + 0.014897129498422146, + 0.014517363160848618, + 0.007942335680127144, + 0.01810937188565731, + 0.011832629330456257, + 0.009576293639838696, + 0.011722189374268055, + 0.010338738560676575, + 0.0057763285003602505, + 0.007768223062157631, + 0.006427491549402475, + 0.006943915504962206, + 0.008169450797140598, + 0.005103519186377525, + 0.008401300758123398, + 0.020969051867723465, + 0.00942511297762394, + 0.014343560673296452, + 0.009814956225454807, + 0.014273873530328274, + 0.018447063863277435, + 0.008413095958530903, + 0.01426964346319437, + 0.0040102009661495686, + 0.0032168107572942972, + 0.005305572412908077, + 0.0036695252638310194, + 0.004977405536919832, + 0.005244037136435509, + 0.005507221445441246, + 0.004578022286295891, + 0.015508032403886318, + 0.01603163219988346, + 0.03618858754634857, + 0.021061135455965996, + 0.030786827206611633, + 0.021605856716632843, + 0.01771778054535389, + 0.02784484624862671, + 0.010262367315590382, + 0.0059864092618227005, + 0.02394404448568821, + 0.00901002623140812, + 0.009593388065695763, + 0.00880963634699583, + 0.008945927955210209, + 0.017715824767947197, + 0.007947729900479317, + 0.004869446624070406, + 0.021993586793541908, + 0.006962082348763943, + 0.012853980995714664, + 0.009984023869037628, + 0.007510403636842966, + 0.02266688086092472, + 0.006357225589454174, + 0.005492453463375568, + 0.008784071542322636, + 0.0059818061999976635, + 0.008573790080845356, + 0.005730851087719202, + 0.005690401885658503, + 0.006700247060507536, + 0.008759607560932636, + 0.007551965769380331, + 0.011938845738768578, + 0.00962988194078207, + 0.008992671966552734, + 0.010979187674820423, + 0.01157181803137064, + 0.009327024221420288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.007646813057363033 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16043522953987122 + ], + "opposite_logprob_tokens": [ + -1.9104351997375488 + ], + "logprob_reward": -0.16043522953987122, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16043522953987122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24996300041675568, + 0.14146798849105835, + 0.05044800415635109, + 0.08488431572914124, + 0.040144503116607666, + 0.21552112698554993, + 0.08733019977807999, + 0.030472954735159874, + 0.07189774513244629, + 0.027870148420333862 + ], + "credit_map": [ + 0.03486574441194534, + 0.03814246132969856, + 0.028626464307308197, + 0.022746216505765915, + 0.023442795500159264, + 0.029764428734779358, + 0.046497151255607605, + 0.025877734646201134, + 0.018570266664028168, + 0.025049259886145592, + 0.016948368400335312, + 0.010538375005126, + 0.019857609644532204, + 0.014079944230616093, + 0.023096317425370216, + 0.01332783978432417, + 0.010736367665231228, + 0.00637274794280529, + 0.005680941045284271, + 0.0038864114321768284, + 0.007211131509393454, + 0.005066509824246168, + 0.006820707116276026, + 0.004673190414905548, + 0.014827810227870941, + 0.007941852323710918, + 0.010361500084400177, + 0.009323670528829098, + 0.00922541692852974, + 0.014839435927569866, + 0.007863053120672703, + 0.010501574724912643, + 0.005664416588842869, + 0.003918612841516733, + 0.004977217875421047, + 0.004944432992488146, + 0.004838332533836365, + 0.005510824732482433, + 0.0058527677319943905, + 0.004437900613993406, + 0.020821914076805115, + 0.020809322595596313, + 0.0391240194439888, + 0.028992190957069397, + 0.0314585417509079, + 0.025562888011336327, + 0.02178437076508999, + 0.026967879384756088, + 0.00768895773217082, + 0.0055990456603467464, + 0.01882064715027809, + 0.008763361722230911, + 0.010758481919765472, + 0.008674119599163532, + 0.00955550279468298, + 0.01747007854282856, + 0.0037426853086799383, + 0.0021045489702373743, + 0.004680270329117775, + 0.00308579346165061, + 0.004263644572347403, + 0.0046742758713662624, + 0.0036452619824558496, + 0.004276473540812731, + 0.005960548762232065, + 0.006347908638417721, + 0.013562152162194252, + 0.00601496547460556, + 0.011987054720520973, + 0.00865495391190052, + 0.006308955140411854, + 0.013061204925179482, + 0.0026763658970594406, + 0.0034876172430813313, + 0.0038766348734498024, + 0.0033782999962568283, + 0.003207430010661483, + 0.004062397871166468, + 0.0037997844628989697, + 0.003381620394065976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.16043522953987122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009230881696566939 + ], + "opposite_logprob_tokens": [ + -7.000923156738281 + ], + "logprob_reward": -0.0009230881696566939, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009230881696566939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24011890590190887, + 0.07822445780038834, + 0.045128945261240005, + 0.04295677691698074, + 0.04133262485265732, + 0.15658895671367645, + 0.07527153193950653, + 0.07733829319477081, + 0.06558609008789062, + 0.17745347321033478 + ], + "credit_map": [ + 0.03254317492246628, + 0.02924107015132904, + 0.02929258905351162, + 0.023825597018003464, + 0.02656232938170433, + 0.03603542596101761, + 0.03020026907324791, + 0.032418444752693176, + 0.012106863781809807, + 0.009903090074658394, + 0.010487117804586887, + 0.0071358042769134045, + 0.012382342480123043, + 0.008392812684178352, + 0.006911438889801502, + 0.010904984548687935, + 0.008004654198884964, + 0.004635038785636425, + 0.005798829719424248, + 0.003326908452436328, + 0.005682704970240593, + 0.007160862907767296, + 0.00511946203187108, + 0.005400483496487141, + 0.008380650542676449, + 0.004429385531693697, + 0.005558268167078495, + 0.003511392278596759, + 0.0058896504342556, + 0.006688558030873537, + 0.0037524413783103228, + 0.004746430553495884, + 0.006000473629683256, + 0.005503636319190264, + 0.00485830707475543, + 0.0035285374615341425, + 0.005767116788774729, + 0.006369305774569511, + 0.004522170405834913, + 0.00478307856246829, + 0.011981107294559479, + 0.011875954456627369, + 0.0366169810295105, + 0.016304006800055504, + 0.021438686177134514, + 0.019167844206094742, + 0.013179441913962364, + 0.026024941354990005, + 0.00753005500882864, + 0.006713536102324724, + 0.013106629252433777, + 0.009375274181365967, + 0.00965954177081585, + 0.008175668306648731, + 0.010520655661821365, + 0.010190167464315891, + 0.006746327970176935, + 0.004097889177501202, + 0.019719665870070457, + 0.0048704808577895164, + 0.009667386300861835, + 0.008224712684750557, + 0.004878269508481026, + 0.019133560359477997, + 0.007273893803358078, + 0.007451948709785938, + 0.008431713096797466, + 0.006506755482405424, + 0.012051539495587349, + 0.008571680635213852, + 0.007381963077932596, + 0.007916595786809921, + 0.015240815468132496, + 0.010444717481732368, + 0.049104735255241394, + 0.010804878547787666, + 0.01469358243048191, + 0.0163356252014637, + 0.016324566677212715, + 0.044504549354314804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0009230881696566939 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the tender yet momentous exchange of an exquisite antique silver ring. The camera remains perfectly still, focused on two pairs of hands suspended above a dark, out-of-focus wooden surface. The giving hand, showing the fine lines and gentle wrinkles of age, reverently holds the ring between thumb and index finger. The receiving hand, youthful and smooth, is slightly open and upward, fingers poised to receive the precious object. The ring itself is a work of intricate filigree, its silver band bearing a soft, time-worn patina in the crevices, contrasting strikingly with the polished surface as it gleams as it catches a single, warm, directional light source. In slow, deliberate micro-movements, the older hand steadily moves forward, placing the ring into the palm of the younger hand, the metal making a soft, almost imperceptible contact. The entire scene is bathed in warm, low-key lighting, creating deep, soft shadows that accentuate the textures of the skin and the uneven metal, evoking a tranquil and profound atmosphere of inheritance.", + "seed": 24840, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_8/index_8822_seed_24840_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": -0.34001387549797074, + "vlm_logprob": -0.34001387549797074, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.34001387549797074, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023998375982046127 + ], + "opposite_logprob_tokens": [ + -4.273998260498047 + ], + "logprob_reward": -0.023998375982046127, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.023998375982046127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08854977041482925, + 0.05241744965314865, + 0.01840779557824135, + 0.06708700954914093, + 0.3498288094997406, + 0.15260744094848633, + 0.05341773107647896, + 0.08438225090503693, + 0.05057426169514656, + 0.08272743225097656 + ], + "credit_map": [ + 0.010032296180725098, + 0.007893158122897148, + 0.01222299225628376, + 0.009242790751159191, + 0.007657795678824186, + 0.009383481927216053, + 0.00745184114202857, + 0.024665411561727524, + 0.006914515979588032, + 0.004407616797834635, + 0.007737739011645317, + 0.0033270069397985935, + 0.0030861361883580685, + 0.006129581946879625, + 0.005174554884433746, + 0.015640301629900932, + 0.003553481074050069, + 0.0019182439427822828, + 0.003358377842232585, + 0.001694075996056199, + 0.0015484297182410955, + 0.002050355775281787, + 0.002036681631579995, + 0.002248148899525404, + 0.012485145591199398, + 0.004811157938092947, + 0.007127452176064253, + 0.00920881051570177, + 0.00652514910325408, + 0.007926683872938156, + 0.006066935136914253, + 0.012935676611959934, + 0.13338762521743774, + 0.010708481073379517, + 0.022496143355965614, + 0.010878142900764942, + 0.012179749086499214, + 0.016212141141295433, + 0.0116749107837677, + 0.13229161500930786, + 0.05467977374792099, + 0.006506144069135189, + 0.008794269524514675, + 0.00792613998055458, + 0.007768852636218071, + 0.010650222189724445, + 0.008672495372593403, + 0.04760954901576042, + 0.008819863200187683, + 0.004910067655146122, + 0.008819754235446453, + 0.006894910708069801, + 0.0063324118964374065, + 0.0059014130383729935, + 0.0046069007366895676, + 0.007132409606128931, + 0.016762642189860344, + 0.0071854800917208195, + 0.01780957356095314, + 0.008278426714241505, + 0.009273653849959373, + 0.008126312866806984, + 0.006263429764658213, + 0.010682730004191399, + 0.008904173038899899, + 0.0033765488769859076, + 0.010126636363565922, + 0.005914711393415928, + 0.004191853106021881, + 0.006058762781322002, + 0.005243686027824879, + 0.006757889408618212, + 0.01833157055079937, + 0.006809341721236706, + 0.0115908058360219, + 0.008648629300296307, + 0.00801214948296547, + 0.010175450704991817, + 0.008035224862396717, + 0.011124253273010254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.023998375982046127 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005513668060303 + ], + "opposite_logprob_tokens": [ + -0.22555135190486908 + ], + "logprob_reward": -1.6005513668060303, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005513668060303, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07446719706058502, + 0.08395113795995712, + 0.0685332864522934, + 0.05261625722050667, + 0.2061973512172699, + 0.3336365818977356, + 0.055268414318561554, + 0.020648576319217682, + 0.06600290536880493, + 0.03867829218506813 + ], + "credit_map": [ + 0.011190414428710938, + 0.0067667183466255665, + 0.010148902423679829, + 0.009119461290538311, + 0.006162307690829039, + 0.007379984948784113, + 0.008507310412824154, + 0.015192091464996338, + 0.012548545375466347, + 0.005532809998840094, + 0.021265236660838127, + 0.0052218446508049965, + 0.005944814067333937, + 0.008379512466490269, + 0.007844094187021255, + 0.017214277759194374, + 0.010607550852000713, + 0.010266530327498913, + 0.012544851750135422, + 0.006078769452869892, + 0.005807369016110897, + 0.008363994769752026, + 0.007648770231753588, + 0.007215441204607487, + 0.01450705248862505, + 0.0036715539172291756, + 0.004925466608256102, + 0.0034598063211888075, + 0.00415112404152751, + 0.0048952531069517136, + 0.0038379535544663668, + 0.01316804625093937, + 0.07857680320739746, + 0.007686725351959467, + 0.015128467231988907, + 0.00795770063996315, + 0.007938306778669357, + 0.008814326487481594, + 0.00854591466486454, + 0.07154907286167145, + 0.13728167116641998, + 0.010319177061319351, + 0.018132362514734268, + 0.011048529297113419, + 0.00962052308022976, + 0.0154883349314332, + 0.010089907795190811, + 0.12165603041648865, + 0.0110611692070961, + 0.005240504164248705, + 0.006498309317976236, + 0.005180342122912407, + 0.006844014395028353, + 0.005921426694840193, + 0.0057607414200901985, + 0.008761904202401638, + 0.003950560465455055, + 0.0023848689161241055, + 0.0021976560819894075, + 0.0017955524381250143, + 0.0021695573814213276, + 0.0036429695319384336, + 0.0017751825507730246, + 0.0027322275564074516, + 0.01240493543446064, + 0.004732078872621059, + 0.01290446612983942, + 0.007251111324876547, + 0.005583095829933882, + 0.007638493552803993, + 0.0067677805200219154, + 0.00872094463557005, + 0.008632777258753777, + 0.0028809423092752695, + 0.006216179113835096, + 0.0043979063630104065, + 0.0033978198189288378, + 0.004105018451809883, + 0.003832170506939292, + 0.005215478129684925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -1.6005513668060303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004655119031667709 + ], + "opposite_logprob_tokens": [ + -5.379654884338379 + ], + "logprob_reward": -0.004655119031667709, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004655119031667709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13702580332756042, + 0.03813793137669563, + 0.07046220451593399, + 0.42190271615982056, + 0.09413352608680725, + 0.07492167502641678, + 0.06869911402463913, + 0.03100859560072422, + 0.03956666961312294, + 0.02414178103208542 + ], + "credit_map": [ + 0.018924573436379433, + 0.007051129825413227, + 0.024469666182994843, + 0.010507668368518353, + 0.009844647720456123, + 0.011247695423662663, + 0.019726110622286797, + 0.03525431826710701, + 0.005973100662231445, + 0.003529380541294813, + 0.005519674625247717, + 0.003229281632229686, + 0.004040815867483616, + 0.0044162692502141, + 0.004084615968167782, + 0.0073447939939796925, + 0.012007013894617558, + 0.005936854053288698, + 0.023429859429597855, + 0.005705459974706173, + 0.0029086759313941, + 0.004927628207951784, + 0.006173446774482727, + 0.009373266249895096, + 0.14435496926307678, + 0.01667701080441475, + 0.0299370177090168, + 0.01620499975979328, + 0.010921483859419823, + 0.020721860229969025, + 0.020609237253665924, + 0.16247618198394775, + 0.02757491171360016, + 0.0037903510965406895, + 0.009729254059493542, + 0.00504453806206584, + 0.007986197248101234, + 0.005534339230507612, + 0.005420126020908356, + 0.029053812846541405, + 0.029631298035383224, + 0.002784566953778267, + 0.004852496553212404, + 0.0026267641223967075, + 0.002373618073761463, + 0.003713218728080392, + 0.0025619249790906906, + 0.026377785950899124, + 0.012663536705076694, + 0.005967535078525543, + 0.009913056157529354, + 0.008731396868824959, + 0.007646187674254179, + 0.007260945159941912, + 0.006783226504921913, + 0.009733229875564575, + 0.00607982836663723, + 0.0024310084991157055, + 0.0071962191723287106, + 0.003316441783681512, + 0.0033440899569541216, + 0.0028157823253422976, + 0.0021249891724437475, + 0.0037002339959144592, + 0.0076942783780395985, + 0.002943714614957571, + 0.008539004251360893, + 0.0047355699352920055, + 0.002712161047384143, + 0.004529914353042841, + 0.0036491260398179293, + 0.004762902855873108, + 0.005137150175869465, + 0.0021298560313880444, + 0.002950878581032157, + 0.002348046051338315, + 0.0024449964985251427, + 0.0031311996281147003, + 0.002585026668384671, + 0.003414624370634556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.004655119031667709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004838436841965 + ], + "opposite_logprob_tokens": [ + -2.6950483322143555 + ], + "logprob_reward": -0.07004838436841965, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004838436841965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10496112704277039, + 0.04279877990484238, + 0.06204001232981682, + 0.03744729235768318, + 0.11595037579536438, + 0.28996121883392334, + 0.13550542294979095, + 0.07648232579231262, + 0.06465684622526169, + 0.07019665092229843 + ], + "credit_map": [ + 0.01235645916312933, + 0.006194962654262781, + 0.015425569377839565, + 0.011137948371469975, + 0.009407097473740578, + 0.009858563542366028, + 0.014224907383322716, + 0.026355605572462082, + 0.006777990609407425, + 0.004835609346628189, + 0.007896563969552517, + 0.003139253007248044, + 0.004239690024405718, + 0.0066438838839530945, + 0.0041588288731873035, + 0.005106955301016569, + 0.009476044215261936, + 0.00881623849272728, + 0.011866732500493526, + 0.005320616532117128, + 0.005152130965143442, + 0.008198922500014305, + 0.00675475038588047, + 0.006454573478549719, + 0.005483250133693218, + 0.002864348003640771, + 0.005416783969849348, + 0.004338241182267666, + 0.004882823210209608, + 0.006126579362899065, + 0.0036212445702403784, + 0.004714018199592829, + 0.03748613968491554, + 0.005237499251961708, + 0.009730623103678226, + 0.005433846730738878, + 0.008137230761349201, + 0.012301607988774776, + 0.005399841815233231, + 0.03222358226776123, + 0.10149848461151123, + 0.01220292504876852, + 0.014687342569231987, + 0.01663919724524021, + 0.014492833986878395, + 0.021562300622463226, + 0.016102828085422516, + 0.09277528524398804, + 0.018347058445215225, + 0.010841547511518002, + 0.024508144706487656, + 0.022618576884269714, + 0.01314688753336668, + 0.014826393686234951, + 0.01536340918391943, + 0.01585339941084385, + 0.01530180312693119, + 0.007022416684776545, + 0.016480300575494766, + 0.008621872402727604, + 0.006931727286428213, + 0.007602248806506395, + 0.0055108461529016495, + 0.009011102840304375, + 0.012389098294079304, + 0.004418781492859125, + 0.014279813505709171, + 0.0073483469896018505, + 0.004874455742537975, + 0.006879126653075218, + 0.006307343952357769, + 0.008159870281815529, + 0.014128812588751316, + 0.005405540112406015, + 0.011701383627951145, + 0.007445399183779955, + 0.007643456105142832, + 0.007464381400495768, + 0.007655572611838579, + 0.008752100169658661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.07004838436841965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000816131301689893 + ], + "opposite_logprob_tokens": [ + -7.125816345214844 + ], + "logprob_reward": -0.000816131301689893, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.000816131301689893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10597588866949081, + 0.1200622096657753, + 0.04548785090446472, + 0.10812343657016754, + 0.24912898242473602, + 0.11644736677408218, + 0.0537455789744854, + 0.05860229581594467, + 0.08370288461446762, + 0.058723531663417816 + ], + "credit_map": [ + 0.018961891531944275, + 0.00834178552031517, + 0.020086869597434998, + 0.014876027591526508, + 0.007479356601834297, + 0.00992454495280981, + 0.013446937315165997, + 0.012858466245234013, + 0.01772136613726616, + 0.009084228426218033, + 0.02974608540534973, + 0.007929747924208641, + 0.008718841709196568, + 0.011607160791754723, + 0.01095926109701395, + 0.024295514449477196, + 0.0072041526436805725, + 0.005541255697607994, + 0.006676572374999523, + 0.004429007414728403, + 0.003977329004555941, + 0.00844782404601574, + 0.005248010158538818, + 0.003963698633015156, + 0.03126616030931473, + 0.005930752493441105, + 0.010865340009331703, + 0.006185495760291815, + 0.0044234939850866795, + 0.008202354423701763, + 0.008152149617671967, + 0.03309767693281174, + 0.09499625116586685, + 0.00913217943161726, + 0.01719001680612564, + 0.011427425779402256, + 0.010225352831184864, + 0.011542433872818947, + 0.011323278769850731, + 0.08329201489686966, + 0.04221341758966446, + 0.005880153737962246, + 0.008897464722394943, + 0.005283175967633724, + 0.006232853047549725, + 0.006136367563158274, + 0.005436984822154045, + 0.03636693209409714, + 0.009767607785761356, + 0.0031150414142757654, + 0.01081900205463171, + 0.005586550571024418, + 0.0060590519569814205, + 0.006110882852226496, + 0.00565206166356802, + 0.006635376252233982, + 0.011800396256148815, + 0.004433717578649521, + 0.013193062506616116, + 0.006634276360273361, + 0.005044559482485056, + 0.006608685478568077, + 0.004007676616311073, + 0.006879914551973343, + 0.017640313133597374, + 0.006499639246612787, + 0.011413044296205044, + 0.009294000454246998, + 0.007976316846907139, + 0.012315632775425911, + 0.008053552359342575, + 0.010510376654565334, + 0.012082619592547417, + 0.005047174170613289, + 0.008319981396198273, + 0.006689927075058222, + 0.004794348031282425, + 0.007200318854302168, + 0.006540013011544943, + 0.008049143478274345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.000816131301689893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the tender yet momentous exchange of an exquisite antique silver ring. The camera remains perfectly still, focused on two pairs of hands suspended above a dark, out-of-focus wooden surface. The giving hand, showing the fine lines and gentle wrinkles of age, reverently holds the ring between thumb and index finger. The receiving hand, youthful and smooth, is slightly open and upward, fingers poised to receive the precious object. The ring itself is a work of intricate filigree, its silver band bearing a soft, time-worn patina in the crevices, contrasting strikingly with the polished surface as it gleams as it catches a single, warm, directional light source. In slow, deliberate micro-movements, the older hand steadily moves forward, placing the ring into the palm of the younger hand, the metal making a soft, almost imperceptible contact. The entire scene is bathed in warm, low-key lighting, creating deep, soft shadows that accentuate the textures of the skin and the uneven metal, evoking a tranquil and profound atmosphere of inheritance.", + "seed": 24841, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_8/index_8822_seed_24841_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": -0.26795938272261993, + "vlm_logprob": -0.26795938272261993, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.26795938272261993, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01973610557615757 + ], + "opposite_logprob_tokens": [ + -4.769736289978027 + ], + "logprob_reward": -0.01973610557615757, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01973610557615757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12491538375616074, + 0.07751011848449707, + 0.05685010179877281, + 0.0355168916285038, + 0.13701091706752777, + 0.03077693283557892, + 0.1207631379365921, + 0.08553346246480942, + 0.2288535088300705, + 0.1022694855928421 + ], + "credit_map": [ + 0.024735404178500175, + 0.012938599102199078, + 0.025516599416732788, + 0.012477683834731579, + 0.007823927327990532, + 0.015522176399827003, + 0.014317373745143414, + 0.01158362440764904, + 0.009265461005270481, + 0.006307889707386494, + 0.013774212449789047, + 0.011799948289990425, + 0.00871832761913538, + 0.01230911910533905, + 0.01045583002269268, + 0.00487933587282896, + 0.008971347473561764, + 0.0057874880731105804, + 0.009822206571698189, + 0.006726549938321114, + 0.004796190652996302, + 0.006953221745789051, + 0.006325042340904474, + 0.0074680582620203495, + 0.0036981599405407906, + 0.004619105719029903, + 0.004622356500476599, + 0.0072893970645964146, + 0.0025984931271523237, + 0.0038424108643084764, + 0.004675114527344704, + 0.004171857610344887, + 0.01421067863702774, + 0.016744451597332954, + 0.01989985816180706, + 0.037280891090631485, + 0.008322697132825851, + 0.013194059021770954, + 0.01488462369889021, + 0.012473671697080135, + 0.004612644203007221, + 0.0036694537848234177, + 0.004671299830079079, + 0.0035482491366565228, + 0.0029577359091490507, + 0.005629148334264755, + 0.002596561098471284, + 0.0030918398406356573, + 0.016604874283075333, + 0.014047068543732166, + 0.018529601395130157, + 0.027137264609336853, + 0.0086679607629776, + 0.013026832602918148, + 0.011203363537788391, + 0.0115461815148592, + 0.01115608960390091, + 0.00932516623288393, + 0.012430076487362385, + 0.01933489367365837, + 0.007584984879940748, + 0.010114594362676144, + 0.00871643703430891, + 0.006871229037642479, + 0.03375106677412987, + 0.032747432589530945, + 0.03291860222816467, + 0.03341710567474365, + 0.021752508357167244, + 0.025363896042108536, + 0.01947655715048313, + 0.029426343739032745, + 0.015804467722773552, + 0.011767419055104256, + 0.01481957919895649, + 0.014936864376068115, + 0.010746647603809834, + 0.013113360852003098, + 0.008710169233381748, + 0.012370985001325607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.01973610557615757 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370744705200195 + ], + "opposite_logprob_tokens": [ + -0.38707447052001953 + ], + "logprob_reward": -1.1370744705200195, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370744705200195, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08615092188119888, + 0.04063102975487709, + 0.028324473649263382, + 0.03486571088433266, + 0.0186166912317276, + 0.05536466836929321, + 0.07897263765335083, + 0.06981313228607178, + 0.4021640121936798, + 0.18509668111801147 + ], + "credit_map": [ + 0.01537881139665842, + 0.010184867307543755, + 0.013523124158382416, + 0.012127957306802273, + 0.005663348827511072, + 0.009679697453975677, + 0.010943891480565071, + 0.008649233728647232, + 0.006333714351058006, + 0.0048171160742640495, + 0.005519608501344919, + 0.005825038533657789, + 0.0033841505646705627, + 0.004422494675964117, + 0.006080555729568005, + 0.004248349461704493, + 0.0034224821720272303, + 0.004140936769545078, + 0.003973611164838076, + 0.003654939355328679, + 0.0026499303057789803, + 0.005225250963121653, + 0.0033487407490611076, + 0.0019085848471149802, + 0.005147023592144251, + 0.004929241258651018, + 0.004428553394973278, + 0.00511889299377799, + 0.0023837019689381123, + 0.004655336961150169, + 0.003555362345650792, + 0.004647598601877689, + 0.002455636626109481, + 0.0025208485312759876, + 0.002821056405082345, + 0.0035716905258595943, + 0.0013373569818213582, + 0.0021981517784297466, + 0.0016094823367893696, + 0.002102471189573407, + 0.005955927539616823, + 0.005887177772819996, + 0.008125856518745422, + 0.01232253946363926, + 0.004950493574142456, + 0.006406433880329132, + 0.005953263025730848, + 0.005762980319559574, + 0.010287699289619923, + 0.00980573333799839, + 0.011145143769681454, + 0.017876891419291496, + 0.00504644587635994, + 0.008733016438782215, + 0.006903993897140026, + 0.00917372852563858, + 0.010041991248726845, + 0.006260761525481939, + 0.009042618796229362, + 0.014929247088730335, + 0.0070933266542851925, + 0.009840860962867737, + 0.007281918078660965, + 0.005322417244315147, + 0.06616441905498505, + 0.05880872532725334, + 0.05894457548856735, + 0.05399712175130844, + 0.03799604997038841, + 0.04766087606549263, + 0.030317237600684166, + 0.04827503114938736, + 0.027661770582199097, + 0.021924858912825584, + 0.026003867387771606, + 0.027246570214629173, + 0.021770978346467018, + 0.024044401943683624, + 0.015583237633109093, + 0.02086099237203598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -1.1370744705200195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024953915271908045 + ], + "opposite_logprob_tokens": [ + -6.002495288848877 + ], + "logprob_reward": -0.0024953915271908045, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024953915271908045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06730975955724716, + 0.1655222475528717, + 0.055915553122758865, + 0.028218429535627365, + 0.05631648004055023, + 0.028064679354429245, + 0.07681156694889069, + 0.14395642280578613, + 0.16311341524124146, + 0.21477141976356506 + ], + "credit_map": [ + 0.01243223249912262, + 0.008145549334585667, + 0.00992677640169859, + 0.009844467043876648, + 0.0050464957021176815, + 0.007750316988676786, + 0.008733722381293774, + 0.005430201999843121, + 0.032770562916994095, + 0.0175319891422987, + 0.022459227591753006, + 0.022841298952698708, + 0.01294154953211546, + 0.018145069479942322, + 0.022470979019999504, + 0.016361579298973083, + 0.012080342508852482, + 0.007558034732937813, + 0.007161451503634453, + 0.007384550757706165, + 0.004206074867397547, + 0.007118498906493187, + 0.005900462623685598, + 0.004506135359406471, + 0.005169795826077461, + 0.0027592158876359463, + 0.003924464341253042, + 0.0055480538867414, + 0.0018775186035782099, + 0.0033264129888266325, + 0.0032015908509492874, + 0.002411377616226673, + 0.0074080815538764, + 0.006194752641022205, + 0.007742681540548801, + 0.015206724405288696, + 0.0030181207694113255, + 0.006569708231836557, + 0.005513164680451155, + 0.004663246218115091, + 0.0026960279792547226, + 0.003765228670090437, + 0.004160403739660978, + 0.008134803734719753, + 0.0017159318085759878, + 0.0028569309506565332, + 0.00262259179726243, + 0.0021127613727003336, + 0.009232593700289726, + 0.0088077737018466, + 0.01157976035028696, + 0.01977921836078167, + 0.00539300125092268, + 0.007550379261374474, + 0.007545123808085918, + 0.006923721171915531, + 0.01696253940463066, + 0.016353929415345192, + 0.020050298422574997, + 0.03770025074481964, + 0.009772456251084805, + 0.018776841461658478, + 0.011968104168772697, + 0.012372001074254513, + 0.024996964260935783, + 0.0228124912828207, + 0.02389056608080864, + 0.022674275562167168, + 0.01538588386029005, + 0.018829144537448883, + 0.013429216109216213, + 0.021094877272844315, + 0.03564400225877762, + 0.027062512934207916, + 0.03066864237189293, + 0.029603907838463783, + 0.025112411007285118, + 0.02480919472873211, + 0.017881356179714203, + 0.023989377543330193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.0024953915271908045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992760241031647 + ], + "opposite_logprob_tokens": [ + -1.8049275875091553 + ], + "logprob_reward": -0.17992760241031647, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992760241031647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11527062207460403, + 0.09317845106124878, + 0.09557522088289261, + 0.031142912805080414, + 0.05422132834792137, + 0.04119446873664856, + 0.026544421911239624, + 0.13382333517074585, + 0.15608879923820496, + 0.2529604136943817 + ], + "credit_map": [ + 0.019558565691113472, + 0.010809570550918579, + 0.01892864517867565, + 0.016376253217458725, + 0.006244702730327845, + 0.012546445243060589, + 0.015480744652450085, + 0.015325701795518398, + 0.010637063533067703, + 0.008438809774816036, + 0.016942763701081276, + 0.01627831533551216, + 0.007163596339523792, + 0.011135569773614407, + 0.012740770354866982, + 0.009841573424637318, + 0.015403853729367256, + 0.00932730920612812, + 0.014166908338665962, + 0.011193416081368923, + 0.007096117362380028, + 0.015476052649319172, + 0.011736275628209114, + 0.011175286956131458, + 0.004999375902116299, + 0.0040866476483643055, + 0.004477696493268013, + 0.004731043707579374, + 0.0020773939322680235, + 0.0035679922439157963, + 0.002971754875034094, + 0.004231007304042578, + 0.008230291306972504, + 0.0069329445250332355, + 0.007951146923005581, + 0.009647082537412643, + 0.0035890548024326563, + 0.006357616279274225, + 0.005249544512480497, + 0.006263643968850374, + 0.004846273921430111, + 0.005799982231110334, + 0.005971562582999468, + 0.00904795341193676, + 0.003158814739435911, + 0.004609744530171156, + 0.003515336662530899, + 0.004244802985340357, + 0.0033391339238733053, + 0.002671134425327182, + 0.0038364273495972157, + 0.004764535930007696, + 0.00236919685266912, + 0.0036569652147591114, + 0.003159171435981989, + 0.0027478565461933613, + 0.01683712936937809, + 0.014973325654864311, + 0.01744326576590538, + 0.029647566378116608, + 0.013017033226788044, + 0.017233893275260925, + 0.011813412420451641, + 0.012857703492045403, + 0.03028181940317154, + 0.02191251516342163, + 0.02421969175338745, + 0.019123056903481483, + 0.013879629783332348, + 0.01963873766362667, + 0.01049025822430849, + 0.016543088480830193, + 0.03730906546115875, + 0.02940197102725506, + 0.03682385012507439, + 0.03663868457078934, + 0.029415525496006012, + 0.03353789821267128, + 0.02062327042222023, + 0.029210137203335762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.17992760241031647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focused on two pairs of hands and a silver ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, out-of-focus wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the older hand move forward to place the ring into the palm of the younger hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, low-key lighting with deep, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005633435794152319 + ], + "opposite_logprob_tokens": [ + -7.500563144683838 + ], + "logprob_reward": -0.0005633435794152319, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005633435794152319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21321642398834229, + 0.10388105362653732, + 0.04628152772784233, + 0.048154931515455246, + 0.03397093713283539, + 0.0428769625723362, + 0.053196340799331665, + 0.18747015297412872, + 0.1857082098722458, + 0.08524347096681595 + ], + "credit_map": [ + 0.03741860017180443, + 0.023640478029847145, + 0.043524328619241714, + 0.021136658266186714, + 0.017485279589891434, + 0.02717885933816433, + 0.023075006902217865, + 0.019757216796278954, + 0.020205872133374214, + 0.009422375820577145, + 0.023193150758743286, + 0.008853989653289318, + 0.006010811310261488, + 0.015359843149781227, + 0.011094113811850548, + 0.00974089466035366, + 0.010864230804145336, + 0.003912824206054211, + 0.007014711853116751, + 0.0065073478035628796, + 0.002732878550887108, + 0.006288505159318447, + 0.005558553617447615, + 0.0034024734050035477, + 0.007567832246422768, + 0.005364384036511183, + 0.006751398555934429, + 0.0075355046428740025, + 0.0047221714630723, + 0.00611945241689682, + 0.005267383996397257, + 0.004826802760362625, + 0.007027717307209969, + 0.004081570077687502, + 0.004710495006293058, + 0.005349441897124052, + 0.0026460662484169006, + 0.0034701963886618614, + 0.003674199106171727, + 0.003011250402778387, + 0.005973647348582745, + 0.004569574724882841, + 0.007403246127068996, + 0.008178308606147766, + 0.0028556897304952145, + 0.005836192052811384, + 0.003948251251131296, + 0.004112048540264368, + 0.008082333952188492, + 0.005466891918331385, + 0.00844813697040081, + 0.01019305270165205, + 0.004777656868100166, + 0.006560534238815308, + 0.005630644038319588, + 0.004037085920572281, + 0.03543813154101372, + 0.01843889057636261, + 0.026358196511864662, + 0.03983232378959656, + 0.014596150256693363, + 0.020555900409817696, + 0.015001452527940273, + 0.017249109223484993, + 0.0302541833370924, + 0.026297220960259438, + 0.027301711961627007, + 0.025470120832324028, + 0.01696023903787136, + 0.021373605355620384, + 0.014706387184560299, + 0.023344745859503746, + 0.012271847575902939, + 0.009197913110256195, + 0.012795863673090935, + 0.012546270154416561, + 0.010412917472422123, + 0.010206270962953568, + 0.0072639212012290955, + 0.01054846215993166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0005633435794152319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_9/index_tensor([4032, 4032])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_9/index_tensor([4032, 4032])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7b712e59bdec51c7064e373a25918c6405242eaa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_9/index_tensor([4032, 4032])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, static shot, shot from a slightly upward angle, captures a chubby, adorable giant panda. It stands in a fairytale-like kitchen, wearing a white chef's hat tilted slightly to one side and a red and white checkered apron tied around its waist. A warm, honey-sweet light slants in from the window on the right side of the frame, illuminating its contented smile. Its small eyes, narrowed blissfully beneath dark circles, and its mouth, slightly raised, cast a golden halo over the fluffy fur on its cheeks. Concentrated over a rustic wood stove, it carefully manipulates the stove with its furry paws: one firmly grasping the handle of a steaming cast iron pot, the other holding a pair of long wooden chopsticks, gently stirring the bubbling milky broth and golden noodles. The rising steam partially obscured its face, making the entire scene seem particularly dreamy. Against the backdrop of a warm cream-colored wall, strings of dried chilies and braided garlic hung from it, and a few brightly colored ceramic jars were placed on a wooden rack nearby. The entire scene was filled with a warm and soothing atmosphere, overflowing with the love of food and the pure joy of it.", + "seed": 13138, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_9/index_4032_seed_13138_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": -0.04882595986127854, + "vlm_logprob": -0.04882595986127854, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04882595986127854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.004824426025152206 + ], + "opposite_logprob_tokens": [ + -5.504824638366699 + ], + "logprob_reward": -0.004824426025152206, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004824426025152206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30271098017692566, + 0.06334667652845383, + 0.03222038224339485, + 0.06533759832382202, + 0.32520100474357605, + 0.020138755440711975, + 0.032535720616579056, + 0.018648335710167885, + 0.030162421986460686, + 0.1096981093287468 + ], + "credit_map": [ + 0.05484725162386894, + 0.05283091962337494, + 0.02376709133386612, + 0.02670900709927082, + 0.03804074600338936, + 0.03446483984589577, + 0.046308647841215134, + 0.025742482393980026, + 0.009880119003355503, + 0.014833077788352966, + 0.006698880344629288, + 0.0059187342412769794, + 0.007670943159610033, + 0.006629082839936018, + 0.006862555164843798, + 0.004853290971368551, + 0.004807492718100548, + 0.003825277555733919, + 0.003947132267057896, + 0.003286086954176426, + 0.004549502395093441, + 0.0034595343749970198, + 0.003821623744443059, + 0.004523734096437693, + 0.00764984218403697, + 0.013502776622772217, + 0.01054265908896923, + 0.006017036270350218, + 0.008431038819253445, + 0.006687423679977655, + 0.004306145943701267, + 0.008200681768357754, + 0.03941522166132927, + 0.0840233862400055, + 0.04622400552034378, + 0.03068014420568943, + 0.041343871504068375, + 0.030705399811267853, + 0.023353245109319687, + 0.02945576421916485, + 0.0038367025554180145, + 0.002592125441879034, + 0.0018132078694179654, + 0.0023259464651346207, + 0.002620196435600519, + 0.00207303068600595, + 0.0018257973715662956, + 0.0030517498962581158, + 0.004682027734816074, + 0.005326343234628439, + 0.0022360810544341803, + 0.00300140050239861, + 0.0035593921784311533, + 0.0028747329488396645, + 0.006921596359461546, + 0.003934148233383894, + 0.0021808713208884, + 0.0033714997116476297, + 0.0018505039624869823, + 0.0018293041503056884, + 0.002593228593468666, + 0.0021848755422979593, + 0.002765193348750472, + 0.001872860244475305, + 0.005004066973924637, + 0.00588305713608861, + 0.0029489672742784023, + 0.0027203925419598818, + 0.003685442730784416, + 0.0036116279661655426, + 0.003535142168402672, + 0.002773726126179099, + 0.02318429946899414, + 0.01718885451555252, + 0.01174437627196312, + 0.012710588984191418, + 0.00952508021146059, + 0.011183598078787327, + 0.015439649112522602, + 0.008721674792468548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.004824426025152206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.026355832815170288 + ], + "opposite_logprob_tokens": [ + -3.651355743408203 + ], + "logprob_reward": -0.026355832815170288, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355832815170288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3109871447086334, + 0.1311357468366623, + 0.17898088693618774, + 0.04980031028389931, + 0.037524983286857605, + 0.08797241002321243, + 0.05395129695534706, + 0.08685237169265747, + 0.03434401750564575, + 0.02845078706741333 + ], + "credit_map": [ + 0.06286882609128952, + 0.050587791949510574, + 0.02554696425795555, + 0.028350168839097023, + 0.038695450872182846, + 0.03288139030337334, + 0.04263441637158394, + 0.029422128573060036, + 0.02323916181921959, + 0.02579330839216709, + 0.012507728300988674, + 0.009042217396199703, + 0.020257288590073586, + 0.015215892344713211, + 0.013720561750233173, + 0.011359588243067265, + 0.02767801098525524, + 0.03476235643029213, + 0.02239440754055977, + 0.014988507144153118, + 0.021014651283621788, + 0.01638719253242016, + 0.013394709676504135, + 0.028361057862639427, + 0.006592409685254097, + 0.009126284159719944, + 0.004937619902193546, + 0.0035866270773112774, + 0.004146404564380646, + 0.005259096156805754, + 0.004555156920105219, + 0.01159671414643526, + 0.00559628801420331, + 0.005497541278600693, + 0.004579878877848387, + 0.0036116144619882107, + 0.005251930095255375, + 0.004471979569643736, + 0.004552587401121855, + 0.003963163122534752, + 0.011495412327349186, + 0.020144375041127205, + 0.008284994401037693, + 0.007702309172600508, + 0.010817068628966808, + 0.008780351839959621, + 0.005885254591703415, + 0.01486264355480671, + 0.011822309345006943, + 0.007463525049388409, + 0.005109014455229044, + 0.0052468134090304375, + 0.005228014662861824, + 0.0049390243366360664, + 0.008663173764944077, + 0.005479424726217985, + 0.009350770153105259, + 0.01667999103665352, + 0.006071043200790882, + 0.012056048028171062, + 0.014404731802642345, + 0.009828687645494938, + 0.008974304422736168, + 0.009486790746450424, + 0.004089812748134136, + 0.006649799179285765, + 0.003912009764462709, + 0.004120875149965286, + 0.0038665437605232, + 0.0038460742216557264, + 0.004131308291107416, + 0.003727594157680869, + 0.00598090048879385, + 0.005196043290197849, + 0.002812782535329461, + 0.004297025967389345, + 0.0024678041227161884, + 0.0028661617543548346, + 0.0022935958113521338, + 0.0025364758912473917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.026355832815170288 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.033700522035360336 + ], + "opposite_logprob_tokens": [ + -3.408700466156006 + ], + "logprob_reward": -0.033700522035360336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033700522035360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08365601301193237, + 0.16486307978630066, + 0.03696718439459801, + 0.07105003297328949, + 0.2944176495075226, + 0.08419515937566757, + 0.10314028710126877, + 0.11352231353521347, + 0.035163335502147675, + 0.013024994172155857 + ], + "credit_map": [ + 0.00870281457901001, + 0.01123588252812624, + 0.009127666242420673, + 0.009854601696133614, + 0.011910687200725079, + 0.00965343788266182, + 0.012702921405434608, + 0.01046799961477518, + 0.017513994127511978, + 0.022914359346032143, + 0.013070622459053993, + 0.015451209619641304, + 0.03491053357720375, + 0.023222103714942932, + 0.021077174693346024, + 0.016703085973858833, + 0.005723930895328522, + 0.00552542507648468, + 0.005949047394096851, + 0.0021884581074118614, + 0.005470287054777145, + 0.0032666195183992386, + 0.00419322494417429, + 0.004650191869586706, + 0.008674101904034615, + 0.0066076889634132385, + 0.006695966701954603, + 0.005823497660458088, + 0.007319408468902111, + 0.007974609732627869, + 0.008838516660034657, + 0.019116243347525597, + 0.031048880890011787, + 0.038666173815727234, + 0.02007710561156273, + 0.026934580877423286, + 0.026328125968575478, + 0.033962432295084, + 0.054909270256757736, + 0.062491074204444885, + 0.01602940820157528, + 0.016434116289019585, + 0.01189336832612753, + 0.006571615114808083, + 0.009193602949380875, + 0.007699218578636646, + 0.006179191637784243, + 0.010194633156061172, + 0.011688508093357086, + 0.016733013093471527, + 0.007124133408069611, + 0.008892955258488655, + 0.015590299852192402, + 0.011732494458556175, + 0.01990131288766861, + 0.01147757563740015, + 0.012182568199932575, + 0.017140425741672516, + 0.009127261117100716, + 0.011708861216902733, + 0.017295490950345993, + 0.012911579571664333, + 0.01669963076710701, + 0.016456488519906998, + 0.006508619524538517, + 0.0042136916890740395, + 0.005594530142843723, + 0.0037420301232486963, + 0.0035871390718966722, + 0.003650410333648324, + 0.0038084674160927534, + 0.0040584467351436615, + 0.001741491723805666, + 0.0023291476536542177, + 0.0011633782414719462, + 0.001643164549022913, + 0.0011877105571329594, + 0.001408173469826579, + 0.002007321920245886, + 0.0015446060569956899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.033700522035360336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.07897115498781204 + ], + "opposite_logprob_tokens": [ + -2.5789711475372314 + ], + "logprob_reward": -0.07897115498781204, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897115498781204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20935457944869995, + 0.24783869087696075, + 0.08993851393461227, + 0.04460630565881729, + 0.09132403880357742, + 0.07510817795991898, + 0.054465461522340775, + 0.07426567375659943, + 0.05124128982424736, + 0.06185721978545189 + ], + "credit_map": [ + 0.020098133012652397, + 0.02122047357261181, + 0.014470907859504223, + 0.024059152230620384, + 0.019239643588662148, + 0.0250187236815691, + 0.056612905114889145, + 0.02863464131951332, + 0.03990115597844124, + 0.04072541370987892, + 0.02248835191130638, + 0.024016540497541428, + 0.04132106900215149, + 0.028969934210181236, + 0.02876184694468975, + 0.021654367446899414, + 0.01782332733273506, + 0.008342196233570576, + 0.01613418571650982, + 0.008019897155463696, + 0.011427904479205608, + 0.009091731160879135, + 0.007181773893535137, + 0.011917498894035816, + 0.007493981625884771, + 0.005326945334672928, + 0.005638845730572939, + 0.0027310913428664207, + 0.005096278619021177, + 0.004865003749728203, + 0.004149574786424637, + 0.009304587729275227, + 0.012196877039968967, + 0.014086424373090267, + 0.00976577028632164, + 0.009733698330819607, + 0.00799246784299612, + 0.009479648433625698, + 0.014346415176987648, + 0.0137227363884449, + 0.012182776816189289, + 0.018031055107712746, + 0.010782580822706223, + 0.00624156603589654, + 0.007877664640545845, + 0.006237595342099667, + 0.005087052937597036, + 0.008667879737913609, + 0.008946986868977547, + 0.009712700732052326, + 0.005208900198340416, + 0.006771032232791185, + 0.007148210424929857, + 0.006098825950175524, + 0.005797583609819412, + 0.0047812191769480705, + 0.00744039798155427, + 0.01233577262610197, + 0.007779343985021114, + 0.006932542659342289, + 0.009283103048801422, + 0.009607146494090557, + 0.008810537867248058, + 0.012076830491423607, + 0.005933435168117285, + 0.010212038643658161, + 0.005808784160763025, + 0.006750880740582943, + 0.0066831219010055065, + 0.006092539522796869, + 0.005056189838796854, + 0.0047043017111718655, + 0.01358798984438181, + 0.008932018652558327, + 0.005950307473540306, + 0.006721528246998787, + 0.00500855129212141, + 0.006530513986945152, + 0.00826440379023552, + 0.006861907429993153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.07897115498781204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.1002778634428978 + ], + "opposite_logprob_tokens": [ + -2.350277900695801 + ], + "logprob_reward": -0.1002778634428978, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002778634428978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21984665095806122, + 0.10703343152999878, + 0.09688476473093033, + 0.07645370811223984, + 0.10721953958272934, + 0.03567180037498474, + 0.08698815107345581, + 0.10281942784786224, + 0.1019372045993805, + 0.06514523178339005 + ], + "credit_map": [ + 0.03412700816988945, + 0.043735697865486145, + 0.01984911970794201, + 0.016157733276486397, + 0.02684096433222294, + 0.023011762648820877, + 0.03346516191959381, + 0.022659216076135635, + 0.018975554034113884, + 0.018851077184081078, + 0.00910460390150547, + 0.011524655856192112, + 0.011630900204181671, + 0.010547034442424774, + 0.014751448296010494, + 0.011648167856037617, + 0.014482350088655949, + 0.018652116879820824, + 0.01638318970799446, + 0.009459647350013256, + 0.011276650242507458, + 0.00887581892311573, + 0.00840812362730503, + 0.009346876293420792, + 0.013539307750761509, + 0.011850784532725811, + 0.014673174358904362, + 0.005632915999740362, + 0.005791691597551107, + 0.006755194626748562, + 0.005198299419134855, + 0.013012347742915154, + 0.014094232581555843, + 0.020775163546204567, + 0.009451213292777538, + 0.009231846779584885, + 0.009983534924685955, + 0.011302580125629902, + 0.014542360790073872, + 0.017838621512055397, + 0.005244521424174309, + 0.003783486783504486, + 0.00464034266769886, + 0.0032862189691513777, + 0.003016349859535694, + 0.003712905803695321, + 0.003988949116319418, + 0.007999028079211712, + 0.010337536223232746, + 0.01452575996518135, + 0.007091319654136896, + 0.007263504434376955, + 0.013284941203892231, + 0.009760314598679543, + 0.015318922698497772, + 0.009405859746038914, + 0.022129539400339127, + 0.011003418825566769, + 0.009897406212985516, + 0.010840160772204399, + 0.010862872935831547, + 0.011317450553178787, + 0.01673494465649128, + 0.01003363449126482, + 0.0216751117259264, + 0.01428391132503748, + 0.012818659655749798, + 0.009076792746782303, + 0.010386037640273571, + 0.011739485897123814, + 0.011970284394919872, + 0.009986934252083302, + 0.013282286934554577, + 0.009166848845779896, + 0.006789004895836115, + 0.006858669687062502, + 0.005493551027029753, + 0.006825578864663839, + 0.009690387174487114, + 0.007038906682282686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.1002778634428978 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, captures a chubby, adorable giant panda. It stands in a fairytale-like kitchen, wearing a white chef's hat tilted slightly to one side and a red and white checkered apron tied around its waist. A warm, honey-sweet light slants in from the window on the right side of the frame, illuminating its contented smile. Its small eyes, narrowed blissfully beneath dark circles, and its mouth, slightly raised, cast a golden halo over the fluffy fur on its cheeks. Concentrated over a rustic wood stove, it carefully manipulates the stove with its furry paws: one firmly grasping the handle of a steaming cast iron pot, the other holding a pair of long wooden chopsticks, gently stirring the bubbling milky broth and golden noodles. The rising steam partially obscured its face, making the entire scene seem particularly dreamy. Against the backdrop of a warm cream-colored wall, strings of dried chilies and braided garlic hung from it, and a few brightly colored ceramic jars were placed on a wooden rack nearby. The entire scene was filled with a warm and soothing atmosphere, overflowing with the love of food and the pure joy of it.", + "seed": 13139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_9/index_4032_seed_13139_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": -0.04430866711772978, + "vlm_logprob": -0.04430866711772978, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.04430866711772978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.003538658143952489 + ], + "opposite_logprob_tokens": [ + -6.003538608551025 + ], + "logprob_reward": -0.003538658143952489, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.003538658143952489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15496909618377686, + 0.06721604615449905, + 0.03627669811248779, + 0.12105321884155273, + 0.18475160002708435, + 0.07395584881305695, + 0.16614140570163727, + 0.03382629528641701, + 0.07767738401889801, + 0.08413244783878326 + ], + "credit_map": [ + 0.023763153702020645, + 0.016084088012576103, + 0.012574966065585613, + 0.021179690957069397, + 0.013927166350185871, + 0.016812490299344063, + 0.026791315525770187, + 0.023836221545934677, + 0.010212493129074574, + 0.005328362807631493, + 0.010815761983394623, + 0.015885639935731888, + 0.004460164345800877, + 0.004237729124724865, + 0.008476321585476398, + 0.007799576036632061, + 0.0020905775018036366, + 0.0032188103068619967, + 0.003924743738025427, + 0.005348077975213528, + 0.00396624393761158, + 0.0045802113600075245, + 0.00845202337950468, + 0.004696009214967489, + 0.01277894340455532, + 0.012551117688417435, + 0.012654666788876057, + 0.017657604068517685, + 0.021892156451940536, + 0.012043542228639126, + 0.017792344093322754, + 0.013682844117283821, + 0.019456570968031883, + 0.014407536014914513, + 0.026573212817311287, + 0.0248212032020092, + 0.016321009024977684, + 0.017464859411120415, + 0.050351377576589584, + 0.01535581424832344, + 0.003399050794541836, + 0.0031008049845695496, + 0.0037654899060726166, + 0.013919204473495483, + 0.003900671610608697, + 0.009462549351155758, + 0.029413102194666862, + 0.006994970142841339, + 0.015803063288331032, + 0.013363811187446117, + 0.02694621868431568, + 0.02388927899301052, + 0.01374528557062149, + 0.008624977432191372, + 0.04954607039690018, + 0.01422269269824028, + 0.003276166506111622, + 0.003220195649191737, + 0.0036143953911960125, + 0.008689392358064651, + 0.0030468872282654047, + 0.0038747501093894243, + 0.0046152533032000065, + 0.003489254042506218, + 0.008425228297710419, + 0.006796357687562704, + 0.00722119677811861, + 0.011699553579092026, + 0.007465131115168333, + 0.013496615923941135, + 0.01352299191057682, + 0.009050309658050537, + 0.011152421124279499, + 0.005130873993039131, + 0.007251210510730743, + 0.01610608585178852, + 0.005321021191775799, + 0.014005945064127445, + 0.016602279618382454, + 0.008562606759369373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.003538658143952489 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.012544889003038406 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012544889003038406, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544889003038406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10577348619699478, + 0.08312496542930603, + 0.14893656969070435, + 0.06678003817796707, + 0.10773496329784393, + 0.06310565769672394, + 0.09523684531450272, + 0.04191840440034866, + 0.15455389022827148, + 0.13283517956733704 + ], + "credit_map": [ + 0.018124012276530266, + 0.012562138959765434, + 0.013430057093501091, + 0.01538100652396679, + 0.011370440945029259, + 0.007348799612373114, + 0.014601008966565132, + 0.012956028804183006, + 0.006647975649684668, + 0.008269171230494976, + 0.008517303504049778, + 0.01402534544467926, + 0.0076861013658344746, + 0.013507797382771969, + 0.016359537839889526, + 0.008111739531159401, + 0.017122091725468636, + 0.01443346869200468, + 0.011760328896343708, + 0.018530556932091713, + 0.021071679890155792, + 0.0250768531113863, + 0.02047470398247242, + 0.020466888323426247, + 0.009937824681401253, + 0.004315968137234449, + 0.014353764243423939, + 0.007258701138198376, + 0.005020200274884701, + 0.008875884115695953, + 0.00975733157247305, + 0.007260361220687628, + 0.010161892510950565, + 0.006465771235525608, + 0.021278563886880875, + 0.014506927691400051, + 0.008841061033308506, + 0.006712186615914106, + 0.030867064371705055, + 0.008901496417820454, + 0.007328195031732321, + 0.0032472158782184124, + 0.0037909408565610647, + 0.010104124434292316, + 0.0037367139011621475, + 0.007386550772935152, + 0.021053722128272057, + 0.0064581939950585365, + 0.007402878720313311, + 0.006761498749256134, + 0.01260998472571373, + 0.020390527322888374, + 0.007636627182364464, + 0.008258203975856304, + 0.0243056770414114, + 0.007871445268392563, + 0.004864636808633804, + 0.003715472063049674, + 0.005349327344447374, + 0.007596235256642103, + 0.003986653406172991, + 0.004286884795874357, + 0.007005154155194759, + 0.005114040803164244, + 0.01691897213459015, + 0.010635478422045708, + 0.009906791150569916, + 0.027508506551384926, + 0.012934297323226929, + 0.01502291765064001, + 0.04219508543610573, + 0.019431831315159798, + 0.009974823333323002, + 0.007932601496577263, + 0.008611715398728848, + 0.04450678080320358, + 0.006995689123868942, + 0.020366190001368523, + 0.02434353157877922, + 0.010103839449584484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.012544889003038406 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.0298001728951931 + ], + "opposite_logprob_tokens": [ + -3.5298001766204834 + ], + "logprob_reward": -0.0298001728951931, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298001728951931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10087142139673233, + 0.08816605806350708, + 0.03547694906592369, + 0.08885113894939423, + 0.12636949121952057, + 0.04782810062170029, + 0.20755426585674286, + 0.08637728542089462, + 0.10893642157316208, + 0.10956881195306778 + ], + "credit_map": [ + 0.020221678540110588, + 0.010401252657175064, + 0.011149194091558456, + 0.0104141253978014, + 0.008343261666595936, + 0.00908550713211298, + 0.0163109190762043, + 0.014945486560463905, + 0.013236881233751774, + 0.006847498472779989, + 0.007706349715590477, + 0.013225274160504341, + 0.004755382891744375, + 0.011434828862547874, + 0.021086527034640312, + 0.009873319417238235, + 0.003454433288425207, + 0.0030712720472365618, + 0.00181508285459131, + 0.004165387712419033, + 0.0037481137551367283, + 0.006652427837252617, + 0.008758805692195892, + 0.0038114264607429504, + 0.009166110306978226, + 0.009361936710774899, + 0.0074072969146072865, + 0.016466481611132622, + 0.014284285716712475, + 0.006926908623427153, + 0.015394061803817749, + 0.009844052605330944, + 0.013240581378340721, + 0.008412067778408527, + 0.02045395039021969, + 0.01615775190293789, + 0.011352203786373138, + 0.01160938385874033, + 0.0339730829000473, + 0.011170471087098122, + 0.003952324390411377, + 0.0031021570321172476, + 0.0027321584057062864, + 0.00708566652610898, + 0.0032064029946923256, + 0.008192524313926697, + 0.015525281429290771, + 0.004031586926430464, + 0.02124760113656521, + 0.016625506803393364, + 0.024775531142950058, + 0.030894435942173004, + 0.017125438898801804, + 0.01133755873888731, + 0.06365807354450226, + 0.021890118718147278, + 0.008287017233669758, + 0.00692377844825387, + 0.01039163675159216, + 0.014595817774534225, + 0.010276724584400654, + 0.00678951945155859, + 0.019476482644677162, + 0.00963631272315979, + 0.013891843147575855, + 0.007017690222710371, + 0.0077515277080237865, + 0.01999395526945591, + 0.010948638431727886, + 0.00995388999581337, + 0.024118229746818542, + 0.015260654501616955, + 0.013673154637217522, + 0.007284914143383503, + 0.00752719771116972, + 0.034076277166604996, + 0.0049287984147667885, + 0.0090503990650177, + 0.024692201986908913, + 0.00833586324006319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.0298001728951931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.04864421859383583 + ], + "opposite_logprob_tokens": [ + -3.0486443042755127 + ], + "logprob_reward": -0.04864421859383583, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864421859383583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08509445190429688, + 0.05991216376423836, + 0.22130697965621948, + 0.059964507818222046, + 0.0684606283903122, + 0.02911556325852871, + 0.054964546114206314, + 0.032370369881391525, + 0.25237298011779785, + 0.13643783330917358 + ], + "credit_map": [ + 0.010900375433266163, + 0.007215170655399561, + 0.017763730138540268, + 0.008833340369164944, + 0.008490529842674732, + 0.00972015131264925, + 0.013563383370637894, + 0.008607769384980202, + 0.007268156856298447, + 0.004401261452585459, + 0.010547406040132046, + 0.006982250139117241, + 0.0055381180718541145, + 0.006834075320512056, + 0.009398614056408405, + 0.008942285552620888, + 0.026333481073379517, + 0.020036842674016953, + 0.017355777323246002, + 0.020859479904174805, + 0.036938391625881195, + 0.03892229124903679, + 0.03363645821809769, + 0.02722427248954773, + 0.009461610578000546, + 0.004622427746653557, + 0.015559638850390911, + 0.006166107952594757, + 0.005148365627974272, + 0.0037773537915199995, + 0.009129749611020088, + 0.006099255755543709, + 0.008968687616288662, + 0.004127824679017067, + 0.01418582908809185, + 0.008439984172582626, + 0.0054983217269182205, + 0.00450278352946043, + 0.01673898659646511, + 0.005998208653181791, + 0.0024946732446551323, + 0.0015246792463585734, + 0.0022910430561751127, + 0.005049657542258501, + 0.002389132510870695, + 0.0030101651791483164, + 0.009759815409779549, + 0.0025963950902223587, + 0.007968172430992126, + 0.005426973570138216, + 0.007010096218436956, + 0.00958879105746746, + 0.004778499715030193, + 0.008068788796663284, + 0.007401334121823311, + 0.004721888341009617, + 0.003581494092941284, + 0.002847091294825077, + 0.0042849029414355755, + 0.004986514337360859, + 0.00374435493722558, + 0.0023112203925848007, + 0.007129942998290062, + 0.0034848523791879416, + 0.03342185541987419, + 0.01569577306509018, + 0.017001673579216003, + 0.04980894923210144, + 0.02209770493209362, + 0.017080094665288925, + 0.06220255792140961, + 0.035064369440078735, + 0.010339122265577316, + 0.008020034991204739, + 0.0057364865206182, + 0.04876960068941116, + 0.0052668265998363495, + 0.02382083423435688, + 0.024485096335411072, + 0.00999982375651598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.04864421859383583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.1270153969526291 + ], + "opposite_logprob_tokens": [ + -2.1270153522491455 + ], + "logprob_reward": -0.1270153969526291, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270153969526291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1261812448501587, + 0.05201435089111328, + 0.026963211596012115, + 0.22650553286075592, + 0.263193279504776, + 0.024458518251776695, + 0.13615044951438904, + 0.039960961788892746, + 0.023435914888978004, + 0.08113648742437363 + ], + "credit_map": [ + 0.020422853529453278, + 0.01210706029087305, + 0.020948858931660652, + 0.014642699621617794, + 0.00905374065041542, + 0.015316788107156754, + 0.021651485934853554, + 0.01203775405883789, + 0.008153267204761505, + 0.0037782983854413033, + 0.01195970643311739, + 0.00735492492094636, + 0.0032245812471956015, + 0.004128899425268173, + 0.007716034073382616, + 0.005698643159121275, + 0.0027002268470823765, + 0.003733792109414935, + 0.0022804350592195988, + 0.004731945227831602, + 0.004468373488634825, + 0.0021552068646997213, + 0.0037876497954130173, + 0.0031055817380547523, + 0.02509283274412155, + 0.018664125353097916, + 0.04401228949427605, + 0.02183639258146286, + 0.02197718247771263, + 0.02595490962266922, + 0.04535330832004547, + 0.02361450530588627, + 0.02853071130812168, + 0.01797530986368656, + 0.042085200548172, + 0.03182759881019592, + 0.022503895685076714, + 0.02479141391813755, + 0.07168519496917725, + 0.023793956264853477, + 0.002495522378012538, + 0.001243287930265069, + 0.001872821245342493, + 0.0035473047755658627, + 0.0020149429328739643, + 0.003481009043753147, + 0.007795145269483328, + 0.0020084858406335115, + 0.012562522664666176, + 0.00941403303295374, + 0.025852885097265244, + 0.026248987764120102, + 0.008411762304604053, + 0.01303314883261919, + 0.029515329748392105, + 0.011111777275800705, + 0.0038414858281612396, + 0.00288363522849977, + 0.002914067590609193, + 0.006834797095507383, + 0.003835200797766447, + 0.005229326896369457, + 0.010315494611859322, + 0.004106955137103796, + 0.00331528065726161, + 0.0016021644696593285, + 0.0020169231574982405, + 0.00427304906770587, + 0.0018158664461225271, + 0.00424670847132802, + 0.0035172244533896446, + 0.0026486986316740513, + 0.011968540959060192, + 0.006222936790436506, + 0.006824992131441832, + 0.01815757341682911, + 0.005905108526349068, + 0.00521840387955308, + 0.019317500293254852, + 0.007521426305174828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.1270153969526291 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, captures a chubby, adorable giant panda. It stands in a fairytale-like kitchen, wearing a white chef's hat tilted slightly to one side and a red and white checkered apron tied around its waist. A warm, honey-sweet light slants in from the window on the right side of the frame, illuminating its contented smile. Its small eyes, narrowed blissfully beneath dark circles, and its mouth, slightly raised, cast a golden halo over the fluffy fur on its cheeks. Concentrated over a rustic wood stove, it carefully manipulates the stove with its furry paws: one firmly grasping the handle of a steaming cast iron pot, the other holding a pair of long wooden chopsticks, gently stirring the bubbling milky broth and golden noodles. The rising steam partially obscured its face, making the entire scene seem particularly dreamy. Against the backdrop of a warm cream-colored wall, strings of dried chilies and braided garlic hung from it, and a few brightly colored ceramic jars were placed on a wooden rack nearby. The entire scene was filled with a warm and soothing atmosphere, overflowing with the love of food and the pure joy of it.", + "seed": 13140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_9/index_4032_seed_13140_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": -0.039705635979771614, + "vlm_logprob": -0.039705635979771614, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.039705635979771614, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.010169712826609612 + ], + "opposite_logprob_tokens": [ + -4.635169506072998 + ], + "logprob_reward": -0.010169712826609612, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010169712826609612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11081424355506897, + 0.1018676832318306, + 0.17210081219673157, + 0.026716146618127823, + 0.0357881598174572, + 0.03620109707117081, + 0.03845681622624397, + 0.1404438316822052, + 0.22122113406658173, + 0.1163901686668396 + ], + "credit_map": [ + 0.013117782771587372, + 0.019012711942195892, + 0.018506662920117378, + 0.009531252086162567, + 0.013371743261814117, + 0.010748973116278648, + 0.012158328667283058, + 0.014366796240210533, + 0.028402451425790787, + 0.008007130585610867, + 0.0065225521102547646, + 0.004660739563405514, + 0.012732255272567272, + 0.028174517676234245, + 0.0072369021363556385, + 0.006131134461611509, + 0.05755062773823738, + 0.011399817653000355, + 0.007408937439322472, + 0.004147907719016075, + 0.013157597742974758, + 0.06406302005052567, + 0.005863053258508444, + 0.008509837090969086, + 0.0068381172604858875, + 0.0018790543545037508, + 0.0019972496666014194, + 0.001574432011693716, + 0.00375921418890357, + 0.005727898329496384, + 0.0031157503835856915, + 0.001824430888518691, + 0.007112281862646341, + 0.0036183639895170927, + 0.0038096418138593435, + 0.002264462411403656, + 0.004352286923676729, + 0.0052642240189015865, + 0.006400624755769968, + 0.002966269850730896, + 0.006875067949295044, + 0.00452266912907362, + 0.004581444896757603, + 0.003248166758567095, + 0.0048206583596765995, + 0.0026029206346720457, + 0.006152241490781307, + 0.0033979243598878384, + 0.00943811610341072, + 0.0035748190712183714, + 0.004639910068362951, + 0.0026684151962399483, + 0.0036578329745680094, + 0.007943395525217056, + 0.0032806789968162775, + 0.0032536459621042013, + 0.04087832570075989, + 0.010333186015486717, + 0.006225001998245716, + 0.004379730671644211, + 0.01586373709142208, + 0.045856863260269165, + 0.007694212719798088, + 0.009212774224579334, + 0.0702032670378685, + 0.013818735256791115, + 0.010849075391888618, + 0.005151316523551941, + 0.020824113860726357, + 0.08306211233139038, + 0.004796605557203293, + 0.012515909969806671, + 0.03552175313234329, + 0.007140504661947489, + 0.0068535711616277695, + 0.002883206820115447, + 0.011703389696776867, + 0.04203462600708008, + 0.0029384903609752655, + 0.007314631715416908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.010169712826609612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.023296575993299484 + ], + "opposite_logprob_tokens": [ + -3.773296594619751 + ], + "logprob_reward": -0.023296575993299484, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296575993299484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06563118100166321, + 0.16065557301044464, + 0.07338359206914902, + 0.1121877059340477, + 0.10823861509561539, + 0.07790501415729523, + 0.05177871882915497, + 0.021144144237041473, + 0.1381024718284607, + 0.1909729689359665 + ], + "credit_map": [ + 0.012428000569343567, + 0.009706850163638592, + 0.015011978335678577, + 0.0042237271554768085, + 0.0064590973779559135, + 0.005832477938383818, + 0.006755524314939976, + 0.00521352794021368, + 0.05042833834886551, + 0.011572564952075481, + 0.008414704352617264, + 0.005277701187878847, + 0.014954224228858948, + 0.05550655350089073, + 0.005720760673284531, + 0.00878072064369917, + 0.02369639091193676, + 0.005941253621131182, + 0.003887284081429243, + 0.001927934936247766, + 0.006053841672837734, + 0.025926826521754265, + 0.002491156104952097, + 0.0034589061979204416, + 0.03682195395231247, + 0.007252479903399944, + 0.006243946496397257, + 0.003244071966037154, + 0.008939040824770927, + 0.04083274304866791, + 0.002990628592669964, + 0.005862849298864603, + 0.03449556976556778, + 0.013926984742283821, + 0.011549222283065319, + 0.004962896928191185, + 0.006639035418629646, + 0.02678651362657547, + 0.004361137747764587, + 0.005517262499779463, + 0.016703372821211815, + 0.017324844375252724, + 0.012097814120352268, + 0.007662812247872353, + 0.008162746205925941, + 0.006450660526752472, + 0.004837306682020426, + 0.004665457643568516, + 0.013635114766657352, + 0.006858239881694317, + 0.005811369512230158, + 0.003562147729098797, + 0.00419803149998188, + 0.012114756740629673, + 0.002149634063243866, + 0.0034494276624172926, + 0.0039764875546097755, + 0.002849272219464183, + 0.0032992414198815823, + 0.001572328619658947, + 0.0028143334202468395, + 0.0036243500653654337, + 0.0015866949688643217, + 0.0014214373659342527, + 0.04405682533979416, + 0.008695395663380623, + 0.006800831761211157, + 0.0035218901466578245, + 0.012692072428762913, + 0.0515618622303009, + 0.00302909966558218, + 0.007744506932795048, + 0.05779064819216728, + 0.012586984783411026, + 0.013118605129420757, + 0.004289083182811737, + 0.01868210732936859, + 0.06814884394407272, + 0.004388713277876377, + 0.01196798775345087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.023296575993299484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.043057918548583984 + ], + "opposite_logprob_tokens": [ + -3.168057918548584 + ], + "logprob_reward": -0.043057918548583984, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057918548583984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033048536628484726, + 0.04956182837486267, + 0.22746534645557404, + 0.06770002841949463, + 0.021963493898510933, + 0.017632735893130302, + 0.025967083871364594, + 0.38746416568756104, + 0.1012628898024559, + 0.0679338201880455 + ], + "credit_map": [ + 0.00459043076261878, + 0.005625144112855196, + 0.0065692998468875885, + 0.002936904551461339, + 0.004144575912505388, + 0.003204780165106058, + 0.0034770057536661625, + 0.0025003955233842134, + 0.009581506252288818, + 0.005154007114470005, + 0.005110387224704027, + 0.0030114396940916777, + 0.009304153732955456, + 0.008085492067039013, + 0.00577161367982626, + 0.0035432293079793453, + 0.06930775940418243, + 0.019152315333485603, + 0.011077731847763062, + 0.008195603266358376, + 0.024041183292865753, + 0.07342811673879623, + 0.010225452482700348, + 0.012037191540002823, + 0.020366055890917778, + 0.004889866337180138, + 0.00354048702865839, + 0.0021077385172247887, + 0.006679572630673647, + 0.02218616008758545, + 0.004173089750111103, + 0.0037570567801594734, + 0.0035127182491123676, + 0.005581330507993698, + 0.0020716285798698664, + 0.0018239188939332962, + 0.0027242768555879593, + 0.0019053685246035457, + 0.0028343037702143192, + 0.0015099495649337769, + 0.0025257396046072245, + 0.004618930630385876, + 0.0019218189409002662, + 0.0011070278706029058, + 0.0025130542926490307, + 0.0015690451255068183, + 0.0021778421942144632, + 0.0011992772342637181, + 0.004356210120022297, + 0.004558683373034, + 0.002835674677044153, + 0.0016501877689734101, + 0.003187622409313917, + 0.002981172874569893, + 0.00464600371196866, + 0.0017515294020995498, + 0.11328817903995514, + 0.02924806997179985, + 0.018811283633112907, + 0.011420759372413158, + 0.0430794283747673, + 0.12786288559436798, + 0.017800448462367058, + 0.025953123345971107, + 0.027964333072304726, + 0.007824509404599667, + 0.007430933881551027, + 0.003521098056808114, + 0.011000317521393299, + 0.0326208658516407, + 0.004948197398334742, + 0.00595263484865427, + 0.020720914006233215, + 0.00410297978669405, + 0.003993352875113487, + 0.0017318365862593055, + 0.0066321734338998795, + 0.02482270821928978, + 0.0017335980664938688, + 0.004196265246719122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.043057918548583984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.04305506497621536 + ], + "opposite_logprob_tokens": [ + -3.1680550575256348 + ], + "logprob_reward": -0.04305506497621536, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305506497621536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06337739527225494, + 0.05098618194460869, + 0.14727453887462616, + 0.05677761510014534, + 0.1500486135482788, + 0.12193715572357178, + 0.024534521624445915, + 0.12892943620681763, + 0.10687050968408585, + 0.14926399290561676 + ], + "credit_map": [ + 0.009301587007939816, + 0.013153056614100933, + 0.008230901323258877, + 0.004841705318540335, + 0.007924932055175304, + 0.00566310016438365, + 0.007393314968794584, + 0.006868802011013031, + 0.011200884357094765, + 0.00412630382925272, + 0.004069947637617588, + 0.0020886939018964767, + 0.008065825328230858, + 0.014070331119000912, + 0.0037200860679149628, + 0.003644106909632683, + 0.04616154357790947, + 0.009651568718254566, + 0.008494418114423752, + 0.005318694282323122, + 0.013869846239686012, + 0.048881836235523224, + 0.007854838855564594, + 0.007041788659989834, + 0.011428236030042171, + 0.004461298696696758, + 0.007882924750447273, + 0.0034159068018198013, + 0.007512337062507868, + 0.013273091055452824, + 0.004589547868818045, + 0.0042142728343605995, + 0.02986953593790531, + 0.042575664818286896, + 0.02297305501997471, + 0.01146871130913496, + 0.013396857306361198, + 0.014856303110718727, + 0.00711094168946147, + 0.007797544356435537, + 0.023970089852809906, + 0.036819618195295334, + 0.01385269220918417, + 0.010195567272603512, + 0.013347143307328224, + 0.01069607026875019, + 0.006222918163985014, + 0.006833063904196024, + 0.005669053178280592, + 0.004255932755768299, + 0.0022032312117516994, + 0.00136725849006325, + 0.0026530171744525433, + 0.00573888560757041, + 0.0013342243619263172, + 0.0013129202416166663, + 0.03677690774202347, + 0.00982888974249363, + 0.0075931367464363575, + 0.004081548657268286, + 0.013576914556324482, + 0.041783981025218964, + 0.006419995333999395, + 0.008868064731359482, + 0.03171572834253311, + 0.006594750564545393, + 0.005877840332686901, + 0.003011893481016159, + 0.010941272601485252, + 0.03770908713340759, + 0.0047791097313165665, + 0.0062408167868852615, + 0.04320685938000679, + 0.010174517519772053, + 0.01372708659619093, + 0.004043474793434143, + 0.014385455287992954, + 0.050303779542446136, + 0.004401432815939188, + 0.009021389298141003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.04305506497621536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.07894890755414963 + ], + "opposite_logprob_tokens": [ + -2.578948974609375 + ], + "logprob_reward": -0.07894890755414963, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894890755414963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07326726615428925, + 0.07985059171915054, + 0.02389848232269287, + 0.06741540133953094, + 0.02381226234138012, + 0.035855818539857864, + 0.04219849407672882, + 0.18721935153007507, + 0.4346453845500946, + 0.03183705732226372 + ], + "credit_map": [ + 0.013048484921455383, + 0.010733526200056076, + 0.014865521341562271, + 0.005942558404058218, + 0.007514042779803276, + 0.0059995511546730995, + 0.0077917012386024, + 0.007371880579739809, + 0.016249440610408783, + 0.009442736394703388, + 0.0073624178767204285, + 0.004194761160761118, + 0.012415215373039246, + 0.017572803422808647, + 0.006691341754049063, + 0.005921872798353434, + 0.005657661240547895, + 0.002408010186627507, + 0.002288133604452014, + 0.001547003397718072, + 0.0032577752135694027, + 0.005320306867361069, + 0.0018087523058056831, + 0.0016108393901959062, + 0.013621344231069088, + 0.006966976914554834, + 0.004892671946436167, + 0.0033816916402429342, + 0.009990242309868336, + 0.016335176303982735, + 0.008081292733550072, + 0.004146010614931583, + 0.0038773752748966217, + 0.00430023530498147, + 0.003088136902078986, + 0.0018103000475093722, + 0.0032134202774614096, + 0.0029744450002908707, + 0.0027724530082195997, + 0.001775896642357111, + 0.005958036985248327, + 0.0077070677652955055, + 0.0034964217338711023, + 0.0027861131820827723, + 0.005224002990871668, + 0.0033819612581282854, + 0.004292144440114498, + 0.0030100683216005564, + 0.006393000017851591, + 0.005967352539300919, + 0.005118453409522772, + 0.0028083971701562405, + 0.0045103877782821655, + 0.004947404842823744, + 0.008214093744754791, + 0.0042394050396978855, + 0.05332288146018982, + 0.01442467700690031, + 0.009900134056806564, + 0.006659673526883125, + 0.02113278955221176, + 0.057800326496362686, + 0.01100426446646452, + 0.012974602170288563, + 0.13795910775661469, + 0.02818140760064125, + 0.020990652963519096, + 0.011945296078920364, + 0.03985613211989403, + 0.16187743842601776, + 0.009871755726635456, + 0.023963572457432747, + 0.006339746527373791, + 0.002735939808189869, + 0.004404526203870773, + 0.001906115678139031, + 0.003957716282457113, + 0.007836909964680672, + 0.0020316625013947487, + 0.0026244428008794785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.07894890755414963 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, captures a chubby, adorable giant panda. It stands in a fairytale-like kitchen, wearing a white chef's hat tilted slightly to one side and a red and white checkered apron tied around its waist. A warm, honey-sweet light slants in from the window on the right side of the frame, illuminating its contented smile. Its small eyes, narrowed blissfully beneath dark circles, and its mouth, slightly raised, cast a golden halo over the fluffy fur on its cheeks. Concentrated over a rustic wood stove, it carefully manipulates the stove with its furry paws: one firmly grasping the handle of a steaming cast iron pot, the other holding a pair of long wooden chopsticks, gently stirring the bubbling milky broth and golden noodles. The rising steam partially obscured its face, making the entire scene seem particularly dreamy. Against the backdrop of a warm cream-colored wall, strings of dried chilies and braided garlic hung from it, and a few brightly colored ceramic jars were placed on a wooden rack nearby. The entire scene was filled with a warm and soothing atmosphere, overflowing with the love of food and the pure joy of it.", + "seed": 13141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_9/index_4032_seed_13141_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": -0.028710604459047318, + "vlm_logprob": -0.028710604459047318, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.028710604459047318, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.004735564347356558 + ], + "opposite_logprob_tokens": [ + -5.504735469818115 + ], + "logprob_reward": -0.004735564347356558, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004735564347356558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27970534563064575, + 0.10370327532291412, + 0.07503774017095566, + 0.07072221487760544, + 0.1405009776353836, + 0.07980615645647049, + 0.05903827026486397, + 0.040284015238285065, + 0.11848871409893036, + 0.03271325305104256 + ], + "credit_map": [ + 0.0539775975048542, + 0.02974557690322399, + 0.044293973594903946, + 0.025907082483172417, + 0.019218415021896362, + 0.03731309995055199, + 0.029828062281012535, + 0.03942156955599785, + 0.014408333227038383, + 0.014064941555261612, + 0.010872610844671726, + 0.010430612601339817, + 0.01007423922419548, + 0.013957086019217968, + 0.01625402458012104, + 0.013641425408422947, + 0.010764751583337784, + 0.009770505130290985, + 0.008239557035267353, + 0.006968749221414328, + 0.007292806636542082, + 0.009357650764286518, + 0.014219235628843307, + 0.0084244878962636, + 0.008653992787003517, + 0.009015720337629318, + 0.010124570690095425, + 0.006777819711714983, + 0.007651578169316053, + 0.010695365257561207, + 0.0077547975815832615, + 0.010048368014395237, + 0.02186707593500614, + 0.017944371327757835, + 0.02077290043234825, + 0.017356816679239273, + 0.013863613829016685, + 0.017033329233527184, + 0.01742382161319256, + 0.014239062555134296, + 0.008574375882744789, + 0.007532940246164799, + 0.010104729793965816, + 0.008084678091108799, + 0.007082069292664528, + 0.01336415484547615, + 0.014836885035037994, + 0.010226313956081867, + 0.010082895867526531, + 0.006644100416451693, + 0.009957353584468365, + 0.0068079340271651745, + 0.005293923430144787, + 0.009341534227132797, + 0.006126856431365013, + 0.004783677868545055, + 0.006602334324270487, + 0.005663574207574129, + 0.00396002409979701, + 0.004405746702104807, + 0.004180596210062504, + 0.0043306234292685986, + 0.0066957673989236355, + 0.004445347934961319, + 0.014875807799398899, + 0.018536442890763283, + 0.012565890327095985, + 0.012031583115458488, + 0.014373164623975754, + 0.022830726578831673, + 0.010503971949219704, + 0.012771128676831722, + 0.005489688832312822, + 0.003461875719949603, + 0.00440549710765481, + 0.0026651069056242704, + 0.003539671655744314, + 0.003644755110144615, + 0.005806661676615477, + 0.003699998836964369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.004735564347356558 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.005263755563646555 + ], + "opposite_logprob_tokens": [ + -5.255263805389404 + ], + "logprob_reward": -0.005263755563646555, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263755563646555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2873850166797638, + 0.1192958876490593, + 0.07161829620599747, + 0.04904917627573013, + 0.11092282831668854, + 0.08143067359924316, + 0.046856172382831573, + 0.06790667772293091, + 0.09343506395816803, + 0.07210022956132889 + ], + "credit_map": [ + 0.048808131366968155, + 0.023419532924890518, + 0.04717884212732315, + 0.026813959702849388, + 0.02053093910217285, + 0.059629328548908234, + 0.027470069006085396, + 0.03353419154882431, + 0.01976587437093258, + 0.013017170131206512, + 0.020795710384845734, + 0.012330573052167892, + 0.010231452994048595, + 0.01932869851589203, + 0.011226318776607513, + 0.012600083835422993, + 0.01220251340419054, + 0.005278481636196375, + 0.012351351790130138, + 0.007008096668869257, + 0.0057001556269824505, + 0.01293360535055399, + 0.007371673826128244, + 0.008772416971623898, + 0.00933099165558815, + 0.006501692812889814, + 0.005453410558402538, + 0.0042103989981114864, + 0.004799261223524809, + 0.006050216034054756, + 0.006888571195304394, + 0.005814635194838047, + 0.015044470317661762, + 0.007920228876173496, + 0.020613519474864006, + 0.00834688264876604, + 0.00812994223088026, + 0.023164117708802223, + 0.016163425520062447, + 0.011540236882865429, + 0.01155112124979496, + 0.005710463505238295, + 0.013901697471737862, + 0.005459359381347895, + 0.006608918309211731, + 0.018887054175138474, + 0.010172481648623943, + 0.009139581583440304, + 0.00667677354067564, + 0.004426314029842615, + 0.008439132012426853, + 0.004106780514121056, + 0.003455385100096464, + 0.0069547682069242, + 0.006938656326383352, + 0.005858364515006542, + 0.012091616168618202, + 0.006507309153676033, + 0.01068056095391512, + 0.0051535931415855885, + 0.0056083062663674355, + 0.013299932703375816, + 0.006662232335656881, + 0.007903123274445534, + 0.013328437693417072, + 0.012334282509982586, + 0.011988041922450066, + 0.007348667364567518, + 0.009626587852835655, + 0.020114703103899956, + 0.00880713015794754, + 0.009887214750051498, + 0.014707518741488457, + 0.004983723163604736, + 0.01086535956710577, + 0.0038451135624200106, + 0.0057355547323822975, + 0.015389692038297653, + 0.007902196608483791, + 0.008671069517731667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005263755563646555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.04865466430783272 + ], + "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": [ + 0.3628127872943878, + 0.07733987271785736, + 0.0793040469288826, + 0.10553283989429474, + 0.07208545506000519, + 0.05859119072556496, + 0.05105011165142059, + 0.0355677530169487, + 0.12743747234344482, + 0.030278390273451805 + ], + "credit_map": [ + 0.050285398960113525, + 0.027606617659330368, + 0.05106223747134209, + 0.02592887356877327, + 0.027009427547454834, + 0.07049492746591568, + 0.053368110209703445, + 0.05705723911523819, + 0.01001252606511116, + 0.009062380529940128, + 0.01064195204526186, + 0.008886835537850857, + 0.007720128633081913, + 0.010299408808350563, + 0.009923586621880531, + 0.010793063789606094, + 0.008883652277290821, + 0.006875406485050917, + 0.010037091560661793, + 0.005844843573868275, + 0.005315100308507681, + 0.018041709437966347, + 0.013654145412147045, + 0.010652103461325169, + 0.013631343841552734, + 0.010987695306539536, + 0.01266862079501152, + 0.009254355914890766, + 0.007738815154880285, + 0.013070573098957539, + 0.018060559406876564, + 0.020120883360505104, + 0.010604373179376125, + 0.007019498851150274, + 0.009965423494577408, + 0.0072381021454930305, + 0.006386539898812771, + 0.008056815713644028, + 0.013620734214782715, + 0.009193976409733295, + 0.007059910334646702, + 0.004418946802616119, + 0.008638717234134674, + 0.004659240134060383, + 0.004311884753406048, + 0.010851569473743439, + 0.010831615887582302, + 0.007819314487278461, + 0.007498631719499826, + 0.004285077564418316, + 0.008928820490837097, + 0.004367776680737734, + 0.003940050955861807, + 0.007382985670119524, + 0.00843453872948885, + 0.006212233100086451, + 0.006392661482095718, + 0.004917926155030727, + 0.004513336345553398, + 0.003723287023603916, + 0.004086349159479141, + 0.003505792934447527, + 0.003922540228813887, + 0.004505862947553396, + 0.02416936680674553, + 0.010527578182518482, + 0.021243620663881302, + 0.0096804890781641, + 0.00857766903936863, + 0.01530931144952774, + 0.023441268131136894, + 0.014488175511360168, + 0.004519602283835411, + 0.0026841494254767895, + 0.004178703762590885, + 0.001989734824746847, + 0.002067100489512086, + 0.00580262765288353, + 0.004797333851456642, + 0.004239141475409269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 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 chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.005945615470409393 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945615470409393, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945615470409393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12654322385787964, + 0.11127561330795288, + 0.0468965582549572, + 0.10708989202976227, + 0.12416259944438934, + 0.23967643082141876, + 0.027369847521185875, + 0.07962135225534439, + 0.0631781816482544, + 0.07418625801801682 + ], + "credit_map": [ + 0.0164821594953537, + 0.010552285239100456, + 0.018739501014351845, + 0.009049336425960064, + 0.010086246766149998, + 0.021915104240179062, + 0.024574536830186844, + 0.015144064091145992, + 0.014715947210788727, + 0.008207464590668678, + 0.016155613586306572, + 0.007071017287671566, + 0.008626304566860199, + 0.02868444286286831, + 0.010739510878920555, + 0.017075318843126297, + 0.0061300997622311115, + 0.0036038970574736595, + 0.005398317240178585, + 0.003597031347453594, + 0.0035612215287983418, + 0.009990059770643711, + 0.008092566393315792, + 0.006523366551846266, + 0.0159269068390131, + 0.007242470979690552, + 0.01756356842815876, + 0.008255144581198692, + 0.006871218793094158, + 0.018190160393714905, + 0.017707565799355507, + 0.015332856215536594, + 0.015090701170265675, + 0.009010237641632557, + 0.018695063889026642, + 0.01007083896547556, + 0.009863292798399925, + 0.03207198157906532, + 0.012123973108828068, + 0.01723652333021164, + 0.025861376896500587, + 0.0167685579508543, + 0.03410845994949341, + 0.017384503036737442, + 0.020933208987116814, + 0.05853070318698883, + 0.029341574758291245, + 0.03674806281924248, + 0.0032569176983088255, + 0.003104547271504998, + 0.0035506486892700195, + 0.0023290689568966627, + 0.002435794798657298, + 0.006318047642707825, + 0.0026055520866066217, + 0.00376926944591105, + 0.00974043644964695, + 0.004789546597748995, + 0.010322240181267262, + 0.0046746195293962955, + 0.006295463070273399, + 0.021840795874595642, + 0.009246385656297207, + 0.012711863033473492, + 0.011537057347595692, + 0.006030605640262365, + 0.009908153675496578, + 0.004345524124801159, + 0.0053883325308561325, + 0.012977377511560917, + 0.005687251221388578, + 0.007303876336663961, + 0.010495896451175213, + 0.004564924165606499, + 0.010380574502050877, + 0.004476642701774836, + 0.005791110452264547, + 0.01915639266371727, + 0.00790154468268156, + 0.011419171467423439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.005945615470409393 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby giant panda wearing a white chef's hat and a red and white checkered apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda standing in a kitchen with a rustic wood stove, cream-colored walls, and hanging dried chilies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda using one paw to hold a steaming cast iron pot and the other to stir bubbling milky broth with wooden chopsticks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm light coming from a window on the right side, creating a golden halo on the panda's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken 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": [ + -0.07895342260599136 + ], + "opposite_logprob_tokens": [ + -2.578953504562378 + ], + "logprob_reward": -0.07895342260599136, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895342260599136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22328323125839233, + 0.19275644421577454, + 0.17514155805110931, + 0.06247551366686821, + 0.06339087337255478, + 0.06217954680323601, + 0.06587911397218704, + 0.041838932782411575, + 0.05227138474583626, + 0.06078337877988815 + ], + "credit_map": [ + 0.03623626381158829, + 0.018853221088647842, + 0.04150211811065674, + 0.01830172725021839, + 0.016360951587557793, + 0.03336696699261665, + 0.03225235268473625, + 0.02640962041914463, + 0.02260657399892807, + 0.02194073796272278, + 0.03491435572504997, + 0.015229938551783562, + 0.014932730235159397, + 0.0308364424854517, + 0.021871693432331085, + 0.030423976480960846, + 0.019858645275235176, + 0.020821155980229378, + 0.021882474422454834, + 0.018933070823550224, + 0.015104318037629128, + 0.034616775810718536, + 0.02164526656270027, + 0.02227983996272087, + 0.007975662127137184, + 0.009465117007493973, + 0.008594660088419914, + 0.007759114261716604, + 0.007625038269907236, + 0.004711219109594822, + 0.01006595604121685, + 0.006278747692704201, + 0.00660996837541461, + 0.006544940639287233, + 0.008818818256258965, + 0.006863432936370373, + 0.005775969009846449, + 0.01102488860487938, + 0.0077261365950107574, + 0.010026722215116024, + 0.008795352652668953, + 0.004119894001632929, + 0.00908926036208868, + 0.005081464070826769, + 0.006278518121689558, + 0.012078283354640007, + 0.010071140713989735, + 0.006665636319667101, + 0.011060306802392006, + 0.007671944797039032, + 0.010058442130684853, + 0.007205082569271326, + 0.006530351005494595, + 0.009451493620872498, + 0.008187157101929188, + 0.00571433175355196, + 0.00548253208398819, + 0.006034443154931068, + 0.005008198320865631, + 0.004383356310427189, + 0.005169504322111607, + 0.0057152677327394485, + 0.005518331658095121, + 0.004527302458882332, + 0.010877121239900589, + 0.004069451708346605, + 0.008065364323556423, + 0.0032124801073223352, + 0.0033101923763751984, + 0.010066642425954342, + 0.00661051832139492, + 0.006059614475816488, + 0.008320092223584652, + 0.0071738059632480145, + 0.008396508172154427, + 0.007762573659420013, + 0.00742159690707922, + 0.009766991250216961, + 0.006227565463632345, + 0.005714251194149256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07895342260599136 + } + ], + "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/0000125/group_0/index_tensor([5886, 5886])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_0/index_tensor([5886, 5886])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2f83427a247951e22ee3cf5229a0e035b7d26f3c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_0/index_tensor([5886, 5886])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "handheld camera, with a subtle, natural shakiness, follows a young man from a slightly lower, level angle as he navigates a narrow Tokyo back alley. The boy has short, shaggy dark brown hair, a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers. The afternoon sun filters through the gaps between buildings and the crisscrossing electrical wires overhead, casting a shifting, interwoven pattern of light and dark on him and the rough concrete walls around him. He stops beside a vintage mint green scooter, scratched and peeling, leaning forward, one leg planted firmly on the damp asphalt, a playful smile on his focused face. He slams the starter with his right foot, but the engine only lets out a weak cough before silencing itself. Undeterred, he grins, and with amusement-filled persistence, he lifts his foot again, ready for another decisive attempt. The whole scene, with its shaky camera and warm lighting, is full of youthful energy and an effortless, casual sense of urban life, creating a relaxed and hopeful leisure atmosphere.", + "seed": 33483, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_0/index_5886_seed_33483_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": -0.1321502260863781, + "vlm_logprob": -0.1321502260863781, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.1321502260863781, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2624218165874481 + ], + "opposite_logprob_tokens": [ + -1.5124218463897705 + ], + "logprob_reward": -0.2624218165874481, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2624218165874481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18540383875370026, + 0.09152156114578247, + 0.1261400282382965, + 0.06512554734945297, + 0.06835483014583588, + 0.0411209762096405, + 0.06439375877380371, + 0.14996564388275146, + 0.08189163357019424, + 0.12608224153518677 + ], + "credit_map": [ + 0.02097090519964695, + 0.026152614504098892, + 0.02124679833650589, + 0.03923279419541359, + 0.019682548940181732, + 0.018310006707906723, + 0.02326161414384842, + 0.01654655672609806, + 0.009897630661725998, + 0.011159362271428108, + 0.01464741863310337, + 0.014493170194327831, + 0.010734694078564644, + 0.009623953141272068, + 0.012276816181838512, + 0.008688513189554214, + 0.015133009292185307, + 0.01597793772816658, + 0.023753268644213676, + 0.01878388412296772, + 0.013183794915676117, + 0.013754233717918396, + 0.011679504998028278, + 0.013874387368559837, + 0.012609517201781273, + 0.006587475538253784, + 0.007360279560089111, + 0.006780566647648811, + 0.007649569772183895, + 0.005771015305072069, + 0.009075322188436985, + 0.009291802532970905, + 0.008181240409612656, + 0.0055303736589848995, + 0.00895154383033514, + 0.01198158971965313, + 0.005522024352103472, + 0.006349372211843729, + 0.012567494995892048, + 0.009271190501749516, + 0.004464492667466402, + 0.00477865943685174, + 0.005814722273498774, + 0.004837361630052328, + 0.003899691393598914, + 0.004532067105174065, + 0.007910424843430519, + 0.004883553367108107, + 0.006812209263443947, + 0.006058193277567625, + 0.0077914646826684475, + 0.00642662076279521, + 0.00862258579581976, + 0.008460238575935364, + 0.008920560590922832, + 0.01130188349634409, + 0.0234553012996912, + 0.018724536523222923, + 0.027338139712810516, + 0.01807153783738613, + 0.012421325780451298, + 0.013020144775509834, + 0.01408145111054182, + 0.02285318821668625, + 0.008860216476023197, + 0.006989675108343363, + 0.012378852814435959, + 0.008247549645602703, + 0.007429290562868118, + 0.01001084316521883, + 0.010695189237594604, + 0.01728001981973648, + 0.014498097822070122, + 0.015676114708185196, + 0.014670899137854576, + 0.01709279604256153, + 0.02135368436574936, + 0.011603177525103092, + 0.012214616872370243, + 0.01897285133600235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.2624218165874481 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206856295466423 + ], + "opposite_logprob_tokens": [ + -2.812068462371826 + ], + "logprob_reward": -0.06206856295466423, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206856295466423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28944268822669983, + 0.13719536364078522, + 0.06470946222543716, + 0.029050029814243317, + 0.11141328513622284, + 0.030885647982358932, + 0.03273576498031616, + 0.16085314750671387, + 0.04074975848197937, + 0.102964848279953 + ], + "credit_map": [ + 0.04550697281956673, + 0.03141380473971367, + 0.035801634192466736, + 0.03098180890083313, + 0.04978673532605171, + 0.04361405223608017, + 0.027255995199084282, + 0.025081660598516464, + 0.01807972602546215, + 0.015002800151705742, + 0.017688438296318054, + 0.014573139138519764, + 0.024336842820048332, + 0.012061504647135735, + 0.01895897276699543, + 0.016493942588567734, + 0.006693545263260603, + 0.0074148280546069145, + 0.008836728520691395, + 0.013048837892711163, + 0.008565311320126057, + 0.004755296278744936, + 0.009415813721716404, + 0.0059790960513055325, + 0.0038177408277988434, + 0.003327929647639394, + 0.0033668838441371918, + 0.004146184306591749, + 0.004958147648721933, + 0.002554970793426037, + 0.002992048626765609, + 0.003886123886331916, + 0.009604454971849918, + 0.00844554789364338, + 0.012698019854724407, + 0.016958540305495262, + 0.013522584922611713, + 0.00958747323602438, + 0.02130451612174511, + 0.01929214783012867, + 0.003079006914049387, + 0.002790620783343911, + 0.004553591832518578, + 0.004814379848539829, + 0.0029443642124533653, + 0.003855519462376833, + 0.0050258943811059, + 0.003822268918156624, + 0.003548581385985017, + 0.002723902463912964, + 0.0038957567885518074, + 0.002674262737855315, + 0.006701911799609661, + 0.003902534954249859, + 0.004044110421091318, + 0.005244705360382795, + 0.03124285489320755, + 0.019865844398736954, + 0.036175716668367386, + 0.018521778285503387, + 0.012799449265003204, + 0.012865772470831871, + 0.012259466573596, + 0.017122268676757812, + 0.0037798122502863407, + 0.004139191471040249, + 0.005506704095751047, + 0.004572324454784393, + 0.0030772911850363016, + 0.0029135963413864374, + 0.006162418518215418, + 0.010598418302834034, + 0.011902881786227226, + 0.01029110699892044, + 0.014020372182130814, + 0.015059880912303925, + 0.014791583642363548, + 0.009502830915153027, + 0.013664633966982365, + 0.013731559738516808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.06206856295466423 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192831702530384 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192831702530384, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192831702530384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15903697907924652, + 0.25569137930870056, + 0.0683426782488823, + 0.04363202303647995, + 0.08769740909337997, + 0.024650314822793007, + 0.02213883027434349, + 0.06791038066148758, + 0.08199957758188248, + 0.1889003962278366 + ], + "credit_map": [ + 0.019842812791466713, + 0.012812544591724873, + 0.025678208097815514, + 0.025789868086576462, + 0.02142529934644699, + 0.0197451114654541, + 0.02053084969520569, + 0.013212287798523903, + 0.03024398349225521, + 0.03205198049545288, + 0.03743424639105797, + 0.03290790319442749, + 0.04249482601881027, + 0.019277021288871765, + 0.03609863668680191, + 0.025182804092764854, + 0.010208582505583763, + 0.009495239704847336, + 0.008112477138638496, + 0.009768599644303322, + 0.008692831732332706, + 0.007064583245664835, + 0.006677268538624048, + 0.008323094807565212, + 0.008230699226260185, + 0.004120077472180128, + 0.006352032534778118, + 0.004309500101953745, + 0.0038631597999483347, + 0.004906304646283388, + 0.006367494352161884, + 0.005482756067067385, + 0.01219957135617733, + 0.009893565438687801, + 0.011131198145449162, + 0.01429221872240305, + 0.006391813512891531, + 0.009868664667010307, + 0.012020254507660866, + 0.011900129728019238, + 0.002831223653629422, + 0.002770849037915468, + 0.0036851451732218266, + 0.003410368924960494, + 0.0021606762893497944, + 0.004280072636902332, + 0.002925899811089039, + 0.002586078830063343, + 0.0034369039349257946, + 0.002458022441715002, + 0.0028768754564225674, + 0.0027852635830640793, + 0.0031811287626624107, + 0.002231452614068985, + 0.0026705421041697264, + 0.002498641377314925, + 0.012076590210199356, + 0.007424760609865189, + 0.014414796605706215, + 0.007641415111720562, + 0.005530505906790495, + 0.005543501581996679, + 0.006572368089109659, + 0.008706443011760712, + 0.007363243028521538, + 0.005278204567730427, + 0.013338077813386917, + 0.007983510382473469, + 0.006213397718966007, + 0.009103487245738506, + 0.010918582789599895, + 0.021801067516207695, + 0.04028524085879326, + 0.014365212991833687, + 0.06355386227369308, + 0.020070357248187065, + 0.010731796734035015, + 0.009150158613920212, + 0.012659415602684021, + 0.018084349110722542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.014192831702530384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21505942940711975, + 0.09090449661016464, + 0.06707338988780975, + 0.0849592536687851, + 0.09523430466651917, + 0.0924837738275528, + 0.06973664462566376, + 0.12073494493961334, + 0.05871684104204178, + 0.10509692132472992 + ], + "credit_map": [ + 0.04278004541993141, + 0.019546546041965485, + 0.0213483814150095, + 0.034408360719680786, + 0.026469562202692032, + 0.031048377975821495, + 0.02349315769970417, + 0.01596498116850853, + 0.01137640792876482, + 0.012366062961518764, + 0.010295548476278782, + 0.008853284642100334, + 0.022421065717935562, + 0.00797540508210659, + 0.010779955424368382, + 0.006836768705397844, + 0.007240604609251022, + 0.005960727576166391, + 0.009951046667993069, + 0.008513744920492172, + 0.009991365484893322, + 0.006539080757647753, + 0.010707403533160686, + 0.008169416338205338, + 0.014774625189602375, + 0.009323498234152794, + 0.013161182403564453, + 0.009390132501721382, + 0.008357977494597435, + 0.011599909514188766, + 0.008886582218110561, + 0.009465348906815052, + 0.011351562105119228, + 0.008095570839941502, + 0.011176292784512043, + 0.011399896815419197, + 0.012219823896884918, + 0.0077101439237594604, + 0.021013617515563965, + 0.012267394922673702, + 0.010265802964568138, + 0.00809541903436184, + 0.011405385099351406, + 0.012777511961758137, + 0.010354568250477314, + 0.00740690715610981, + 0.022773239761590958, + 0.009404941461980343, + 0.00927787646651268, + 0.007712778635323048, + 0.00868943240493536, + 0.007552234455943108, + 0.011633484624326229, + 0.007230172865092754, + 0.009348928928375244, + 0.008291737176477909, + 0.019946662709116936, + 0.01560228131711483, + 0.02280147559940815, + 0.014529854990541935, + 0.011399365961551666, + 0.008525955490767956, + 0.01059317123144865, + 0.017336174845695496, + 0.0061812493950128555, + 0.0052813817746937275, + 0.007147522643208504, + 0.005433310754597187, + 0.00393369235098362, + 0.004919013474136591, + 0.008177613839507103, + 0.01764305680990219, + 0.013342021033167839, + 0.011805923655629158, + 0.012799250893294811, + 0.014156423509120941, + 0.01628272980451584, + 0.007822929881513119, + 0.009724502451717854, + 0.01916315034031868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.008639576844871044 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134283423423767 + ], + "opposite_logprob_tokens": [ + -1.3134284019470215 + ], + "logprob_reward": -0.3134283423423767, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134283423423767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1637464463710785, + 0.08128704875707626, + 0.1419023722410202, + 0.055984821170568466, + 0.0902315080165863, + 0.07096193730831146, + 0.09102135896682739, + 0.07325331121683121, + 0.06787100434303284, + 0.16374026238918304 + ], + "credit_map": [ + 0.0235330518335104, + 0.01439186092466116, + 0.01749146357178688, + 0.020463852211833, + 0.01776888221502304, + 0.03937278315424919, + 0.014583525247871876, + 0.01614101231098175, + 0.010411767289042473, + 0.00899096205830574, + 0.01249586883932352, + 0.011332593858242035, + 0.007953460328280926, + 0.007632681168615818, + 0.009049292653799057, + 0.013420415110886097, + 0.024659501388669014, + 0.020219413563609123, + 0.01896880753338337, + 0.024074669927358627, + 0.0162173081189394, + 0.009768196381628513, + 0.012426541186869144, + 0.015567914582788944, + 0.0060195717960596085, + 0.004011219367384911, + 0.006786861456930637, + 0.006171959452331066, + 0.00828086119145155, + 0.006108634173870087, + 0.007308743428438902, + 0.011296968907117844, + 0.008986311964690685, + 0.008806483820080757, + 0.009544329717755318, + 0.012338083237409592, + 0.014767756685614586, + 0.007997274398803711, + 0.014989769086241722, + 0.012801488861441612, + 0.006597314029932022, + 0.005708955228328705, + 0.0091175502166152, + 0.010492328554391861, + 0.008875759318470955, + 0.006072294898331165, + 0.016358153894543648, + 0.007739577908068895, + 0.007397457957267761, + 0.006887542549520731, + 0.01163262128829956, + 0.009268105030059814, + 0.012654110789299011, + 0.007878516800701618, + 0.015284002758562565, + 0.020018992945551872, + 0.012172981165349483, + 0.015468357130885124, + 0.00742051936686039, + 0.008137494325637817, + 0.007087663281708956, + 0.007439807988703251, + 0.006586645729839802, + 0.0089398343116045, + 0.007840004749596119, + 0.00624111806973815, + 0.010164663195610046, + 0.012709622271358967, + 0.004800384398549795, + 0.008543003350496292, + 0.007660709321498871, + 0.00991148967295885, + 0.019678881391882896, + 0.017575472593307495, + 0.0371549017727375, + 0.01714351214468479, + 0.019684333354234695, + 0.011286227032542229, + 0.01371214259415865, + 0.027504775673151016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.3134283423423767 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, with a subtle, natural shakiness, follows a young man from a slightly lower, level angle as he navigates a narrow Tokyo back alley. The boy has short, shaggy dark brown hair, a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers. The afternoon sun filters through the gaps between buildings and the crisscrossing electrical wires overhead, casting a shifting, interwoven pattern of light and dark on him and the rough concrete walls around him. He stops beside a vintage mint green scooter, scratched and peeling, leaning forward, one leg planted firmly on the damp asphalt, a playful smile on his focused face. He slams the starter with his right foot, but the engine only lets out a weak cough before silencing itself. Undeterred, he grins, and with amusement-filled persistence, he lifts his foot again, ready for another decisive attempt. The whole scene, with its shaky camera and warm lighting, is full of youthful energy and an effortless, casual sense of urban life, creating a relaxed and hopeful leisure atmosphere.", + "seed": 33484, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_0/index_5886_seed_33484_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": -0.15876443777233362, + "vlm_logprob": -0.15876443777233362, + "vlm_yes_no_margin": 2.4750000119209288, + "vlm_token_credit": -0.15876443777233362, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23359593749046326 + ], + "opposite_logprob_tokens": [ + -1.6085959672927856 + ], + "logprob_reward": -0.23359593749046326, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23359593749046326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30209577083587646, + 0.10467072576284409, + 0.24338850378990173, + 0.02978876233100891, + 0.08952130377292633, + 0.04863674193620682, + 0.07060330361127853, + 0.05124339088797569, + 0.03053741529583931, + 0.02951410785317421 + ], + "credit_map": [ + 0.04750509560108185, + 0.037510886788368225, + 0.031291455030441284, + 0.031768884509801865, + 0.025705836713314056, + 0.0363851822912693, + 0.04490363970398903, + 0.04702479764819145, + 0.009485689923167229, + 0.0062684048898518085, + 0.020348699763417244, + 0.010939527302980423, + 0.009041005745530128, + 0.020826537162065506, + 0.009991900064051151, + 0.01776895858347416, + 0.02702726051211357, + 0.013928147032856941, + 0.07000212371349335, + 0.01929152011871338, + 0.015173164196312428, + 0.06862635910511017, + 0.015087180770933628, + 0.014252757653594017, + 0.0029264348559081554, + 0.003316126996651292, + 0.00427718972787261, + 0.003906418103724718, + 0.004390941467136145, + 0.0046428111381828785, + 0.003438706509768963, + 0.0028901335317641497, + 0.009858449921011925, + 0.005561106838285923, + 0.009959721937775612, + 0.015339209698140621, + 0.009457581676542759, + 0.012205883860588074, + 0.016696065664291382, + 0.01044328324496746, + 0.0049048722721636295, + 0.007263678126037121, + 0.007650830317288637, + 0.006845742464065552, + 0.005776113830506802, + 0.006680516991764307, + 0.005616632755845785, + 0.0038983558770269156, + 0.005880632903426886, + 0.009838147088885307, + 0.008847572840750217, + 0.011426500976085663, + 0.010277027264237404, + 0.010211437940597534, + 0.00878631416708231, + 0.005335671827197075, + 0.00497306976467371, + 0.004598177503794432, + 0.004353333730250597, + 0.005364234559237957, + 0.004945705179125071, + 0.01447850838303566, + 0.007488258183002472, + 0.0050420998595654964, + 0.003244981402531266, + 0.0023746283259242773, + 0.004243370611220598, + 0.004900306463241577, + 0.004992224276065826, + 0.004359812010079622, + 0.004338088445365429, + 0.0020840021315962076, + 0.004087832756340504, + 0.002667258260771632, + 0.0038831590209156275, + 0.003028452629223466, + 0.003088930621743202, + 0.004182447679340839, + 0.0044893198646605015, + 0.004086705856025219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.23359593749046326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285212755203247 + ], + "opposite_logprob_tokens": [ + -2.0178520679473877 + ], + "logprob_reward": -0.14285212755203247, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14285212755203247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4887402057647705, + 0.03876899182796478, + 0.04649342969059944, + 0.06084951013326645, + 0.09925547242164612, + 0.1054580807685852, + 0.04630470648407936, + 0.04229653999209404, + 0.05435808748006821, + 0.01747497171163559 + ], + "credit_map": [ + 0.05559109151363373, + 0.0610491968691349, + 0.1278344988822937, + 0.052255284041166306, + 0.040709465742111206, + 0.05318209156394005, + 0.03811348229646683, + 0.06000513955950737, + 0.00392838753759861, + 0.0021116111893206835, + 0.010053733363747597, + 0.004061954095959663, + 0.001841081422753632, + 0.00759238563477993, + 0.0033570225350558758, + 0.0058228145353496075, + 0.003361945739015937, + 0.0019456230802461505, + 0.013710631988942623, + 0.004773053340613842, + 0.0024625773075968027, + 0.014482816681265831, + 0.002589282812550664, + 0.003167500486597419, + 0.007234477438032627, + 0.006750962231308222, + 0.008421134203672409, + 0.007043932564556599, + 0.006374521646648645, + 0.013928711414337158, + 0.006222263444215059, + 0.004873510450124741, + 0.01216424535959959, + 0.006741868332028389, + 0.009841599501669407, + 0.016162028536200523, + 0.01247863844037056, + 0.012377206236124039, + 0.016967102885246277, + 0.012522779405117035, + 0.011314583010971546, + 0.006140721030533314, + 0.00951218418776989, + 0.017389070242643356, + 0.011268699541687965, + 0.017601581290364265, + 0.01983490027487278, + 0.012396341189742088, + 0.0030683379154652357, + 0.0031381992157548666, + 0.0036181912291795015, + 0.006846773903816938, + 0.0066771782003343105, + 0.01136494055390358, + 0.006897279527038336, + 0.004693807102739811, + 0.0028435909189283848, + 0.004249753430485725, + 0.0032123562414199114, + 0.004877737257629633, + 0.0035117422230541706, + 0.01282445713877678, + 0.006867393851280212, + 0.003909509629011154, + 0.004368387162685394, + 0.0049770246259868145, + 0.00385749782435596, + 0.008602390997111797, + 0.0060760523192584515, + 0.011669379658997059, + 0.011025071144104004, + 0.0037822872400283813, + 0.002854305785149336, + 0.001499456469900906, + 0.0016068211989477277, + 0.002780183916911483, + 0.002225829754024744, + 0.002132458845153451, + 0.002556174760684371, + 0.0018197426106780767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.14285212755203247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072005033493042 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016072005033493042, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072005033493042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3616611659526825, + 0.05586659908294678, + 0.11912993341684341, + 0.14156503975391388, + 0.07026193290948868, + 0.08819032460451126, + 0.03722168505191803, + 0.057085126638412476, + 0.026768170297145844, + 0.04225001111626625 + ], + "credit_map": [ + 0.04914698749780655, + 0.028048796579241753, + 0.07495883107185364, + 0.03548910468816757, + 0.025021042674779892, + 0.0432148352265358, + 0.056587088853120804, + 0.04919447749853134, + 0.011234687641263008, + 0.00470395851880312, + 0.006124124396592379, + 0.0079683568328619, + 0.003383151488378644, + 0.008067003451287746, + 0.0065129417926073074, + 0.007872376590967178, + 0.011340768076479435, + 0.004297605250030756, + 0.03994818776845932, + 0.008152393624186516, + 0.005999155342578888, + 0.03934108465909958, + 0.005061029922217131, + 0.004989717621356249, + 0.021364599466323853, + 0.01737034320831299, + 0.017931997776031494, + 0.01367365475744009, + 0.016016624867916107, + 0.034972142428159714, + 0.009087943471968174, + 0.011147742159664631, + 0.008217881433665752, + 0.007098858710378408, + 0.010175087489187717, + 0.010522708296775818, + 0.0072620040737092495, + 0.009010237641632557, + 0.011633328162133694, + 0.006341833621263504, + 0.009651236236095428, + 0.008045990951359272, + 0.012075723148882389, + 0.010601609945297241, + 0.010830903425812721, + 0.014705561101436615, + 0.015801560133695602, + 0.006477741524577141, + 0.003324703313410282, + 0.003612702013924718, + 0.006159754935652018, + 0.005472932010889053, + 0.004111518617719412, + 0.008045639842748642, + 0.003971852362155914, + 0.002522583119571209, + 0.006163644604384899, + 0.005345371086150408, + 0.006427801679819822, + 0.00961312185972929, + 0.009100762195885181, + 0.009619701653718948, + 0.00652120029553771, + 0.004293526988476515, + 0.003561797784641385, + 0.0022316842805594206, + 0.0035003679804503918, + 0.005259722471237183, + 0.0027983500622212887, + 0.004355376586318016, + 0.0034959937911480665, + 0.0015648800181224942, + 0.007239511236548424, + 0.005331024061888456, + 0.005109468940645456, + 0.004181610886007547, + 0.004428474698215723, + 0.0065556904301047325, + 0.004912929143756628, + 0.004491305444389582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.016072005033493042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204819686710835 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014204819686710835, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204819686710835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2876826524734497, + 0.05357198789715767, + 0.18307892978191376, + 0.09602659940719604, + 0.06105305626988411, + 0.12338219583034515, + 0.0583723783493042, + 0.0464666523039341, + 0.04358905181288719, + 0.046776510775089264 + ], + "credit_map": [ + 0.045796431601047516, + 0.03485187143087387, + 0.0820474922657013, + 0.032257214188575745, + 0.022410543635487556, + 0.030021047219634056, + 0.01897290162742138, + 0.021325161680579185, + 0.004868719261139631, + 0.0037504229694604874, + 0.005752106197178364, + 0.005693022161722183, + 0.006025198381394148, + 0.008994965814054012, + 0.009239287115633488, + 0.009248262271285057, + 0.011306105181574821, + 0.004888540133833885, + 0.06560616195201874, + 0.011189470067620277, + 0.00918755866587162, + 0.06856541335582733, + 0.005450718104839325, + 0.006884965114295483, + 0.013663720339536667, + 0.008880998939275742, + 0.009264926426112652, + 0.012883796356618404, + 0.008578920736908913, + 0.015484261326491833, + 0.015087125822901726, + 0.012182843871414661, + 0.00551297701895237, + 0.004915392491966486, + 0.006598019506782293, + 0.0071847871877253056, + 0.007694514933973551, + 0.008947637863457203, + 0.01255807001143694, + 0.007641658652573824, + 0.010815669782459736, + 0.007058596238493919, + 0.010944212786853313, + 0.022043606266379356, + 0.014279503375291824, + 0.019112735986709595, + 0.02603493072092533, + 0.01309294905513525, + 0.0047469851560890675, + 0.007167746778577566, + 0.004604575224220753, + 0.0067171258851885796, + 0.006964351516216993, + 0.012287946417927742, + 0.010025577619671822, + 0.005858068820089102, + 0.0039746412076056, + 0.005181028041988611, + 0.003991852048784494, + 0.005375943146646023, + 0.0052359201945364475, + 0.013549520634114742, + 0.005719665437936783, + 0.003438082290813327, + 0.004099324811249971, + 0.003828370710834861, + 0.0031413089018315077, + 0.0062926518730819225, + 0.007337396964430809, + 0.005441262852400541, + 0.010103024542331696, + 0.00334571348503232, + 0.0029343394562602043, + 0.0029907303396612406, + 0.003371813101693988, + 0.004735181108117104, + 0.007921746000647545, + 0.010882768779993057, + 0.00829895120114088, + 0.00564097985625267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.014204819686710835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870972990989685 + ], + "opposite_logprob_tokens": [ + -1.1370973587036133 + ], + "logprob_reward": -0.3870972990989685, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870972990989685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36459672451019287, + 0.08811411261558533, + 0.10043002665042877, + 0.0659392848610878, + 0.06934870779514313, + 0.08696923404932022, + 0.07358675450086594, + 0.08909807354211807, + 0.0336119569838047, + 0.02830510400235653 + ], + "credit_map": [ + 0.0405486598610878, + 0.021943649277091026, + 0.08401218801736832, + 0.03232502192258835, + 0.02970423921942711, + 0.04513663426041603, + 0.05318676307797432, + 0.057739563286304474, + 0.009889407083392143, + 0.0058537633158266544, + 0.019192837178707123, + 0.010720686987042427, + 0.006238117814064026, + 0.015908565372228622, + 0.010397205129265785, + 0.009913532063364983, + 0.007511861156672239, + 0.004002903588116169, + 0.03133722394704819, + 0.007938364520668983, + 0.00545057188719511, + 0.033670149743556976, + 0.004564192611724138, + 0.0059547643177211285, + 0.006987099535763264, + 0.007961391471326351, + 0.005496039986610413, + 0.009899335913360119, + 0.007562323939055204, + 0.01250202115625143, + 0.008908947929739952, + 0.006622123531997204, + 0.010362595319747925, + 0.006364126689732075, + 0.00819447822868824, + 0.011733352206647396, + 0.00834192056208849, + 0.009952083230018616, + 0.00837971642613411, + 0.006020435597747564, + 0.008491397835314274, + 0.005395783577114344, + 0.007730287034064531, + 0.010262242518365383, + 0.013337340205907822, + 0.014779871329665184, + 0.01768932119011879, + 0.009282993152737617, + 0.005972789600491524, + 0.0114608034491539, + 0.008191263303160667, + 0.012463938444852829, + 0.009745795279741287, + 0.009423096664249897, + 0.011052467860281467, + 0.005276595242321491, + 0.006120398174971342, + 0.009976486675441265, + 0.008615926839411259, + 0.014630851335823536, + 0.01002865843474865, + 0.02419423498213291, + 0.009598535485565662, + 0.0059329853393137455, + 0.002900052350014448, + 0.0027070462238043547, + 0.003950578160583973, + 0.004887819290161133, + 0.004522397648543119, + 0.006619586609303951, + 0.005391074810177088, + 0.0026334014255553484, + 0.0038884805981069803, + 0.0034433696419000626, + 0.002546143950894475, + 0.003113890765234828, + 0.0035015845205634832, + 0.003969389013946056, + 0.004682439845055342, + 0.0031598082277923822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.3870972990989685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, with a subtle, natural shakiness, follows a young man from a slightly lower, level angle as he navigates a narrow Tokyo back alley. The boy has short, shaggy dark brown hair, a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers. The afternoon sun filters through the gaps between buildings and the crisscrossing electrical wires overhead, casting a shifting, interwoven pattern of light and dark on him and the rough concrete walls around him. He stops beside a vintage mint green scooter, scratched and peeling, leaning forward, one leg planted firmly on the damp asphalt, a playful smile on his focused face. He slams the starter with his right foot, but the engine only lets out a weak cough before silencing itself. Undeterred, he grins, and with amusement-filled persistence, he lifts his foot again, ready for another decisive attempt. The whole scene, with its shaky camera and warm lighting, is full of youthful energy and an effortless, casual sense of urban life, creating a relaxed and hopeful leisure atmosphere.", + "seed": 33485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_0/index_5886_seed_33485_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": -0.09910860247910022, + "vlm_logprob": -0.09910860247910022, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.09910860247910022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11108055710792542 + ], + "opposite_logprob_tokens": [ + -2.3610806465148926 + ], + "logprob_reward": -0.11108055710792542, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11108055710792542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11248045414686203, + 0.08680684864521027, + 0.11828619986772537, + 0.10620604455471039, + 0.06141458824276924, + 0.1539878249168396, + 0.10358891636133194, + 0.062211569398641586, + 0.1272994726896286, + 0.06771809607744217 + ], + "credit_map": [ + 0.012965884990990162, + 0.012163316830992699, + 0.014829589053988457, + 0.02223191410303116, + 0.010752297937870026, + 0.012990882620215416, + 0.011742061004042625, + 0.014804506674408913, + 0.009181007742881775, + 0.0052677663043141365, + 0.017988678067922592, + 0.00932249054312706, + 0.010832126252353191, + 0.018114833161234856, + 0.010007137432694435, + 0.006092808675020933, + 0.014334742911159992, + 0.03852354735136032, + 0.03010108508169651, + 0.012218140065670013, + 0.007241716142743826, + 0.0052100433968007565, + 0.005114169791340828, + 0.0055427527986466885, + 0.018658794462680817, + 0.014848397113382816, + 0.0177022535353899, + 0.011347479186952114, + 0.01547364704310894, + 0.007706284988671541, + 0.008757085539400578, + 0.011712094768881798, + 0.009793463163077831, + 0.007636537775397301, + 0.010175695642828941, + 0.005305194761604071, + 0.0073194908909499645, + 0.005057468079030514, + 0.005220548249781132, + 0.010906187817454338, + 0.03685929626226425, + 0.014127010479569435, + 0.03207772970199585, + 0.01738561876118183, + 0.009379399009048939, + 0.010500694625079632, + 0.006591705605387688, + 0.02706637606024742, + 0.019584767520427704, + 0.008700727485120296, + 0.014905594289302826, + 0.009063010104000568, + 0.012241390533745289, + 0.012879171408712864, + 0.008436236530542374, + 0.01777801290154457, + 0.007877346128225327, + 0.005542149301618338, + 0.009004254825413227, + 0.006734069436788559, + 0.010529429651796818, + 0.010100717656314373, + 0.007507681846618652, + 0.004915924277156591, + 0.022877033799886703, + 0.01508468110114336, + 0.014355257153511047, + 0.011034361086785793, + 0.013388956896960735, + 0.01612694375216961, + 0.01613207906484604, + 0.018300160765647888, + 0.006254115141928196, + 0.005754288751631975, + 0.007682158146053553, + 0.00988531019538641, + 0.015622690320014954, + 0.008414272218942642, + 0.008568582125008106, + 0.005536677315831184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.11108055710792542 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307573288679123 + ], + "opposite_logprob_tokens": [ + -3.1680757999420166 + ], + "logprob_reward": -0.04307573288679123, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307573288679123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07743930071592331, + 0.07507336139678955, + 0.134757399559021, + 0.10635163635015488, + 0.0273281279951334, + 0.06213511526584625, + 0.0515625961124897, + 0.07417090982198715, + 0.15772300958633423, + 0.2334585338830948 + ], + "credit_map": [ + 0.012318112887442112, + 0.0075135743245482445, + 0.012692788615822792, + 0.01430831290781498, + 0.007815290242433548, + 0.0077494969591498375, + 0.008986018598079681, + 0.006055720616132021, + 0.00653826491907239, + 0.0036498429253697395, + 0.019008025527000427, + 0.005140847526490688, + 0.008852693252265453, + 0.01936383545398712, + 0.007001389283686876, + 0.005518467631191015, + 0.01772608608007431, + 0.04117525368928909, + 0.03563593700528145, + 0.013580260798335075, + 0.008944803848862648, + 0.005899949464946985, + 0.006088524591177702, + 0.0057065957225859165, + 0.014607960358262062, + 0.012030229903757572, + 0.017587140202522278, + 0.01446184515953064, + 0.014470896683633327, + 0.007739338092505932, + 0.013536136597394943, + 0.011918102391064167, + 0.0037491547409445047, + 0.0028777741827070713, + 0.004932676907628775, + 0.002802525879815221, + 0.0033872697968035936, + 0.002243871334940195, + 0.003006973769515753, + 0.004327884875237942, + 0.006730834022164345, + 0.005225157830864191, + 0.01327231340110302, + 0.006864908151328564, + 0.006619611289352179, + 0.006930972915142775, + 0.0051465388387441635, + 0.011344782076776028, + 0.006061592139303684, + 0.005449492949992418, + 0.008072826080024242, + 0.005539769772440195, + 0.006410801783204079, + 0.006142808124423027, + 0.005982913542538881, + 0.007902395911514759, + 0.007135952822864056, + 0.006214079912751913, + 0.008417136035859585, + 0.006667531561106443, + 0.015295248478651047, + 0.009559229016304016, + 0.014426754787564278, + 0.006454981863498688, + 0.028580540791153908, + 0.019499097019433975, + 0.018395306542515755, + 0.014080574735999107, + 0.013711616396903992, + 0.019808785989880562, + 0.020180869847536087, + 0.02346622943878174, + 0.020201025530695915, + 0.018867874518036842, + 0.014631790108978748, + 0.029524924233555794, + 0.06256989389657974, + 0.023198846727609634, + 0.04620686173439026, + 0.018257347866892815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.04307573288679123 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186016008257866 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186016008257866, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186016008257866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11039941012859344, + 0.14692345261573792, + 0.12321241199970245, + 0.05463715270161629, + 0.041446175426244736, + 0.03298414498567581, + 0.20067910850048065, + 0.017803587019443512, + 0.22437430918216705, + 0.04754018038511276 + ], + "credit_map": [ + 0.013794777914881706, + 0.013128694146871567, + 0.017342541366815567, + 0.014994386583566666, + 0.013973710127174854, + 0.015677250921726227, + 0.01268344558775425, + 0.008804616518318653, + 0.014305836521089077, + 0.009312281385064125, + 0.03807349503040314, + 0.013176040723919868, + 0.014598515816032887, + 0.035845473408699036, + 0.011361140757799149, + 0.010250681079924107, + 0.015216488391160965, + 0.03362984210252762, + 0.02336922660470009, + 0.015646321699023247, + 0.009320107288658619, + 0.0072670066729187965, + 0.009858423843979836, + 0.00890500657260418, + 0.010245375335216522, + 0.00705245416611433, + 0.009653152897953987, + 0.006232286337763071, + 0.006066957954317331, + 0.004121062811464071, + 0.004468773957341909, + 0.006797095760703087, + 0.006372714880853891, + 0.004376876633614302, + 0.006922421045601368, + 0.005301486235111952, + 0.004626401234418154, + 0.0024706623516976833, + 0.004301098641008139, + 0.007074519991874695, + 0.007554671261459589, + 0.002725825645029545, + 0.005045642610639334, + 0.0034012869000434875, + 0.00236298656091094, + 0.0029493377078324556, + 0.0038310771342366934, + 0.005113319028168917, + 0.033064115792512894, + 0.02067871019244194, + 0.030463026836514473, + 0.024343138560652733, + 0.021801015362143517, + 0.02385200560092926, + 0.01868157461285591, + 0.027795542031526566, + 0.0018849126063287258, + 0.0018213249277323484, + 0.003210953902453184, + 0.002724963705986738, + 0.0018257537158206105, + 0.0020992399659007788, + 0.001989686628803611, + 0.0022467519156634808, + 0.04542786628007889, + 0.02615361288189888, + 0.023536689579486847, + 0.01787159964442253, + 0.018779972568154335, + 0.023187492042779922, + 0.03377251699566841, + 0.03564458712935448, + 0.0031226573046296835, + 0.004654539283365011, + 0.0038049619179219007, + 0.009095996618270874, + 0.007796620484441519, + 0.007411465980112553, + 0.006068084388971329, + 0.005585858132690191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.018186016008257866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782824665307999 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782824665307999, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782824665307999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04397953301668167, + 0.05934556946158409, + 0.4359206259250641, + 0.055003516376018524, + 0.05866536498069763, + 0.08444835245609283, + 0.07924117147922516, + 0.052782297134399414, + 0.07323696464300156, + 0.05737666040658951 + ], + "credit_map": [ + 0.005043035838752985, + 0.004973615985363722, + 0.005918446462601423, + 0.007719126529991627, + 0.004263914655894041, + 0.0033526390325278044, + 0.007085621356964111, + 0.0056231277994811535, + 0.005547233857214451, + 0.002348215552046895, + 0.01568526402115822, + 0.0047075944021344185, + 0.006561733782291412, + 0.016338225454092026, + 0.005031653214246035, + 0.0031256452202796936, + 0.054096732288599014, + 0.14728765189647675, + 0.12040684372186661, + 0.036865174770355225, + 0.02678179368376732, + 0.01964164897799492, + 0.014508930966258049, + 0.016331780701875687, + 0.008318133652210236, + 0.007888441905379295, + 0.011693919077515602, + 0.005629390943795443, + 0.006182725541293621, + 0.0035694283433258533, + 0.005036584101617336, + 0.006684888619929552, + 0.009608227759599686, + 0.005745967850089073, + 0.010112069547176361, + 0.007019581738859415, + 0.005995723884552717, + 0.005207517649978399, + 0.005690382327884436, + 0.009285888634622097, + 0.018187707290053368, + 0.009880728088319302, + 0.011037185788154602, + 0.011232172138988972, + 0.00651725847274065, + 0.009399738162755966, + 0.00644204206764698, + 0.011751512065529823, + 0.011507680639624596, + 0.007252973970025778, + 0.009659895673394203, + 0.008619635365903378, + 0.008914828300476074, + 0.01164171565324068, + 0.009582585655152798, + 0.012061848305165768, + 0.005220609251409769, + 0.004693868104368448, + 0.004812346305698156, + 0.004745949991047382, + 0.010650976561009884, + 0.008110087364912033, + 0.0093157310038805, + 0.005232721567153931, + 0.011539280414581299, + 0.0068716141395270824, + 0.0069760470651090145, + 0.004645046778023243, + 0.014420959167182446, + 0.010813003405928612, + 0.009138570167124271, + 0.008832426741719246, + 0.00468558119609952, + 0.005124442745000124, + 0.005055765155702829, + 0.007861376740038395, + 0.01240037102252245, + 0.007399794179946184, + 0.010110050439834595, + 0.004739274270832539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.009782824665307999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341788172721863 + ], + "opposite_logprob_tokens": [ + -1.313417911529541 + ], + "logprob_reward": -0.31341788172721863, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341788172721863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054036758840084076, + 0.0832664743065834, + 0.41308170557022095, + 0.05588145554065704, + 0.06403648108243942, + 0.052533648908138275, + 0.04759357124567032, + 0.12034530192613602, + 0.058552250266075134, + 0.05067238584160805 + ], + "credit_map": [ + 0.006940076593309641, + 0.005764253903180361, + 0.00696862768381834, + 0.009292985312640667, + 0.0043519968166947365, + 0.0061977277509868145, + 0.005893684923648834, + 0.008627410046756268, + 0.007995150052011013, + 0.004152223002165556, + 0.023542556911706924, + 0.005960510112345219, + 0.008305059745907784, + 0.020270805805921555, + 0.008911823853850365, + 0.004128348082304001, + 0.04315410554409027, + 0.1475333273410797, + 0.1109209954738617, + 0.03262747824192047, + 0.026266107335686684, + 0.0180099755525589, + 0.016225645318627357, + 0.018344057723879814, + 0.005190219264477491, + 0.005880032666027546, + 0.007908632978796959, + 0.007664320059120655, + 0.00891032349318266, + 0.003743209410458803, + 0.009277284145355225, + 0.007307433523237705, + 0.006547159980982542, + 0.006876178085803986, + 0.007957159541547298, + 0.008764495141804218, + 0.007802958600223064, + 0.00563444197177887, + 0.009571559727191925, + 0.010882524773478508, + 0.008509061299264431, + 0.00655152415856719, + 0.008261535316705704, + 0.007026108913123608, + 0.004160647746175528, + 0.005775481928139925, + 0.0046567292883992195, + 0.0075925602577626705, + 0.0060643646866083145, + 0.0036670067347586155, + 0.005962823983281851, + 0.005285684484988451, + 0.006182982586324215, + 0.0044804844073951244, + 0.004692640155553818, + 0.011257583275437355, + 0.010025829076766968, + 0.011452654376626015, + 0.013394678942859173, + 0.011042363941669464, + 0.02441062405705452, + 0.014748561196029186, + 0.02112102136015892, + 0.014149566181004047, + 0.011367824859917164, + 0.009269600734114647, + 0.006402172148227692, + 0.004635010380297899, + 0.006036275532096624, + 0.008037266321480274, + 0.00799716915935278, + 0.004806932993233204, + 0.004696992691606283, + 0.004851022735238075, + 0.007390458602458239, + 0.009262977167963982, + 0.0067100971937179565, + 0.006449697073549032, + 0.005792108830064535, + 0.0055190338753163815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.31341788172721863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, with a subtle, natural shakiness, follows a young man from a slightly lower, level angle as he navigates a narrow Tokyo back alley. The boy has short, shaggy dark brown hair, a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers. The afternoon sun filters through the gaps between buildings and the crisscrossing electrical wires overhead, casting a shifting, interwoven pattern of light and dark on him and the rough concrete walls around him. He stops beside a vintage mint green scooter, scratched and peeling, leaning forward, one leg planted firmly on the damp asphalt, a playful smile on his focused face. He slams the starter with his right foot, but the engine only lets out a weak cough before silencing itself. Undeterred, he grins, and with amusement-filled persistence, he lifts his foot again, ready for another decisive attempt. The whole scene, with its shaky camera and warm lighting, is full of youthful energy and an effortless, casual sense of urban life, creating a relaxed and hopeful leisure atmosphere.", + "seed": 33486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_0/index_5886_seed_33486_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": -0.14017801452428102, + "vlm_logprob": -0.14017801452428102, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.14017801452428102, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31864020228385925 + ], + "opposite_logprob_tokens": [ + -1.3186402320861816 + ], + "logprob_reward": -0.31864020228385925, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31864020228385925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19799628853797913, + 0.12288729101419449, + 0.09723635017871857, + 0.09641281515359879, + 0.0818268358707428, + 0.08405209332704544, + 0.060945186764001846, + 0.030626516789197922, + 0.10207576304674149, + 0.12594084441661835 + ], + "credit_map": [ + 0.015094141475856304, + 0.019333912059664726, + 0.024302644655108452, + 0.03481199964880943, + 0.035746391862630844, + 0.02758733555674553, + 0.019279811531305313, + 0.021840043365955353, + 0.007339556235820055, + 0.010297983884811401, + 0.012654207646846771, + 0.029212161898612976, + 0.009763737209141254, + 0.021244633942842484, + 0.010560269467532635, + 0.021814744919538498, + 0.007052227854728699, + 0.007291278336197138, + 0.015539249405264854, + 0.012189614586532116, + 0.009024073369801044, + 0.013791502453386784, + 0.013602785766124725, + 0.018745621666312218, + 0.009544525295495987, + 0.008949648588895798, + 0.018728377297520638, + 0.01157724391669035, + 0.015479917638003826, + 0.01346083078533411, + 0.007446147035807371, + 0.011226121336221695, + 0.0077675278298556805, + 0.01229909434914589, + 0.013612269423902035, + 0.012243195436894894, + 0.013333424925804138, + 0.011743537150323391, + 0.0064783524721860886, + 0.004349429160356522, + 0.007156106643378735, + 0.008819539099931717, + 0.008773526176810265, + 0.012328725308179855, + 0.019481360912322998, + 0.011155536398291588, + 0.00949866883456707, + 0.006838632747530937, + 0.004122128710150719, + 0.005512637551873922, + 0.006563278380781412, + 0.0034639283549040556, + 0.007473522797226906, + 0.019223572686314583, + 0.004085536114871502, + 0.010500581003725529, + 0.002641637809574604, + 0.002103401580825448, + 0.0038798912428319454, + 0.0032938530202955008, + 0.004331214819103479, + 0.006859165616333485, + 0.003017426235601306, + 0.0044999257661402225, + 0.014633415266871452, + 0.009924369864165783, + 0.021187206730246544, + 0.013408539816737175, + 0.014330863021314144, + 0.01368597149848938, + 0.008621936663985252, + 0.0062834639102220535, + 0.016189610585570335, + 0.012736096978187561, + 0.01773504726588726, + 0.016169684007763863, + 0.026260389015078545, + 0.014875304885208607, + 0.013732126913964748, + 0.008242592215538025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.31864020228385925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867476224899292 + ], + "opposite_logprob_tokens": [ + -3.0486748218536377 + ], + "logprob_reward": -0.04867476224899292, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867476224899292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2024700790643692, + 0.2850170135498047, + 0.08898881822824478, + 0.07421466708183289, + 0.0362832210958004, + 0.03920971602201462, + 0.1115620881319046, + 0.03689118102192879, + 0.05144178494811058, + 0.07392138242721558 + ], + "credit_map": [ + 0.015467878431081772, + 0.026680827140808105, + 0.021510250866413116, + 0.030264876782894135, + 0.033835336565971375, + 0.02650330774486065, + 0.021703200414776802, + 0.026504412293434143, + 0.00900649931281805, + 0.014994644559919834, + 0.0240675937384367, + 0.06087498739361763, + 0.025853941217064857, + 0.03929978236556053, + 0.03673887625336647, + 0.0741807147860527, + 0.00739674549549818, + 0.007579408120363951, + 0.014444798231124878, + 0.010844134725630283, + 0.009732713922858238, + 0.011752644553780556, + 0.012142376974225044, + 0.01509600505232811, + 0.006231061182916164, + 0.00576512748375535, + 0.01873512752354145, + 0.00697722053155303, + 0.011771977879106998, + 0.009632510133087635, + 0.005509287118911743, + 0.00959235429763794, + 0.003548873821273446, + 0.003635536879301071, + 0.007253879681229591, + 0.0038452509324997663, + 0.005680181551724672, + 0.005288007669150829, + 0.00349175906740129, + 0.0035397340543568134, + 0.0036581067834049463, + 0.003847007406875491, + 0.0069452193565666676, + 0.006104704923927784, + 0.003691824385896325, + 0.005979431793093681, + 0.004828776698559523, + 0.004154647700488567, + 0.0067748744040727615, + 0.009992504492402077, + 0.012825115583837032, + 0.007606037892401218, + 0.021131642162799835, + 0.03298093006014824, + 0.007232424803078175, + 0.013018563389778137, + 0.0023089051246643066, + 0.0032282883767038584, + 0.004385252017527819, + 0.0043832287192344666, + 0.008445879444479942, + 0.008733455091714859, + 0.0026672312524169683, + 0.0027389442548155785, + 0.0038822703063488007, + 0.004327031318098307, + 0.007500463631004095, + 0.006457281764596701, + 0.007769453339278698, + 0.010329761542379856, + 0.005854856222867966, + 0.005320672877132893, + 0.00776560977101326, + 0.00755619490519166, + 0.010089882649481297, + 0.010139779187738895, + 0.01427516806870699, + 0.008968035690486431, + 0.009072071872651577, + 0.0060546486638486385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04867476224899292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.12210340052843094, + 0.12879282236099243, + 0.0531260147690773, + 0.10911878943443298, + 0.1013859286904335, + 0.07227756828069687, + 0.10090748220682144, + 0.056075796484947205, + 0.15194235742092133, + 0.1042698547244072 + ], + "credit_map": [ + 0.009616261348128319, + 0.01632072776556015, + 0.014231415465474129, + 0.020381886512041092, + 0.020343342795968056, + 0.014039517380297184, + 0.013980947434902191, + 0.013189302757382393, + 0.005210766103118658, + 0.009366830810904503, + 0.014875278808176517, + 0.026884382590651512, + 0.012673008255660534, + 0.03280377388000488, + 0.01048047374933958, + 0.01649831235408783, + 0.00484096584841609, + 0.007657763548195362, + 0.008602970279753208, + 0.007614284288138151, + 0.004588793497532606, + 0.00653837388381362, + 0.006652352400124073, + 0.006630513351410627, + 0.015824073925614357, + 0.011900359764695168, + 0.022190600633621216, + 0.013605616055428982, + 0.01240132562816143, + 0.014930774457752705, + 0.008252481929957867, + 0.010013559833168983, + 0.008436956442892551, + 0.012490157037973404, + 0.01862151362001896, + 0.013196456246078014, + 0.014631472527980804, + 0.019649742171168327, + 0.008729524910449982, + 0.005630103405565023, + 0.00706844637170434, + 0.010223029181361198, + 0.014242008328437805, + 0.01488155871629715, + 0.006175810005515814, + 0.009888251312077045, + 0.005240539088845253, + 0.0045579276047647, + 0.0066964770667254925, + 0.008805948309600353, + 0.013042021542787552, + 0.007250951603055, + 0.011271363124251366, + 0.03282159939408302, + 0.00654213922098279, + 0.01447698287665844, + 0.004599286243319511, + 0.006159517448395491, + 0.008769466541707516, + 0.008006343618035316, + 0.012872472405433655, + 0.008341637440025806, + 0.003955270629376173, + 0.003371799597516656, + 0.01727425493299961, + 0.012654142454266548, + 0.02327364683151245, + 0.021921375766396523, + 0.01732581853866577, + 0.03078649565577507, + 0.014263242483139038, + 0.01444337423890829, + 0.012350055389106274, + 0.01375845167785883, + 0.01942308247089386, + 0.012904640287160873, + 0.012833427637815475, + 0.01397367101162672, + 0.009847895242273808, + 0.009178630076348782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.014191891998052597 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059472741559147835 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.0059472741559147835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059472741559147835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15772059559822083, + 0.13650096952915192, + 0.07541164755821228, + 0.11060357093811035, + 0.08137965202331543, + 0.11242847889661789, + 0.10830151289701462, + 0.06583619117736816, + 0.03387902304530144, + 0.11793839931488037 + ], + "credit_map": [ + 0.013294395990669727, + 0.02746022678911686, + 0.016159547492861748, + 0.018798982724547386, + 0.03307546675205231, + 0.023713866248726845, + 0.010744638741016388, + 0.01447346992790699, + 0.006711537018418312, + 0.010226060636341572, + 0.016681542620062828, + 0.02732853963971138, + 0.017453745007514954, + 0.028218941763043404, + 0.014579837210476398, + 0.015300768427550793, + 0.005536968354135752, + 0.006512932479381561, + 0.011252133175730705, + 0.009631508029997349, + 0.01188859436661005, + 0.011594478972256184, + 0.00855079386383295, + 0.01044423971325159, + 0.010779013857245445, + 0.011713068932294846, + 0.02495270036160946, + 0.011668986640870571, + 0.015033052302896976, + 0.01530655100941658, + 0.010013381019234657, + 0.011136817745864391, + 0.004764769226312637, + 0.006669595371931791, + 0.011978082358837128, + 0.0071891434490680695, + 0.0120597705245018, + 0.020069530233740807, + 0.009542756713926792, + 0.009106002748012543, + 0.00566954305395484, + 0.008154422976076603, + 0.010897668078541756, + 0.013712044805288315, + 0.027657486498355865, + 0.019037431105971336, + 0.017092199996113777, + 0.01020768377929926, + 0.005691729951649904, + 0.008625165559351444, + 0.01059967651963234, + 0.008196590468287468, + 0.019706828519701958, + 0.02381693571805954, + 0.015056315809488297, + 0.016608264297246933, + 0.0050999317318201065, + 0.005090442951768637, + 0.011205642484128475, + 0.007341780699789524, + 0.010304590687155724, + 0.013486829586327076, + 0.0050860000774264336, + 0.008220969699323177, + 0.0036310080904513597, + 0.0028525860980153084, + 0.006727208849042654, + 0.00431325426325202, + 0.005771694704890251, + 0.005098416935652494, + 0.0026116534136235714, + 0.0028732018545269966, + 0.01719564199447632, + 0.012639355845749378, + 0.019311001524329185, + 0.014565753750503063, + 0.01860162615776062, + 0.014233004301786423, + 0.013557838276028633, + 0.007834179326891899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0059472741559147835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with a subtle, natural shakiness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with short, shaggy dark brown hair wearing a loose-fitting gray hoodie, faded jeans, and scuffed white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow Tokyo back alley with crisscrossing electrical wires overhead and rough concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vintage mint green scooter that is scratched and peeling, positioned next to the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slam the scooter starter with his right foot, resulting in a weak cough before the engine silences?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343594193458557 + ], + "opposite_logprob_tokens": [ + -1.3134359121322632 + ], + "logprob_reward": -0.31343594193458557, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343594193458557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08638451993465424, + 0.15594495832920074, + 0.23337103426456451, + 0.05217749997973442, + 0.041003379970788956, + 0.08987251669168472, + 0.1278965175151825, + 0.03696122765541077, + 0.07294481992721558, + 0.10344352573156357 + ], + "credit_map": [ + 0.005638706963509321, + 0.009610315784811974, + 0.01059415191411972, + 0.016344059258699417, + 0.011666333302855492, + 0.011308053508400917, + 0.008954023942351341, + 0.012268883176147938, + 0.00907843466848135, + 0.009700408205389977, + 0.017206525430083275, + 0.025465726852416992, + 0.01360850315541029, + 0.0401436910033226, + 0.02047625370323658, + 0.020265433937311172, + 0.017223751172423363, + 0.023022912442684174, + 0.025276796892285347, + 0.024419009685516357, + 0.02863895706832409, + 0.02777193859219551, + 0.03091106191277504, + 0.05610661208629608, + 0.006558320485055447, + 0.0054389857687056065, + 0.010620711371302605, + 0.0064580729231238365, + 0.00533866323530674, + 0.0062786429189145565, + 0.004789254628121853, + 0.0066948505118489265, + 0.003795407246798277, + 0.002999840071424842, + 0.006698825862258673, + 0.004787512589246035, + 0.005790111143141985, + 0.00803329050540924, + 0.004756506532430649, + 0.004141887184232473, + 0.00656034704297781, + 0.006105727981775999, + 0.009460262022912502, + 0.011861409060657024, + 0.023377884179353714, + 0.014660485088825226, + 0.010570601560175419, + 0.007275800220668316, + 0.007875462993979454, + 0.010622178204357624, + 0.014293507672846317, + 0.009629095904529095, + 0.013673827983438969, + 0.03747427463531494, + 0.011454392224550247, + 0.02287379838526249, + 0.0027832724153995514, + 0.0026725493371486664, + 0.004805200267583132, + 0.00450014416128397, + 0.00459302868694067, + 0.005317086353898048, + 0.006338498555123806, + 0.00595144834369421, + 0.005576269701123238, + 0.005767321679741144, + 0.011843536049127579, + 0.008141233585774899, + 0.010527962818741798, + 0.01584448292851448, + 0.0066685364581644535, + 0.008575479499995708, + 0.01630403846502304, + 0.011310522444546223, + 0.018042167648673058, + 0.009464707225561142, + 0.015231247060000896, + 0.013267910107970238, + 0.011038370430469513, + 0.008784574456512928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.31343594193458557 + } + ], + "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/0000125/group_1/index_tensor([911, 911])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_1/index_tensor([911, 911])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5d45b67df488f0d2aef7bdebe13dc4e89e4cbcbb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_1/index_tensor([911, 911])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, shot from a horizontal angle, depicts a serene, gender-ambiguous angel bathed in dreamy light. A pair of vast, pristine white wings unfurl behind them, each feather clearly visible and edged with a soft silver halo. A cascade of silver-white hair gently falls over their shoulders. The angel, clad in a white robe of translucent silk, holds a gaze filled with compassion and warmth. Slowly and firmly, they extend their right hand toward the camera. A warm, luminous golden glow emanates from their palm and fingertips, like a candlelight shrouded in mist. This light gently illuminates the surrounding space and casts dynamic shadows on the divine being's face and robes. The entire frame is in soft focus, with the background blurred, leaving only a faint, abstract bokeh of golden and pale blue specks. This inviting and soothing gesture, combined with the warm tones and tranquil atmosphere, creates a sacred, healing, and otherworldly visual experience.", + "seed": 5349, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_1/index_911_seed_5349_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": -0.026201892807148398, + "vlm_logprob": -0.026201892807148398, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.026201892807148398, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0394199900329113 + ], + "opposite_logprob_tokens": [ + -3.2894198894500732 + ], + "logprob_reward": -0.0394199900329113, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0394199900329113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15341709554195404, + 0.019272612407803535, + 0.023361044004559517, + 0.010606198571622372, + 0.017812350764870644, + 0.4782584309577942, + 0.13953545689582825, + 0.035908084362745285, + 0.08824069797992706, + 0.03358807414770126 + ], + "credit_map": [ + 0.010557138361036777, + 0.010570290498435497, + 0.05110277980566025, + 0.05255495756864548, + 0.008175610564649105, + 0.007009518798440695, + 0.006239903625100851, + 0.00720688421279192, + 0.0023038946092128754, + 0.002639270154759288, + 0.0032031587325036526, + 0.0021041170693933964, + 0.0021553384140133858, + 0.002585255540907383, + 0.0020266552455723286, + 0.0022549210116267204, + 0.0027251599822193384, + 0.0030514562968164682, + 0.0023947556037455797, + 0.0022801768500357866, + 0.0053755114786326885, + 0.00196364289149642, + 0.0029631888028234243, + 0.002607149537652731, + 0.0012884736061096191, + 0.001180312829092145, + 0.0014735240256413817, + 0.0014209810178726912, + 0.0017930951435118914, + 0.0012442533625289798, + 0.0012460720026865602, + 0.0009594859438948333, + 0.0016647365409880877, + 0.0017655326519161463, + 0.0025107930414378643, + 0.0042940424755215645, + 0.0017776450840756297, + 0.0021403031423687935, + 0.0019997155759483576, + 0.0016595791094005108, + 0.03264429792761803, + 0.03305600583553314, + 0.1730598509311676, + 0.1915324330329895, + 0.008117415942251682, + 0.015478291548788548, + 0.013892008922994137, + 0.010478077456355095, + 0.019444171339273453, + 0.011576851829886436, + 0.04050559550523758, + 0.016105590388178825, + 0.010938088409602642, + 0.01496225968003273, + 0.011630457825958729, + 0.014372424222528934, + 0.009903151541948318, + 0.003168189898133278, + 0.0030992699321359396, + 0.009034553542733192, + 0.0028497257735580206, + 0.0026262213941663504, + 0.002946696011349559, + 0.002280274871736765, + 0.00837624166160822, + 0.010564553551375866, + 0.009580559097230434, + 0.01384298037737608, + 0.016684845089912415, + 0.010285595431923866, + 0.011496343649923801, + 0.007409570273011923, + 0.002990748966112733, + 0.003009541891515255, + 0.0029422533698379993, + 0.006280609406530857, + 0.0050758132711052895, + 0.0035307896323502064, + 0.002837198553606868, + 0.006921113934367895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0394199900329113 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640542928129435 + ], + "opposite_logprob_tokens": [ + -4.882640361785889 + ], + "logprob_reward": -0.007640542928129435, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640542928129435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0658416599035263, + 0.015622618608176708, + 0.013370866887271404, + 0.009181506931781769, + 0.017790280282497406, + 0.6740418076515198, + 0.07631255686283112, + 0.07348011434078217, + 0.037148259580135345, + 0.0172103401273489 + ], + "credit_map": [ + 0.006912996061146259, + 0.00459388829767704, + 0.01966189034283161, + 0.022439725697040558, + 0.002863594563677907, + 0.0027806819416582584, + 0.003073649713769555, + 0.003515239804983139, + 0.0025076933670789003, + 0.001958058914169669, + 0.0018700615037232637, + 0.0016095638275146484, + 0.0017414819449186325, + 0.001546006416901946, + 0.0022781668230891228, + 0.002111587440595031, + 0.0017963792197406292, + 0.0017177866538986564, + 0.0014237918658182025, + 0.0014428923605009913, + 0.0019503324292600155, + 0.0009406639146618545, + 0.002291445154696703, + 0.0018075763946399093, + 0.001095385174266994, + 0.0009378525428473949, + 0.0010757518466562033, + 0.0008760885102674365, + 0.0020629114005714655, + 0.0008018426015041769, + 0.0013831760734319687, + 0.0009484993061050773, + 0.001769471331499517, + 0.0016487122047692537, + 0.0038099861703813076, + 0.0036319191567599773, + 0.001970923040062189, + 0.0014433537144213915, + 0.0018049993086606264, + 0.00171091768424958, + 0.045918840914964676, + 0.04571584612131119, + 0.24644839763641357, + 0.26978152990341187, + 0.011362416669726372, + 0.021349726244807243, + 0.019227055832743645, + 0.01423803623765707, + 0.013105932623147964, + 0.00700807711109519, + 0.017623623833060265, + 0.015145461075007915, + 0.005093690473586321, + 0.007556037977337837, + 0.004813685547560453, + 0.005966059397906065, + 0.022170744836330414, + 0.00643760384991765, + 0.004519451875239611, + 0.025834282860159874, + 0.0033173353876918554, + 0.0037046121433377266, + 0.004177620634436607, + 0.0033184653148055077, + 0.003725233720615506, + 0.00478760851547122, + 0.0042070550844073296, + 0.006733907386660576, + 0.005724023096263409, + 0.004769562743604183, + 0.004007040522992611, + 0.0031938303727656603, + 0.0014524566940963268, + 0.0018359312089160085, + 0.001339161884970963, + 0.003926802426576614, + 0.0023233334068208933, + 0.0016462516505271196, + 0.0012534934794530272, + 0.0034329122863709927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.007640542928129435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204066798090935 + ], + "opposite_logprob_tokens": [ + -2.8120405673980713 + ], + "logprob_reward": -0.06204066798090935, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204066798090935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07960595190525055, + 0.026747776195406914, + 0.024903008714318275, + 0.01689995266497135, + 0.09404116123914719, + 0.217606320977211, + 0.038968514651060104, + 0.2011912614107132, + 0.08784496784210205, + 0.21219103038311005 + ], + "credit_map": [ + 0.012815034948289394, + 0.008209985680878162, + 0.019162369892001152, + 0.01038455218076706, + 0.008478578180074692, + 0.005520286038517952, + 0.005663767922669649, + 0.009371381253004074, + 0.00406888360157609, + 0.002139607910066843, + 0.004558780696243048, + 0.002770394319668412, + 0.004852835088968277, + 0.002261134097352624, + 0.002805774100124836, + 0.0032903708051890135, + 0.0026565841399133205, + 0.003929244354367256, + 0.0033047401811927557, + 0.00228948169387877, + 0.005011141300201416, + 0.0023072545882314444, + 0.0031788675114512444, + 0.0022256984375417233, + 0.0021136566065251827, + 0.0018840719712898135, + 0.0022034901194274426, + 0.0027818847447633743, + 0.002945677610114217, + 0.0016503368970006704, + 0.0015174878062680364, + 0.0018033486558124423, + 0.007632717955857515, + 0.006508363410830498, + 0.024477357044816017, + 0.030048221349716187, + 0.006012982223182917, + 0.005028498824685812, + 0.006689158733934164, + 0.007643865421414375, + 0.015700574964284897, + 0.014416621066629887, + 0.07136600464582443, + 0.07550179213285446, + 0.01015148963779211, + 0.010779506526887417, + 0.010198140516877174, + 0.009492217563092709, + 0.0034645507112145424, + 0.003443522145971656, + 0.008861127309501171, + 0.006672359071671963, + 0.0038790658582001925, + 0.004358163569122553, + 0.0047185844741761684, + 0.003571145934984088, + 0.05439118295907974, + 0.021768759936094284, + 0.012981920503079891, + 0.08279139548540115, + 0.005152713507413864, + 0.008552202954888344, + 0.007086697965860367, + 0.008466408587992191, + 0.01224518846720457, + 0.013012324459850788, + 0.009792876429855824, + 0.010547713376581669, + 0.012001272290945053, + 0.013017428107559681, + 0.00941920280456543, + 0.007808974012732506, + 0.03852546587586403, + 0.019137170165777206, + 0.02117958292365074, + 0.06272409111261368, + 0.01624477282166481, + 0.019100796431303024, + 0.01264876313507557, + 0.022630419582128525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.06204066798090935 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013338964199647307 + ], + "opposite_logprob_tokens": [ + -6.626333713531494 + ], + "logprob_reward": -0.0013338964199647307, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013338964199647307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15590670704841614, + 0.06099238246679306, + 0.058966293931007385, + 0.04227957874536514, + 0.052621182054281235, + 0.36551612615585327, + 0.07600487023591995, + 0.12001677602529526, + 0.029617996886372566, + 0.038078051060438156 + ], + "credit_map": [ + 0.017230510711669922, + 0.013881010003387928, + 0.036074526607990265, + 0.04514176771044731, + 0.00992070697247982, + 0.009297206066548824, + 0.01245430763810873, + 0.011906680651009083, + 0.009205958805978298, + 0.006207237485796213, + 0.0059744310565292835, + 0.005965434480458498, + 0.008875129744410515, + 0.005579035729169846, + 0.01077378448098898, + 0.008411369286477566, + 0.006524933036416769, + 0.008773785084486008, + 0.005608674138784409, + 0.006003555376082659, + 0.008398881182074547, + 0.004205853678286076, + 0.011227726936340332, + 0.008222883567214012, + 0.004828609060496092, + 0.005683918483555317, + 0.004103378858417273, + 0.004095430485904217, + 0.00832549948245287, + 0.003827387699857354, + 0.006327275652438402, + 0.005088078323751688, + 0.003996522165834904, + 0.003414764767512679, + 0.01237273309379816, + 0.014099014922976494, + 0.0045196786522865295, + 0.002838431391865015, + 0.006072138901799917, + 0.00530789652839303, + 0.023431535810232162, + 0.023220228031277657, + 0.1325569748878479, + 0.13601715862751007, + 0.012712519615888596, + 0.011891414411365986, + 0.012401140294969082, + 0.013285167515277863, + 0.01326669380068779, + 0.006437557749450207, + 0.020276295021176338, + 0.010895224288105965, + 0.006253004539757967, + 0.007685147225856781, + 0.004686548840254545, + 0.006504397373646498, + 0.030057689175009727, + 0.013456618413329124, + 0.00786874070763588, + 0.0467853769659996, + 0.004286190494894981, + 0.005387210752815008, + 0.006654683966189623, + 0.0055202702060341835, + 0.003027202794328332, + 0.00420643575489521, + 0.0034845995251089334, + 0.006888898089528084, + 0.0034403533209115267, + 0.003423525020480156, + 0.002208539517596364, + 0.0029384419322013855, + 0.006702026817947626, + 0.003364073345437646, + 0.0032137534581124783, + 0.012374664656817913, + 0.0027014336083084345, + 0.003103567287325859, + 0.0021126095671206713, + 0.004505923017859459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0013338964199647307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057436667382717 + ], + "opposite_logprob_tokens": [ + -3.8955743312835693 + ], + "logprob_reward": -0.02057436667382717, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057436667382717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10244638472795486, + 0.020995158702135086, + 0.011676911264657974, + 0.02054072543978691, + 0.026858368888497353, + 0.07148906588554382, + 0.584670901298523, + 0.1054370105266571, + 0.013859361410140991, + 0.04202605038881302 + ], + "credit_map": [ + 0.008878928609192371, + 0.0066576553508639336, + 0.031367380172014236, + 0.027420027181506157, + 0.010097087360918522, + 0.006011637859046459, + 0.006277804262936115, + 0.00573586393147707, + 0.003021832089871168, + 0.0017795312451198697, + 0.002570329699665308, + 0.0023017891217023134, + 0.00513010798022151, + 0.0020973961800336838, + 0.002042237436398864, + 0.002051933202892542, + 0.0014754607109352946, + 0.0011264930944889784, + 0.0015332287875935435, + 0.0011842481326311827, + 0.0016554435715079308, + 0.0010834287386387587, + 0.001928630517795682, + 0.0016899779438972473, + 0.002616819692775607, + 0.0030188921373337507, + 0.002616734243929386, + 0.0026959010865539312, + 0.002614699536934495, + 0.0016769067151471972, + 0.0027830717153847218, + 0.0025177004281431437, + 0.0036217495799064636, + 0.0021279011853039265, + 0.004810039885342121, + 0.004627373069524765, + 0.0039792414754629135, + 0.0025323606096208096, + 0.00271380296908319, + 0.002445901045575738, + 0.007340400479733944, + 0.005646063946187496, + 0.017453555017709732, + 0.015464193187654018, + 0.007699326612055302, + 0.006441619712859392, + 0.005355439614504576, + 0.006088468246161938, + 0.0806720033288002, + 0.04434357210993767, + 0.18180207908153534, + 0.07922855764627457, + 0.04269038513302803, + 0.06157892942428589, + 0.04025166481733322, + 0.054103702306747437, + 0.032615743577480316, + 0.01003136858344078, + 0.005768528673797846, + 0.04220413416624069, + 0.002335775876417756, + 0.004233678802847862, + 0.003836150746792555, + 0.004411625675857067, + 0.002649025060236454, + 0.0014160634018480778, + 0.0014730303082615137, + 0.001505028922110796, + 0.0020397689659148455, + 0.0011501635890454054, + 0.0019090953283011913, + 0.001717186183668673, + 0.008749696426093578, + 0.003498856443911791, + 0.004590772092342377, + 0.012144478969275951, + 0.003259560326114297, + 0.003670671721920371, + 0.0026110203471034765, + 0.003500994760543108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.02057436667382717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot from a horizontal angle, depicts a serene, gender-ambiguous angel bathed in dreamy light. A pair of vast, pristine white wings unfurl behind them, each feather clearly visible and edged with a soft silver halo. A cascade of silver-white hair gently falls over their shoulders. The angel, clad in a white robe of translucent silk, holds a gaze filled with compassion and warmth. Slowly and firmly, they extend their right hand toward the camera. A warm, luminous golden glow emanates from their palm and fingertips, like a candlelight shrouded in mist. This light gently illuminates the surrounding space and casts dynamic shadows on the divine being's face and robes. The entire frame is in soft focus, with the background blurred, leaving only a faint, abstract bokeh of golden and pale blue specks. This inviting and soothing gesture, combined with the warm tones and tranquil atmosphere, creates a sacred, healing, and otherworldly visual experience.", + "seed": 5350, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_1/index_911_seed_5350_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": -0.009785236325114965, + "vlm_logprob": -0.009785236325114965, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.009785236325114965, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017018355429172516 + ], + "opposite_logprob_tokens": [ + -4.2670183181762695 + ], + "logprob_reward": -0.017018355429172516, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017018355429172516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21610736846923828, + 0.18154780566692352, + 0.19170315563678741, + 0.07278775423765182, + 0.10041913390159607, + 0.021434813737869263, + 0.033007778227329254, + 0.06689605116844177, + 0.04160839691758156, + 0.07448767125606537 + ], + "credit_map": [ + 0.05064811930060387, + 0.02017957717180252, + 0.030312571674585342, + 0.024206554517149925, + 0.020402904599905014, + 0.017318500205874443, + 0.016705399379134178, + 0.036333758383989334, + 0.014335297979414463, + 0.0067385113798081875, + 0.05690814182162285, + 0.05188096687197685, + 0.006498169619590044, + 0.012165234424173832, + 0.01167947705835104, + 0.021342018619179726, + 0.013825483620166779, + 0.011455715633928776, + 0.018797224387526512, + 0.06996600329875946, + 0.009527587331831455, + 0.011755941435694695, + 0.0129332784563303, + 0.04344194382429123, + 0.006846219301223755, + 0.006498393602669239, + 0.02058948576450348, + 0.005217249039560556, + 0.005435158032923937, + 0.005931489169597626, + 0.004519427195191383, + 0.017750333994627, + 0.00772100267931819, + 0.006789564620703459, + 0.011812019161880016, + 0.021719496697187424, + 0.005205594934523106, + 0.009386156685650349, + 0.005169137846678495, + 0.03261617198586464, + 0.003831773530691862, + 0.001893893931992352, + 0.002684986684471369, + 0.0022165297996252775, + 0.0015818871324881911, + 0.0024876194074749947, + 0.001681957277469337, + 0.005056166090071201, + 0.0027583898045122623, + 0.0027296796906739473, + 0.003972547594457865, + 0.007445738185197115, + 0.0035502654500305653, + 0.0028700947295874357, + 0.005941013805568218, + 0.003740053391084075, + 0.01027968991547823, + 0.005486434791237116, + 0.010006219148635864, + 0.010357892140746117, + 0.005720661021769047, + 0.009407338686287403, + 0.007084387354552746, + 0.008553430438041687, + 0.005776038393378258, + 0.003297958755865693, + 0.0059804716147482395, + 0.00731565011665225, + 0.003732898738235235, + 0.00426857266575098, + 0.007630010135471821, + 0.003606799291446805, + 0.013145800679922104, + 0.005796864163130522, + 0.011876745149493217, + 0.008709562942385674, + 0.008639131672680378, + 0.009212817065417767, + 0.006447240710258484, + 0.010659519582986832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.017018355429172516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561134994029999 + ], + "opposite_logprob_tokens": [ + -4.387561321258545 + ], + "logprob_reward": -0.012561134994029999, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561134994029999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03876562416553497, + 0.13252192735671997, + 0.035408783704042435, + 0.07763515412807465, + 0.0911538153886795, + 0.06936313211917877, + 0.01714882254600525, + 0.3635293245315552, + 0.13485237956047058, + 0.03962103649973869 + ], + "credit_map": [ + 0.006132259964942932, + 0.003627755679190159, + 0.0052320086397230625, + 0.006447840016335249, + 0.004851824138313532, + 0.0034043577034026384, + 0.004113318398594856, + 0.004956257529556751, + 0.009149940684437752, + 0.004986075684428215, + 0.04154764115810394, + 0.03789253532886505, + 0.004234286490827799, + 0.008464585058391094, + 0.008492100983858109, + 0.01775476336479187, + 0.003995963837951422, + 0.0033988317009061575, + 0.0055910637602210045, + 0.007245552726089954, + 0.005127358715981245, + 0.003095826134085655, + 0.0041320607997477055, + 0.0028221248649060726, + 0.006887895055115223, + 0.004913282580673695, + 0.02738090790808201, + 0.016977017745375633, + 0.0028412521351128817, + 0.003495729062706232, + 0.0038061554078012705, + 0.011332917958498001, + 0.008391785435378551, + 0.006544061936438084, + 0.02779628336429596, + 0.005422550719231367, + 0.004226448014378548, + 0.00653923861682415, + 0.005478995386511087, + 0.02675444819033146, + 0.011865736916661263, + 0.006112640257924795, + 0.009093040600419044, + 0.007728108670562506, + 0.007462129462510347, + 0.009432150050997734, + 0.006009206175804138, + 0.011660116724669933, + 0.0016745188040658832, + 0.0014662259491160512, + 0.0023912531323730946, + 0.003361563663929701, + 0.0017767451936379075, + 0.001280802651308477, + 0.003730278927832842, + 0.0014674323610961437, + 0.05481972545385361, + 0.028677545487880707, + 0.05574227496981621, + 0.049902964383363724, + 0.03269324451684952, + 0.05232293903827667, + 0.040891360491514206, + 0.04847924783825874, + 0.01842624694108963, + 0.010102261789143085, + 0.02017030492424965, + 0.02178906835615635, + 0.011612070724368095, + 0.014223155565559864, + 0.026221703737974167, + 0.012307549826800823, + 0.005334229674190283, + 0.0031337966211140156, + 0.005557684693485498, + 0.008630428463220596, + 0.0040747253224253654, + 0.00323045183904469, + 0.005644736345857382, + 0.004014982841908932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.012561134994029999 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952843930572271 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952843930572271, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952843930572271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06057910993695259, + 0.1568618267774582, + 0.12693467736244202, + 0.06333888322114944, + 0.057880181819200516, + 0.010123468935489655, + 0.02320769801735878, + 0.032985154539346695, + 0.06528334319591522, + 0.4028056859970093 + ], + "credit_map": [ + 0.01232580654323101, + 0.005718223284929991, + 0.008086592890322208, + 0.00628268625587225, + 0.007895758375525475, + 0.006071867421269417, + 0.00604486558586359, + 0.008153310976922512, + 0.010455292649567127, + 0.006592070683836937, + 0.044447291642427444, + 0.061794862151145935, + 0.005235944874584675, + 0.008684518747031689, + 0.012927737087011337, + 0.006724107079207897, + 0.008076753467321396, + 0.0071922545321285725, + 0.011991296894848347, + 0.0476042814552784, + 0.005266305059194565, + 0.00804939866065979, + 0.008398125879466534, + 0.030356265604496002, + 0.0051419297233223915, + 0.0036875733640044928, + 0.02249869331717491, + 0.014533616602420807, + 0.002290604868903756, + 0.0030053025111556053, + 0.0032002010848373175, + 0.008980964310467243, + 0.005444637965410948, + 0.0041074687615036964, + 0.013194626197218895, + 0.005348282400518656, + 0.0028163448441773653, + 0.005142693407833576, + 0.002581452950835228, + 0.01924467459321022, + 0.0014310519909486175, + 0.0014266542857512832, + 0.0011124092852696776, + 0.001486757886596024, + 0.001009072526358068, + 0.0012304690899327397, + 0.0010785351041704416, + 0.001348518650047481, + 0.0024422151036560535, + 0.0020342934876680374, + 0.0031072606798261404, + 0.004741689655929804, + 0.0020511432085186243, + 0.0017983942525461316, + 0.004655768163502216, + 0.0023769333492964506, + 0.004491229075938463, + 0.00287855532951653, + 0.005033215042203665, + 0.004446769133210182, + 0.0029093995690345764, + 0.004924638196825981, + 0.004780255723744631, + 0.003521092003211379, + 0.010848969221115112, + 0.005581020377576351, + 0.00905730202794075, + 0.00938086211681366, + 0.004768000915646553, + 0.007627705577760935, + 0.01090392004698515, + 0.007115561049431562, + 0.035817794501781464, + 0.028335319831967354, + 0.06704321503639221, + 0.10006070137023926, + 0.036945879459381104, + 0.025445256382226944, + 0.07347248494625092, + 0.03568504378199577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.005952843930572271 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036151548847556114 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036151548847556114, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036151548847556114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04340227320790291, + 0.0831729993224144, + 0.0574742890894413, + 0.06703030318021774, + 0.03006156161427498, + 0.012618733569979668, + 0.02056233398616314, + 0.21595610678195953, + 0.39662614464759827, + 0.07309524714946747 + ], + "credit_map": [ + 0.0099337138235569, + 0.004404525272548199, + 0.00761173851788044, + 0.005036415997892618, + 0.0020406548865139484, + 0.0034376392140984535, + 0.002914030570536852, + 0.008023555390536785, + 0.00527470838278532, + 0.0034859615843743086, + 0.017469754442572594, + 0.034535326063632965, + 0.002727295970544219, + 0.0038517224602401257, + 0.006567561067640781, + 0.009260665625333786, + 0.0035646979231387377, + 0.0033023625146597624, + 0.005121217109262943, + 0.02170761488378048, + 0.002052395138889551, + 0.0035736209247261286, + 0.003461143933236599, + 0.01469123363494873, + 0.00624722708016634, + 0.00452815555036068, + 0.02374206855893135, + 0.01274049747735262, + 0.002339782193303108, + 0.0030913662631064653, + 0.0030728187412023544, + 0.011268390342593193, + 0.002375267678871751, + 0.0020073349587619305, + 0.004350672941654921, + 0.005712543614208698, + 0.001425706665031612, + 0.0027080699801445007, + 0.0015415706438943744, + 0.009940393269062042, + 0.002170110121369362, + 0.0015234352322295308, + 0.0015454163076356053, + 0.0013721673749387264, + 0.0012197218602523208, + 0.0014925012364983559, + 0.0009992298437282443, + 0.0022961513604968786, + 0.002559957792982459, + 0.0016184905543923378, + 0.003173337783664465, + 0.0031526382081210613, + 0.002016545971855521, + 0.001755199278704822, + 0.004248726647347212, + 0.0020374380983412266, + 0.03304941579699516, + 0.01662345789372921, + 0.03364339470863342, + 0.03119601681828499, + 0.018351929262280464, + 0.029919078573584557, + 0.02385787107050419, + 0.029314935207366943, + 0.05438537895679474, + 0.030731264501810074, + 0.058713577687740326, + 0.061700548976659775, + 0.03622009605169296, + 0.0426512248814106, + 0.07489008456468582, + 0.03733396530151367, + 0.00974289420992136, + 0.006007035728543997, + 0.008403845131397247, + 0.014939501881599426, + 0.009015356190502644, + 0.007116096094250679, + 0.009169050492346287, + 0.008701472543179989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.0036151548847556114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977869238704443 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.00977869238704443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977869238704443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05383670702576637, + 0.12638016045093536, + 0.07925094664096832, + 0.18014082312583923, + 0.06706033647060394, + 0.056161727756261826, + 0.026571203023195267, + 0.18802058696746826, + 0.15205521881580353, + 0.07052228599786758 + ], + "credit_map": [ + 0.007842207327485085, + 0.0040819100104272366, + 0.013394227251410484, + 0.011018084362149239, + 0.0033721833024173975, + 0.002809954108670354, + 0.0034158756025135517, + 0.007902259938418865, + 0.012130018323659897, + 0.005160334520041943, + 0.03979664295911789, + 0.042091064155101776, + 0.005378939211368561, + 0.006480293348431587, + 0.008380480110645294, + 0.0069623771123588085, + 0.004754591267555952, + 0.004405671264976263, + 0.007017566356807947, + 0.02852673828601837, + 0.004636448808014393, + 0.005259147845208645, + 0.005776089150458574, + 0.018874680623412132, + 0.01570972613990307, + 0.01268120389431715, + 0.06539095938205719, + 0.02846432849764824, + 0.004030917305499315, + 0.008621697314083576, + 0.0066801332868635654, + 0.03856184333562851, + 0.005815610755234957, + 0.0038952920585870743, + 0.012976367957890034, + 0.020134784281253815, + 0.0034635430201888084, + 0.005339180584996939, + 0.004350120201706886, + 0.011085429228842258, + 0.010022052563726902, + 0.005850164219737053, + 0.00784118939191103, + 0.005342752672731876, + 0.005584316328167915, + 0.007003491278737783, + 0.004478741902858019, + 0.010039013810455799, + 0.0025378677528351545, + 0.0021973983384668827, + 0.003947672434151173, + 0.006889520678669214, + 0.002417836105450988, + 0.0014719319297000766, + 0.005139337386935949, + 0.001969632226973772, + 0.02951459027826786, + 0.014928539283573627, + 0.028978029265999794, + 0.02566945180296898, + 0.015861593186855316, + 0.026861857622861862, + 0.021242346614599228, + 0.024964163079857826, + 0.02131860889494419, + 0.011413096450269222, + 0.022557450458407402, + 0.023875553160905838, + 0.012739935889840126, + 0.01631801202893257, + 0.02939555048942566, + 0.014436988160014153, + 0.006995752919465303, + 0.005142254289239645, + 0.01086920965462923, + 0.018061233684420586, + 0.006748904939740896, + 0.004584026988595724, + 0.011861627921462059, + 0.006259268615394831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.00977869238704443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot from a horizontal angle, depicts a serene, gender-ambiguous angel bathed in dreamy light. A pair of vast, pristine white wings unfurl behind them, each feather clearly visible and edged with a soft silver halo. A cascade of silver-white hair gently falls over their shoulders. The angel, clad in a white robe of translucent silk, holds a gaze filled with compassion and warmth. Slowly and firmly, they extend their right hand toward the camera. A warm, luminous golden glow emanates from their palm and fingertips, like a candlelight shrouded in mist. This light gently illuminates the surrounding space and casts dynamic shadows on the divine being's face and robes. The entire frame is in soft focus, with the background blurred, leaving only a faint, abstract bokeh of golden and pale blue specks. This inviting and soothing gesture, combined with the warm tones and tranquil atmosphere, creates a sacred, healing, and otherworldly visual experience.", + "seed": 5351, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_1/index_911_seed_5351_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": -0.030189206823706627, + "vlm_logprob": -0.030189206823706627, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.030189206823706627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018299441784620285 + ], + "opposite_logprob_tokens": [ + -4.268299579620361 + ], + "logprob_reward": -0.018299441784620285, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018299441784620285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045185454189777374, + 0.04289233312010765, + 0.03148903697729111, + 0.15968354046344757, + 0.023127451539039612, + 0.01931389793753624, + 0.05522317439317703, + 0.5935078859329224, + 0.01629602164030075, + 0.013281209394335747 + ], + "credit_map": [ + 0.004199721850454807, + 0.005767017137259245, + 0.0055204229429364204, + 0.006541032809764147, + 0.009048474952578545, + 0.004722996614873409, + 0.00450379541143775, + 0.004881991073489189, + 0.003770550712943077, + 0.006012418773025274, + 0.004516685381531715, + 0.005298632197082043, + 0.010266738012433052, + 0.003895883448421955, + 0.004767183680087328, + 0.004364240914583206, + 0.002342677442356944, + 0.005464354529976845, + 0.002562374109402299, + 0.003271367633715272, + 0.006905112415552139, + 0.002483133226633072, + 0.004222497344017029, + 0.004237520508468151, + 0.01449486892670393, + 0.018081380054354668, + 0.028722012415528297, + 0.02898034267127514, + 0.013690367341041565, + 0.011656152084469795, + 0.019603529945015907, + 0.024454891681671143, + 0.002696610987186432, + 0.002925829729065299, + 0.0024884419981390238, + 0.0016140507068485022, + 0.0032144004944711924, + 0.001989741576835513, + 0.004497440531849861, + 0.0037009355146437883, + 0.0017539911204949021, + 0.003371257334947586, + 0.0018686390249058604, + 0.002025244291871786, + 0.0036127541679888964, + 0.001982045127078891, + 0.002184569137170911, + 0.0025153968017548323, + 0.006006214767694473, + 0.006436378229409456, + 0.006645761895924807, + 0.0055531165562570095, + 0.009207148104906082, + 0.005257701501250267, + 0.007784462533891201, + 0.008332389406859875, + 0.03774638473987579, + 0.023979028686881065, + 0.24112018942832947, + 0.03432219475507736, + 0.015591224655508995, + 0.022716054692864418, + 0.028163470327854156, + 0.18986929953098297, + 0.0035417056642472744, + 0.0015207974938675761, + 0.0013906271196901798, + 0.001856358372606337, + 0.0018272126326337457, + 0.0013552213786169887, + 0.0025176217313855886, + 0.002286478178575635, + 0.0012942769099026918, + 0.0021044625900685787, + 0.001748790848068893, + 0.0016011070692911744, + 0.0014540001284331083, + 0.0012391441268846393, + 0.0019685979932546616, + 0.0018708299612626433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.018299441784620285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608913205564022 + ], + "opposite_logprob_tokens": [ + -4.141088962554932 + ], + "logprob_reward": -0.01608913205564022, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608913205564022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05533721670508385, + 0.07512769103050232, + 0.042428743094205856, + 0.35158610343933105, + 0.02110622636973858, + 0.01846533641219139, + 0.04538704454898834, + 0.35396304726600647, + 0.01613476127386093, + 0.020463844761252403 + ], + "credit_map": [ + 0.003987904638051987, + 0.006289628334343433, + 0.005662606097757816, + 0.004203273914754391, + 0.01620057225227356, + 0.005177034065127373, + 0.009209935553371906, + 0.004606261849403381, + 0.008503776043653488, + 0.017607228830456734, + 0.008459885604679585, + 0.010992663912475109, + 0.00893181748688221, + 0.005122124217450619, + 0.007820702157914639, + 0.007689496036618948, + 0.0032554310746490955, + 0.007510330993682146, + 0.004549357108771801, + 0.00427083857357502, + 0.00646252790465951, + 0.003131320932880044, + 0.006623635534197092, + 0.006625301670283079, + 0.02751540206372738, + 0.023804040625691414, + 0.08962266147136688, + 0.04487071558833122, + 0.024871712550520897, + 0.023856069892644882, + 0.0670851394534111, + 0.04996038228273392, + 0.002435150556266308, + 0.0018219256307929754, + 0.0020082828123122454, + 0.0018175676232203841, + 0.0018840320408344269, + 0.0017338501056656241, + 0.0037452997639775276, + 0.005660117603838444, + 0.0014446048298850656, + 0.00280238245613873, + 0.002142030280083418, + 0.002099149627611041, + 0.0028392854146659374, + 0.002264057518914342, + 0.002305858302861452, + 0.0025679669342935085, + 0.003796023316681385, + 0.00552832055836916, + 0.004585444927215576, + 0.005726031959056854, + 0.007490319665521383, + 0.005296314135193825, + 0.005380683578550816, + 0.007583905477076769, + 0.02590571716427803, + 0.013035207986831665, + 0.149560809135437, + 0.02014043554663658, + 0.010238219983875751, + 0.011499418877065182, + 0.014744522050023079, + 0.10883874446153641, + 0.001923281466588378, + 0.0018395247170701623, + 0.0014842927921563387, + 0.0018788970774039626, + 0.0023843988310545683, + 0.0013706726022064686, + 0.0026907206047326326, + 0.0025629724841564894, + 0.0024912087246775627, + 0.0026345120277255774, + 0.0022300872951745987, + 0.0028284445870667696, + 0.003192067611962557, + 0.0021657771430909634, + 0.002467402955517173, + 0.0024543441832065582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.01608913205564022 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029792040586472 + ], + "opposite_logprob_tokens": [ + -2.3502979278564453 + ], + "logprob_reward": -0.10029792040586472, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029792040586472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052236851304769516, + 0.02327083982527256, + 0.05132313445210457, + 0.15090543031692505, + 0.03156832978129387, + 0.02778550796210766, + 0.07620223611593246, + 0.5490625500679016, + 0.01789555326104164, + 0.019749557599425316 + ], + "credit_map": [ + 0.004305544774979353, + 0.006395838223397732, + 0.005803040228784084, + 0.004704883787781, + 0.012212976813316345, + 0.005689718294888735, + 0.006257583387196064, + 0.006867268588393927, + 0.00208674231544137, + 0.002922485815361142, + 0.00214947946369648, + 0.0024406088050454855, + 0.0046836091205477715, + 0.0024745122063905, + 0.002611174713820219, + 0.0039022271521389484, + 0.00354013848118484, + 0.01115301251411438, + 0.005512523930519819, + 0.005302445497363806, + 0.006284973584115505, + 0.005327249877154827, + 0.007910473272204399, + 0.006292319856584072, + 0.01821332797408104, + 0.019242655485868454, + 0.011533339507877827, + 0.022523684427142143, + 0.015880540013313293, + 0.013057918287813663, + 0.026088187471032143, + 0.024365784600377083, + 0.0032847982365638018, + 0.003922570962458849, + 0.0031009602826088667, + 0.0029036691412329674, + 0.006330918520689011, + 0.0034362440928816795, + 0.0035931896418333054, + 0.004995977971702814, + 0.002647974994033575, + 0.002506494289264083, + 0.0028414379339665174, + 0.00239368062466383, + 0.0035095647908747196, + 0.0019964263774454594, + 0.008565657772123814, + 0.0033242690842598677, + 0.010506760329008102, + 0.008624961599707603, + 0.014089323580265045, + 0.01113145425915718, + 0.008530797436833382, + 0.007292599882930517, + 0.0068521443754434586, + 0.009174197912216187, + 0.04063672944903374, + 0.019706599414348602, + 0.23483708500862122, + 0.033366858959198, + 0.01136537455022335, + 0.018042244017124176, + 0.020016955211758614, + 0.1710907369852066, + 0.003521594451740384, + 0.0022816061973571777, + 0.0014326798263937235, + 0.002116067335009575, + 0.0023863837122917175, + 0.001427310868166387, + 0.002487344201654196, + 0.0022425672505050898, + 0.0030614195857197046, + 0.0023368196561932564, + 0.002669824520125985, + 0.0022992314770817757, + 0.0035183716099709272, + 0.001366351032629609, + 0.0024482342414557934, + 0.0020493040792644024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.10029792040586472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624690420925617 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624690420925617, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624690420925617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03862706199288368, + 0.029982078820466995, + 0.03222740814089775, + 0.3760579228401184, + 0.01832459308207035, + 0.01943015865981579, + 0.020820310339331627, + 0.436504989862442, + 0.014635911211371422, + 0.013389508239924908 + ], + "credit_map": [ + 0.0033354961778968573, + 0.00495610386133194, + 0.004446864128112793, + 0.0042692311108112335, + 0.008899078704416752, + 0.004357881844043732, + 0.004566561430692673, + 0.003795847063884139, + 0.002249059733003378, + 0.0039275409653782845, + 0.0030003853607922792, + 0.003009531646966934, + 0.007242040242999792, + 0.0028475637082010508, + 0.0036160657182335854, + 0.0040898919105529785, + 0.002454048953950405, + 0.004640428349375725, + 0.0022144580725580454, + 0.0037769556511193514, + 0.008692242205142975, + 0.0025095276068896055, + 0.0034542677458375692, + 0.0044854795560240746, + 0.029261106625199318, + 0.021860139444470406, + 0.10895171016454697, + 0.03578526899218559, + 0.022314537316560745, + 0.01897544413805008, + 0.08854766935110092, + 0.05036206915974617, + 0.0016396469436585903, + 0.002362787490710616, + 0.0012950921664014459, + 0.0015189307741820812, + 0.003530662739649415, + 0.0015533360419794917, + 0.0035120754037052393, + 0.0029120631515979767, + 0.00187083357013762, + 0.002865624614059925, + 0.0019843881018459797, + 0.0016464374493807554, + 0.004057764541357756, + 0.001992719015106559, + 0.0028537786565721035, + 0.002158614806830883, + 0.0019521177746355534, + 0.002955992938950658, + 0.002228215103968978, + 0.002713648369535804, + 0.0033601776231080294, + 0.0026149593759328127, + 0.002516978420317173, + 0.0024782230611890554, + 0.031606145203113556, + 0.01613924466073513, + 0.18746605515480042, + 0.026120075955986977, + 0.008083569817245007, + 0.014557690359652042, + 0.015743296593427658, + 0.13678893446922302, + 0.0028514056466519833, + 0.0016078075859695673, + 0.0018567874794825912, + 0.0015260451473295689, + 0.001311799162067473, + 0.0013220796827226877, + 0.0020262617617845535, + 0.0021337252110242844, + 0.0016843327321112156, + 0.001911481493152678, + 0.002067860681563616, + 0.0014801722718402743, + 0.0019996128976345062, + 0.0009608714026398957, + 0.0019851047545671463, + 0.0013000735780224204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.007624690420925617 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634849451482296 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634849451482296, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634849451482296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11026369035243988, + 0.06913834065198898, + 0.04115471616387367, + 0.16186660528182983, + 0.11561024188995361, + 0.18329016864299774, + 0.10301228612661362, + 0.13556557893753052, + 0.04436876252293587, + 0.03572968766093254 + ], + "credit_map": [ + 0.011891365982592106, + 0.011643165722489357, + 0.014447811990976334, + 0.014835754409432411, + 0.01716259866952896, + 0.011074925772845745, + 0.012602263130247593, + 0.01660579815506935, + 0.007393101695924997, + 0.016117025166749954, + 0.004620284773409367, + 0.009333536960184574, + 0.010556904599070549, + 0.006009313743561506, + 0.006749389227479696, + 0.008358776569366455, + 0.004066199064254761, + 0.005573695059865713, + 0.003810511203482747, + 0.00561210373416543, + 0.006150231696665287, + 0.004512414801865816, + 0.00591388251632452, + 0.005515672732144594, + 0.019810719415545464, + 0.02254546247422695, + 0.01603674329817295, + 0.02845814824104309, + 0.019880633801221848, + 0.014548738487064838, + 0.017432156950235367, + 0.023153988644480705, + 0.012530890293419361, + 0.01317543350160122, + 0.012504921294748783, + 0.009807806462049484, + 0.016213122755289078, + 0.012202708050608635, + 0.023006482049822807, + 0.01616887003183365, + 0.017020149156451225, + 0.019502539187669754, + 0.017271623015403748, + 0.015951819717884064, + 0.017321892082691193, + 0.015305754728615284, + 0.0347786545753479, + 0.04613771289587021, + 0.012821388430893421, + 0.012764335609972477, + 0.018140759319067, + 0.016325442120432854, + 0.011221558786928654, + 0.010480526834726334, + 0.010586829856038094, + 0.010671432130038738, + 0.011468240059912205, + 0.009772926568984985, + 0.04303964599967003, + 0.014623358845710754, + 0.008246166631579399, + 0.005837541073560715, + 0.011593339033424854, + 0.030984332785010338, + 0.009094920009374619, + 0.0051386416889727116, + 0.005723214708268642, + 0.004433035850524902, + 0.0054998937994241714, + 0.002890929812565446, + 0.005636304616928101, + 0.005951814819127321, + 0.004534805193543434, + 0.0036269198171794415, + 0.003123511327430606, + 0.005736051592975855, + 0.005943249445408583, + 0.0032912432216107845, + 0.00554911233484745, + 0.00392478983849287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.008634849451482296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot from a horizontal angle, depicts a serene, gender-ambiguous angel bathed in dreamy light. A pair of vast, pristine white wings unfurl behind them, each feather clearly visible and edged with a soft silver halo. A cascade of silver-white hair gently falls over their shoulders. The angel, clad in a white robe of translucent silk, holds a gaze filled with compassion and warmth. Slowly and firmly, they extend their right hand toward the camera. A warm, luminous golden glow emanates from their palm and fingertips, like a candlelight shrouded in mist. This light gently illuminates the surrounding space and casts dynamic shadows on the divine being's face and robes. The entire frame is in soft focus, with the background blurred, leaving only a faint, abstract bokeh of golden and pale blue specks. This inviting and soothing gesture, combined with the warm tones and tranquil atmosphere, creates a sacred, healing, and otherworldly visual experience.", + "seed": 5352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_1/index_911_seed_5352_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": -0.016697357315570115, + "vlm_logprob": -0.016697357315570115, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.016697357315570115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041279386729002 + ], + "opposite_logprob_tokens": [ + -3.2912793159484863 + ], + "logprob_reward": -0.041279386729002, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.041279386729002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10906019806861877, + 0.0726499930024147, + 0.09533219784498215, + 0.076999232172966, + 0.059059981256723404, + 0.0785682201385498, + 0.04416636377573013, + 0.03179752826690674, + 0.10214431583881378, + 0.3302219808101654 + ], + "credit_map": [ + 0.00956901628524065, + 0.014369660057127476, + 0.009948431514203548, + 0.01695670187473297, + 0.016480309888720512, + 0.012081078253686428, + 0.018598586320877075, + 0.01105640921741724, + 0.010502584278583527, + 0.009549137204885483, + 0.013016587123274803, + 0.01120018307119608, + 0.006790024694055319, + 0.0052540116012096405, + 0.009531731717288494, + 0.006805730052292347, + 0.005179175175726414, + 0.024769138544797897, + 0.005292849149554968, + 0.02626754157245159, + 0.012991073541343212, + 0.005381786730140448, + 0.008396401070058346, + 0.007054230198264122, + 0.006110329180955887, + 0.009534671902656555, + 0.009657700546085835, + 0.014000799506902695, + 0.012131066992878914, + 0.005342712160199881, + 0.011003033258020878, + 0.009218916296958923, + 0.005353282671421766, + 0.011493516154587269, + 0.011116334237158298, + 0.009190335869789124, + 0.006894765887409449, + 0.003292144276201725, + 0.006156584247946739, + 0.005563014652580023, + 0.006281489506363869, + 0.02024269290268421, + 0.02481752261519432, + 0.009060054086148739, + 0.005726249888539314, + 0.004903252236545086, + 0.0032356781885027885, + 0.0043012844398617744, + 0.005964763928204775, + 0.006173902191221714, + 0.0077444687485694885, + 0.007688331417739391, + 0.004497944377362728, + 0.0025988523848354816, + 0.005048797000199556, + 0.004449302796274424, + 0.0042702811770141125, + 0.00415249215438962, + 0.003445038106292486, + 0.007309808395802975, + 0.0033091064542531967, + 0.001751376548781991, + 0.004512251820415258, + 0.0030471719801425934, + 0.02040984109044075, + 0.01998097077012062, + 0.007667933590710163, + 0.012893731705844402, + 0.011319457553327084, + 0.008126222528517246, + 0.010842667892575264, + 0.010903490707278252, + 0.12526826560497284, + 0.12000315636396408, + 0.016212375834584236, + 0.014924686402082443, + 0.021158920601010323, + 0.011537916027009487, + 0.0117555046454072, + 0.009361188858747482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.041279386729002 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982400916516781 + ], + "opposite_logprob_tokens": [ + -3.5298240184783936 + ], + "logprob_reward": -0.02982400916516781, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982400916516781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12464382499456406, + 0.058110311627388, + 0.14185790717601776, + 0.033944305032491684, + 0.02690269611775875, + 0.20113521814346313, + 0.016827352344989777, + 0.024837471544742584, + 0.03307749330997467, + 0.3386634290218353 + ], + "credit_map": [ + 0.016172096133232117, + 0.020799128338694572, + 0.015920301899313927, + 0.019265251234173775, + 0.012717456556856632, + 0.009215911850333214, + 0.019208397716283798, + 0.011345279403030872, + 0.007012900896370411, + 0.009447935037314892, + 0.0077971587888896465, + 0.009121043607592583, + 0.005667849909514189, + 0.0038250619545578957, + 0.007629294414073229, + 0.0076090688817203045, + 0.008127938956022263, + 0.0372280515730381, + 0.0063020093366503716, + 0.038469526916742325, + 0.01921762526035309, + 0.007849127054214478, + 0.01248048897832632, + 0.012183132581412792, + 0.002652114490047097, + 0.0046752216294407845, + 0.0033463644795119762, + 0.006174345966428518, + 0.005689042620360851, + 0.0029540574178099632, + 0.004871733952313662, + 0.0035814254079014063, + 0.0020080399699509144, + 0.004860801622271538, + 0.004503708798438311, + 0.004365272354334593, + 0.0026615848764777184, + 0.0020851856097579002, + 0.003380696289241314, + 0.0030374079942703247, + 0.016163013875484467, + 0.05572809278964996, + 0.062456924468278885, + 0.024833673611283302, + 0.01004671212285757, + 0.010701312683522701, + 0.008377950638532639, + 0.012827528640627861, + 0.0014167913468554616, + 0.0018853519577533007, + 0.0019024715293198824, + 0.0019308528862893581, + 0.003299061208963394, + 0.0017432412132620811, + 0.0023938757367432117, + 0.002255707047879696, + 0.003108615754172206, + 0.002643310697749257, + 0.00302138808183372, + 0.004780132789164782, + 0.0035822009667754173, + 0.001844708458520472, + 0.0029039958026260138, + 0.002953121205791831, + 0.00504524540156126, + 0.00577873969450593, + 0.0027077114209532738, + 0.006374109070748091, + 0.0033207384403795004, + 0.0027336017228662968, + 0.004046136047691107, + 0.0030712103471159935, + 0.13259705901145935, + 0.12395691871643066, + 0.014492953196167946, + 0.014078876003623009, + 0.01960880681872368, + 0.011693624779582024, + 0.011150420643389225, + 0.011084778234362602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.02982400916516781 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525640370324254 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.00525640370324254, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525640370324254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024701492860913277, + 0.016128288581967354, + 0.014252079650759697, + 0.011972853913903236, + 0.03790506720542908, + 0.033122044056653976, + 0.01142085250467062, + 0.02710181288421154, + 0.03903261199593544, + 0.7843629121780396 + ], + "credit_map": [ + 0.00210797181352973, + 0.0034805666655302048, + 0.002793302061036229, + 0.003702363232150674, + 0.0026495058555155993, + 0.003384787356480956, + 0.0037020938470959663, + 0.0028809034265577793, + 0.002208577934652567, + 0.0023554167710244656, + 0.0023857050109654665, + 0.0024752363096922636, + 0.0012333958875387907, + 0.0011176374973729253, + 0.0025390824303030968, + 0.001813238370232284, + 0.001532415859401226, + 0.002841686364263296, + 0.0014657004503533244, + 0.0021528275683522224, + 0.0020387289114296436, + 0.0011334777809679508, + 0.0014256469439715147, + 0.001661596354097128, + 0.0011449455050751567, + 0.002646897453814745, + 0.0014461971586570144, + 0.0030445021111518145, + 0.0012844394659623504, + 0.0007956023328006268, + 0.0009332912159152329, + 0.0006769797764718533, + 0.0033338069915771484, + 0.011763457208871841, + 0.01079869456589222, + 0.004960259888321161, + 0.002239241963252425, + 0.00153943186160177, + 0.0014272400876507163, + 0.001842940109781921, + 0.002819749526679516, + 0.00740509107708931, + 0.008925225585699081, + 0.004606715403497219, + 0.0022964829113334417, + 0.0021619333419948816, + 0.002328661736100912, + 0.002578187035396695, + 0.001615546876564622, + 0.0015313902404159307, + 0.0019390969537198544, + 0.0022067511454224586, + 0.0009503819746896625, + 0.0009237159392796457, + 0.0011426741257309914, + 0.0011112961219623685, + 0.003757348284125328, + 0.003988180309534073, + 0.0031231516040861607, + 0.005729347001761198, + 0.0030626843217760324, + 0.0015269857831299305, + 0.003497566794976592, + 0.0024165520444512367, + 0.005757723934948444, + 0.006425406783819199, + 0.0043565575033426285, + 0.008294138126075268, + 0.004059275612235069, + 0.0026079847011715174, + 0.004237367305904627, + 0.003294161055237055, + 0.3183557987213135, + 0.2981576919555664, + 0.02864622138440609, + 0.03961990401148796, + 0.0374426431953907, + 0.02448776550590992, + 0.01990911178290844, + 0.017743844538927078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.00525640370324254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011793570592999458 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.0011793570592999458, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011793570592999458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05904841795563698, + 0.02830519899725914, + 0.08026450872421265, + 0.09318304061889648, + 0.2706819772720337, + 0.09371387958526611, + 0.03152352571487427, + 0.11714866757392883, + 0.08353200554847717, + 0.14259877800941467 + ], + "credit_map": [ + 0.01113204937428236, + 0.006838381290435791, + 0.010117931291460991, + 0.008531332015991211, + 0.004970068112015724, + 0.006547985598444939, + 0.004851496312767267, + 0.0060591744258999825, + 0.0041966913267970085, + 0.004189787432551384, + 0.004174182657152414, + 0.004202463664114475, + 0.002709722612053156, + 0.002395049436017871, + 0.003060760674998164, + 0.0033765400294214487, + 0.004109248984605074, + 0.02949872799217701, + 0.003192370757460594, + 0.029254846274852753, + 0.003393066581338644, + 0.003793229814618826, + 0.0035714716650545597, + 0.0034515466541051865, + 0.006151835434138775, + 0.023988615721464157, + 0.007734075654298067, + 0.02381308563053608, + 0.010083808563649654, + 0.006303232628852129, + 0.008901619352400303, + 0.0062067690305411816, + 0.0239117294549942, + 0.08392228186130524, + 0.08219847083091736, + 0.03395695239305496, + 0.013915074989199638, + 0.01175936684012413, + 0.008676821365952492, + 0.01234125904738903, + 0.007695832289755344, + 0.02463800273835659, + 0.027763549238443375, + 0.01401292160153389, + 0.004730246961116791, + 0.005026589147746563, + 0.005146335810422897, + 0.00470039714127779, + 0.0038870503194630146, + 0.003419626271352172, + 0.0053466190584003925, + 0.00443490082398057, + 0.0035026180557906628, + 0.0027627863455563784, + 0.0045960405841469765, + 0.0035738847218453884, + 0.01693025417625904, + 0.015563009306788445, + 0.012093141674995422, + 0.029159776866436005, + 0.013915840536355972, + 0.006254512816667557, + 0.012555843219161034, + 0.010676287114620209, + 0.012398866005241871, + 0.013409130275249481, + 0.005909385159611702, + 0.02636094205081463, + 0.007138408254832029, + 0.004407033324241638, + 0.008268223144114017, + 0.005640014540404081, + 0.04187370091676712, + 0.03883863613009453, + 0.008461312390863895, + 0.021229643374681473, + 0.009926289319992065, + 0.006970181129872799, + 0.009628626517951488, + 0.005670397542417049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.0011793570592999458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gender-ambiguous angel with vast, pristine white wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the angel wearing a white robe of translucent silk with silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the angel slowly and firmly extend their right hand toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a warm, luminous golden glow emanate from the angel's palm and fingertips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot with a soft focus and blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947629921138287 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947629921138287, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947629921138287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0240450631827116, + 0.02204122394323349, + 0.0631011351943016, + 0.11996138840913773, + 0.13189581036567688, + 0.01041566114872694, + 0.008241720497608185, + 0.01943969540297985, + 0.05649213865399361, + 0.544366180896759 + ], + "credit_map": [ + 0.001973699312657118, + 0.003716219449415803, + 0.0021826918236911297, + 0.0034564726520329714, + 0.003291873261332512, + 0.0021426365710794926, + 0.004537480883300304, + 0.0027439873665571213, + 0.0020849306602030993, + 0.002721872879192233, + 0.0029181591235101223, + 0.0027510656509548426, + 0.0021655773743987083, + 0.0015658382326364517, + 0.004310875199735165, + 0.0035229010973125696, + 0.0036926802713423967, + 0.02110178954899311, + 0.003567791311070323, + 0.02138659358024597, + 0.003398402128368616, + 0.003175541991367936, + 0.003515796735882759, + 0.0032625305466353893, + 0.007443296257406473, + 0.02769516408443451, + 0.010239393450319767, + 0.029483554884791374, + 0.014036817476153374, + 0.008512144908308983, + 0.013378288596868515, + 0.0091727115213871, + 0.011494925245642662, + 0.03178422525525093, + 0.031084444373846054, + 0.018604284152388573, + 0.013885977678000927, + 0.006670756731182337, + 0.008778649382293224, + 0.009592526592314243, + 0.0008220337913371623, + 0.002104756888002157, + 0.0027821678668260574, + 0.0014506644802168012, + 0.0009474061080254614, + 0.0006255782209336758, + 0.0007701880531385541, + 0.0009128644596785307, + 0.0010484905214980245, + 0.0010206842562183738, + 0.0010629688622429967, + 0.0017302111955359578, + 0.0010703136213123798, + 0.000571001844946295, + 0.001074108644388616, + 0.0006639403873123229, + 0.002379824174568057, + 0.002390605863183737, + 0.002038379432633519, + 0.005720451474189758, + 0.001893006730824709, + 0.001188150024972856, + 0.0022619457449764013, + 0.0015673304442316294, + 0.013149874284863472, + 0.01277208048850298, + 0.005720625631511211, + 0.004469274077564478, + 0.006023366004228592, + 0.004159687086939812, + 0.004343379754573107, + 0.005853850860148668, + 0.2185278981924057, + 0.20416082441806793, + 0.020895633846521378, + 0.015163874253630638, + 0.030605841428041458, + 0.022485695779323578, + 0.01766110397875309, + 0.014865308068692684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.005947629921138287 + } + ], + "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/0000125/group_10/index_tensor([4264, 4264])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_10/index_tensor([4264, 4264])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa7edef3fb5d5cd8350883c2fc3e7ac47abcf57e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_10/index_tensor([4264, 4264])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme wide-angle lens, looking upward from the ground, tilts the camera upward in a slow, majestic rhythm, gradually revealing a cosmic spectacle. Against the velvety, deep indigo sky, countless fiery meteors intertwine densely, forming a veritable waterfall of light. These meteors do not streak haphazardly, but rather cascade vertically and unstoppably in a grand, unified rhythm. Each meteor trails a long, shimmering trail of blazing white, icy blue, and fleeting gold. These trails briefly illuminate the faint, distant stars and nebulous dust in the background as they fall. This luminous torrent, the sole source of light for the entire scene, casts dynamic, flowing shadows upon the faintly visible, rolling mountain silhouettes at the bottom of the frame, creating a breathtaking, almost sacred atmosphere of grandeur, imbuing the image with a sense of cosmic tranquility and awe.", + "seed": 24788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_10/index_4264_seed_24788_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": -0.07826528269797564, + "vlm_logprob": -0.07826528269797564, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.07826528269797564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015044321306049824 + ], + "opposite_logprob_tokens": [ + -5.265044212341309 + ], + "logprob_reward": -0.015044321306049824, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.015044321306049824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11303803324699402, + 0.09877122938632965, + 0.11609134823083878, + 0.06537958979606628, + 0.09236662089824677, + 0.0425877682864666, + 0.05504777282476425, + 0.15754693746566772, + 0.2177286446094513, + 0.04144205152988434 + ], + "credit_map": [ + 0.012337399646639824, + 0.013630845583975315, + 0.01648748852312565, + 0.015986915677785873, + 0.013976610265672207, + 0.012589454650878906, + 0.013026890344917774, + 0.015002429485321045, + 0.008905143477022648, + 0.011233573779463768, + 0.006214513909071684, + 0.009588695131242275, + 0.02370697259902954, + 0.020668046548962593, + 0.009377689100801945, + 0.009076596237719059, + 0.014973622746765614, + 0.01043094601482153, + 0.00893489085137844, + 0.017752325162291527, + 0.024867771193385124, + 0.022506391629576683, + 0.007913290522992611, + 0.008712111040949821, + 0.00651786383241415, + 0.006920596584677696, + 0.0107127008959651, + 0.010833615437150002, + 0.00897188950330019, + 0.008101468905806541, + 0.007189014460891485, + 0.006132436450570822, + 0.01086189690977335, + 0.01233814936131239, + 0.0070875720120966434, + 0.006822647992521524, + 0.017285365611314774, + 0.013278735801577568, + 0.010720565915107727, + 0.01397168543189764, + 0.004170069471001625, + 0.008918465115129948, + 0.006345676723867655, + 0.003075302578508854, + 0.006436025258153677, + 0.0055418554693460464, + 0.004059274215251207, + 0.00404109712690115, + 0.0073434156365692616, + 0.006623715627938509, + 0.009888763539493084, + 0.006828846875578165, + 0.006823081523180008, + 0.0074437265284359455, + 0.005398374982178211, + 0.004697846248745918, + 0.018172333016991615, + 0.010673899203538895, + 0.02516675740480423, + 0.023632798343896866, + 0.02479744330048561, + 0.026978492736816406, + 0.013387516140937805, + 0.01473770011216402, + 0.009180955588817596, + 0.009260427206754684, + 0.016943734139204025, + 0.012081613764166832, + 0.06355907768011093, + 0.07395225018262863, + 0.019319899380207062, + 0.013430685736238956, + 0.0035568391904234886, + 0.004034791607409716, + 0.007524012587964535, + 0.005680784583091736, + 0.005808061454445124, + 0.005253978073596954, + 0.005744989961385727, + 0.003838593838736415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.015044321306049824 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284634590148926 + ], + "opposite_logprob_tokens": [ + -2.0178463459014893 + ], + "logprob_reward": -0.14284634590148926, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284634590148926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06134903430938721, + 0.0741647332906723, + 0.15219175815582275, + 0.10778289288282394, + 0.15806470811367035, + 0.19770905375480652, + 0.07382870465517044, + 0.034742292016744614, + 0.047188900411129, + 0.09297793358564377 + ], + "credit_map": [ + 0.00500199431553483, + 0.008487275801599026, + 0.009565779939293861, + 0.007053579669445753, + 0.008595999330282211, + 0.008966374211013317, + 0.005719561595469713, + 0.007958467118442059, + 0.006807978730648756, + 0.008484411053359509, + 0.006996964570134878, + 0.011180024594068527, + 0.014180894941091537, + 0.011708704754710197, + 0.00846429169178009, + 0.006341464351862669, + 0.020082710310816765, + 0.011178714223206043, + 0.014494319446384907, + 0.020899977535009384, + 0.03455482795834541, + 0.03405047580599785, + 0.007579568307846785, + 0.009351147338747978, + 0.010793416760861874, + 0.01139132771641016, + 0.01696048676967621, + 0.015422262251377106, + 0.016316069290041924, + 0.0144749591127038, + 0.012340103276073933, + 0.010084269568324089, + 0.0236971378326416, + 0.02167677879333496, + 0.013984833844006062, + 0.011592881754040718, + 0.02658986486494541, + 0.02130337432026863, + 0.016744747757911682, + 0.022475097328424454, + 0.022531477734446526, + 0.031612787395715714, + 0.028261592611670494, + 0.024289892986416817, + 0.025245770812034607, + 0.025190310552716255, + 0.01609111577272415, + 0.024486109614372253, + 0.008108153007924557, + 0.010784799233078957, + 0.012454052455723286, + 0.009319684468209743, + 0.00995421502739191, + 0.009970963932573795, + 0.006827645469456911, + 0.006409191060811281, + 0.00277133472263813, + 0.0018474995158612728, + 0.004437573254108429, + 0.0032858310732990503, + 0.009236744605004787, + 0.008859356865286827, + 0.0021620739717036486, + 0.0021418766118586063, + 0.0014840412186458707, + 0.0014306423254311085, + 0.0014340410707518458, + 0.0014114034129306674, + 0.018055183812975883, + 0.019278578460216522, + 0.0018683384405449033, + 0.002226673299446702, + 0.008568410761654377, + 0.012553418055176735, + 0.013618347235023975, + 0.009083844721317291, + 0.015270421281456947, + 0.012037080712616444, + 0.014575967565178871, + 0.00727044278755784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.14284634590148926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283372461795807 + ], + "opposite_logprob_tokens": [ + -2.017833709716797 + ], + "logprob_reward": -0.14283372461795807, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283372461795807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09224013239145279, + 0.06919986009597778, + 0.11480184644460678, + 0.13704082369804382, + 0.15104985237121582, + 0.15813691914081573, + 0.11518379300832748, + 0.054607830941677094, + 0.0449119433760643, + 0.06282699853181839 + ], + "credit_map": [ + 0.012696145102381706, + 0.016947759315371513, + 0.0076531535014510155, + 0.014663969166576862, + 0.011635659262537956, + 0.01435880083590746, + 0.006717316340655088, + 0.007567322812974453, + 0.006261195056140423, + 0.009068048559129238, + 0.0064296359196305275, + 0.0107683464884758, + 0.011373454704880714, + 0.010907712392508984, + 0.0076581877656280994, + 0.006733281537890434, + 0.019554417580366135, + 0.009259853512048721, + 0.01402554102241993, + 0.00920138880610466, + 0.020965509116649628, + 0.020902110263705254, + 0.010397729463875294, + 0.010495293885469437, + 0.010656469501554966, + 0.01093264389783144, + 0.020078541710972786, + 0.0131028788164258, + 0.02456502430140972, + 0.019952166825532913, + 0.01918705180287361, + 0.018566060811281204, + 0.018733233213424683, + 0.02346811816096306, + 0.012079448439180851, + 0.011590219102799892, + 0.03136397898197174, + 0.02227228507399559, + 0.01658543013036251, + 0.014957132749259472, + 0.01860794797539711, + 0.026460539549589157, + 0.021872393786907196, + 0.01811831444501877, + 0.02097686380147934, + 0.020619332790374756, + 0.012513306923210621, + 0.01896822825074196, + 0.012829428538680077, + 0.009147095493972301, + 0.01651819422841072, + 0.014623641967773438, + 0.016231879591941833, + 0.015733962878584862, + 0.01612081378698349, + 0.013978774659335613, + 0.005804616957902908, + 0.003799624741077423, + 0.0057509951293468475, + 0.006675551179796457, + 0.010240456089377403, + 0.009971980936825275, + 0.005868210457265377, + 0.006496392656117678, + 0.003820358542725444, + 0.0034638340584933758, + 0.004419487901031971, + 0.004456018563359976, + 0.007606294006109238, + 0.00890528317540884, + 0.007133273407816887, + 0.005107394885271788, + 0.004973693285137415, + 0.006152656860649586, + 0.007141696289181709, + 0.007167024537920952, + 0.007687292993068695, + 0.008448726497590542, + 0.013196119107306004, + 0.008059792220592499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.14283372461795807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594101399183273 + ], + "opposite_logprob_tokens": [ + -3.8955941200256348 + ], + "logprob_reward": -0.020594101399183273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594101399183273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09875055402517319, + 0.08811963349580765, + 0.11650744825601578, + 0.058814968913793564, + 0.06925003230571747, + 0.1667870730161667, + 0.12861758470535278, + 0.050544623285532, + 0.16211362183094025, + 0.0604945532977581 + ], + "credit_map": [ + 0.012717739678919315, + 0.01879454031586647, + 0.010773543268442154, + 0.01579396426677704, + 0.012535455636680126, + 0.013756736181676388, + 0.006857429165393114, + 0.007521137595176697, + 0.006956501863896847, + 0.009396745823323727, + 0.009191102348268032, + 0.009129993617534637, + 0.020135436207056046, + 0.01675451174378395, + 0.00951546709984541, + 0.007039868272840977, + 0.00714983930811286, + 0.007682307623326778, + 0.005901428870856762, + 0.012977703474462032, + 0.035250723361968994, + 0.034431811422109604, + 0.007328988518565893, + 0.005784626584500074, + 0.006394567899405956, + 0.00677164550870657, + 0.008337249979376793, + 0.008981628343462944, + 0.008365136571228504, + 0.007747167721390724, + 0.007554356940090656, + 0.0046632057055830956, + 0.010700780898332596, + 0.00833019707351923, + 0.006961882580071688, + 0.006804111413657665, + 0.007762570399791002, + 0.008770973421633244, + 0.006259630434215069, + 0.013659871183335781, + 0.014836156740784645, + 0.025810204446315765, + 0.019241824746131897, + 0.014686712063848972, + 0.02784753032028675, + 0.020971061661839485, + 0.015458734706044197, + 0.027934836223721504, + 0.011516823433339596, + 0.013990207575261593, + 0.020165879279375076, + 0.013510620221495628, + 0.018795333802700043, + 0.017893776297569275, + 0.013379815965890884, + 0.019365103915333748, + 0.003472744720056653, + 0.00293045025318861, + 0.004643633030354977, + 0.007059503812342882, + 0.010383483953773975, + 0.011217101477086544, + 0.006313811056315899, + 0.004523890558630228, + 0.00775787141174078, + 0.007358700502663851, + 0.00725061958655715, + 0.010726459324359894, + 0.047945741564035416, + 0.05098363757133484, + 0.014496339485049248, + 0.015594231896102428, + 0.005205544177442789, + 0.006573771126568317, + 0.008151279762387276, + 0.0064702206291258335, + 0.010988184250891209, + 0.008130875416100025, + 0.00970009621232748, + 0.005274576134979725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.020594101399183273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000792026519775 + ], + "opposite_logprob_tokens": [ + -2.695007801055908 + ], + "logprob_reward": -0.07000792026519775, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000792026519775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28449133038520813, + 0.08155930042266846, + 0.13509324193000793, + 0.07660400867462158, + 0.09854762256145477, + 0.033371977508068085, + 0.02933214232325554, + 0.06632807105779648, + 0.1371585726737976, + 0.05751372128725052 + ], + "credit_map": [ + 0.0207455325871706, + 0.025875480845570564, + 0.04438946396112442, + 0.029398728162050247, + 0.0362510085105896, + 0.03647942095994949, + 0.038666948676109314, + 0.052684735506772995, + 0.006295931991189718, + 0.008602045476436615, + 0.006982704624533653, + 0.0072868820279836655, + 0.01652509532868862, + 0.013457918539643288, + 0.012262989766895771, + 0.010145732201635838, + 0.007578750606626272, + 0.006485413294285536, + 0.004866192117333412, + 0.010013876482844353, + 0.04664633795619011, + 0.046584486961364746, + 0.0075748274102807045, + 0.005343359895050526, + 0.007063411641865969, + 0.006118617486208677, + 0.008474951609969139, + 0.008866479620337486, + 0.01553569920361042, + 0.011668271385133266, + 0.009983453899621964, + 0.008893119171261787, + 0.015578098595142365, + 0.01386730745434761, + 0.009193710051476955, + 0.0075288331136107445, + 0.017301853746175766, + 0.013338853605091572, + 0.00967144500464201, + 0.012067526578903198, + 0.003104847390204668, + 0.005424192175269127, + 0.004954732954502106, + 0.002811259590089321, + 0.006079556420445442, + 0.004121323581784964, + 0.0033059276174753904, + 0.0035701387096196413, + 0.00345471384935081, + 0.003834296250715852, + 0.005796645302325487, + 0.004016762599349022, + 0.003519524121657014, + 0.0037602242082357407, + 0.0023825757671147585, + 0.002567399526014924, + 0.004194850567728281, + 0.00302825472317636, + 0.006970779504626989, + 0.006903310306370258, + 0.017950285226106644, + 0.017029160633683205, + 0.005180662963539362, + 0.005070765037089586, + 0.0027438148390501738, + 0.003827768610790372, + 0.002964598825201392, + 0.003499462967738509, + 0.054766640067100525, + 0.05847436562180519, + 0.005926643498241901, + 0.004955273121595383, + 0.005044303368777037, + 0.006558674853295088, + 0.007389756850898266, + 0.005129282828420401, + 0.010654236190021038, + 0.008345712907612324, + 0.009236713871359825, + 0.005155038554221392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.07000792026519775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme wide-angle lens, looking upward from the ground, tilts the camera upward in a slow, majestic rhythm, gradually revealing a cosmic spectacle. Against the velvety, deep indigo sky, countless fiery meteors intertwine densely, forming a veritable waterfall of light. These meteors do not streak haphazardly, but rather cascade vertically and unstoppably in a grand, unified rhythm. Each meteor trails a long, shimmering trail of blazing white, icy blue, and fleeting gold. These trails briefly illuminate the faint, distant stars and nebulous dust in the background as they fall. This luminous torrent, the sole source of light for the entire scene, casts dynamic, flowing shadows upon the faintly visible, rolling mountain silhouettes at the bottom of the frame, creating a breathtaking, almost sacred atmosphere of grandeur, imbuing the image with a sense of cosmic tranquility and awe.", + "seed": 24789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_10/index_4264_seed_24789_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": -0.05293195219710469, + "vlm_logprob": -0.05293195219710469, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05293195219710469, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021180162206292152 + ], + "opposite_logprob_tokens": [ + -4.396180152893066 + ], + "logprob_reward": -0.021180162206292152, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.021180162206292152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29258033633232117, + 0.049474913626909256, + 0.07166770100593567, + 0.02245619148015976, + 0.0749998688697815, + 0.054151661694049835, + 0.14086055755615234, + 0.04089479148387909, + 0.04631326347589493, + 0.20660072565078735 + ], + "credit_map": [ + 0.05460198596119881, + 0.015372156165540218, + 0.06636709719896317, + 0.017590157687664032, + 0.019984355196356773, + 0.09844478964805603, + 0.009203048422932625, + 0.011016743257641792, + 0.006427527405321598, + 0.009044716134667397, + 0.0052611385472118855, + 0.007094886619597673, + 0.005720758810639381, + 0.0073915040120482445, + 0.005746189039200544, + 0.0027881928253918886, + 0.007772604934871197, + 0.011948027648031712, + 0.006026287097483873, + 0.00890736747533083, + 0.011402688920497894, + 0.009624354541301727, + 0.009700674563646317, + 0.006285694427788258, + 0.0035395282320678234, + 0.002662254963070154, + 0.0026062773540616035, + 0.003318483242765069, + 0.002082016784697771, + 0.0050762421451509, + 0.0020560771226882935, + 0.0011153104715049267, + 0.007906988263130188, + 0.010857509449124336, + 0.015208314172923565, + 0.009164627641439438, + 0.00865521002560854, + 0.007747035939246416, + 0.0074373153038322926, + 0.008022861555218697, + 0.006463463418185711, + 0.0053281947039067745, + 0.008023588918149471, + 0.005752802360802889, + 0.007135000545531511, + 0.010031885467469692, + 0.006696494296193123, + 0.004720231518149376, + 0.016187191009521484, + 0.011072749271988869, + 0.02824833244085312, + 0.011891629546880722, + 0.014626754447817802, + 0.03382095694541931, + 0.01069718413054943, + 0.014315761625766754, + 0.005040918011218309, + 0.0036312732845544815, + 0.006204220931977034, + 0.005612688604742289, + 0.0031293167266994715, + 0.006612091325223446, + 0.00576807139441371, + 0.004896209109574556, + 0.006133254151791334, + 0.005483327433466911, + 0.007913988083600998, + 0.005588050931692123, + 0.003608783008530736, + 0.0064897313714027405, + 0.004943112842738628, + 0.006153016351163387, + 0.03281522914767265, + 0.027866005897521973, + 0.03422916308045387, + 0.0340445451438427, + 0.015641504898667336, + 0.015528463758528233, + 0.019137036055326462, + 0.027338769286870956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.021180162206292152 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270795464515686 + ], + "opposite_logprob_tokens": [ + -2.127079486846924 + ], + "logprob_reward": -0.1270795464515686, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270795464515686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3010699152946472, + 0.07300388067960739, + 0.07768836617469788, + 0.03832412511110306, + 0.11533408612012863, + 0.0845993384718895, + 0.08398724347352982, + 0.03748304769396782, + 0.0500623919069767, + 0.13844750821590424 + ], + "credit_map": [ + 0.07790468633174896, + 0.02578653208911419, + 0.04207901656627655, + 0.009877552278339863, + 0.022948656231164932, + 0.10426891595125198, + 0.007715270388871431, + 0.010489306412637234, + 0.007709876634180546, + 0.015515161678195, + 0.010268203914165497, + 0.010247470811009407, + 0.006537390872836113, + 0.008813324384391308, + 0.00906929187476635, + 0.0048431591130793095, + 0.01035340130329132, + 0.014267117716372013, + 0.006548569072037935, + 0.01297274511307478, + 0.00858632568269968, + 0.010184414684772491, + 0.00878805760294199, + 0.005987740587443113, + 0.005481462925672531, + 0.004729108419269323, + 0.005852874368429184, + 0.004160753451287746, + 0.006165096536278725, + 0.005553882569074631, + 0.0035451564472168684, + 0.0028357920236885548, + 0.012845314107835293, + 0.013014791533350945, + 0.026223549619317055, + 0.015205946750938892, + 0.014035526663064957, + 0.011963429860770702, + 0.013536098413169384, + 0.008509427309036255, + 0.013133944012224674, + 0.008530247956514359, + 0.015629185363650322, + 0.008332392200827599, + 0.0062525589019060135, + 0.021009335294365883, + 0.00712564354762435, + 0.004586036782711744, + 0.009646263904869556, + 0.011558039113879204, + 0.01203496940433979, + 0.010170748457312584, + 0.0126447519287467, + 0.010774663649499416, + 0.00722144078463316, + 0.00993636529892683, + 0.003972348757088184, + 0.004720637574791908, + 0.005080525763332844, + 0.006088175345212221, + 0.002834013430401683, + 0.003200699109584093, + 0.006884471513330936, + 0.004702176433056593, + 0.006262029986828566, + 0.006235896144062281, + 0.007627677638083696, + 0.005586476996541023, + 0.004931486211717129, + 0.0077889771200716496, + 0.005915502086281776, + 0.005714341998100281, + 0.021457437425851822, + 0.02029481530189514, + 0.02224692888557911, + 0.02175389975309372, + 0.01113265659660101, + 0.010609096847474575, + 0.013158163987100124, + 0.017794517800211906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.1270795464515686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813595116138458 + ], + "opposite_logprob_tokens": [ + -3.529813528060913 + ], + "logprob_reward": -0.029813595116138458, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813595116138458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.269561767578125, + 0.060823846608400345, + 0.04350269213318825, + 0.03629612922668457, + 0.33832207322120667, + 0.034119512885808945, + 0.060047101229429245, + 0.0253453366458416, + 0.06549979746341705, + 0.06648170948028564 + ], + "credit_map": [ + 0.08474467694759369, + 0.027637118473649025, + 0.015175282955169678, + 0.011726776137948036, + 0.02348697930574417, + 0.08753149956464767, + 0.009997235611081123, + 0.009262202307581902, + 0.007055488880723715, + 0.00997439119964838, + 0.008267746306955814, + 0.009883066639304161, + 0.005879399366676807, + 0.006984116043895483, + 0.008705333806574345, + 0.004074300639331341, + 0.005458462983369827, + 0.005505278706550598, + 0.0049187070690095425, + 0.006814093794673681, + 0.007555530406534672, + 0.005167972296476364, + 0.005254581104964018, + 0.0028280685655772686, + 0.005157020408660173, + 0.004929347429424524, + 0.003508202964439988, + 0.00637346226722002, + 0.004081858787685633, + 0.005332562606781721, + 0.005009324289858341, + 0.0019043494248762727, + 0.039544813334941864, + 0.035847071558237076, + 0.08023379743099213, + 0.04515279084444046, + 0.0413656085729599, + 0.029146356508135796, + 0.04271763190627098, + 0.0243140310049057, + 0.0037074172869324684, + 0.004339586012065411, + 0.004368412308394909, + 0.004219230730086565, + 0.004409091547131538, + 0.0047867028042674065, + 0.005065940786153078, + 0.003223131876438856, + 0.006684673950076103, + 0.009436244145035744, + 0.006352519616484642, + 0.006915575824677944, + 0.010609176941215992, + 0.0080190971493721, + 0.005322461947798729, + 0.006707351189106703, + 0.0028329542838037014, + 0.0029283210169523954, + 0.0038614978548139334, + 0.0025172678288072348, + 0.0018984676571562886, + 0.00408268254250288, + 0.004338046535849571, + 0.0028861011378467083, + 0.008685607463121414, + 0.0077167390845716, + 0.011213237419724464, + 0.0071672797203063965, + 0.0068608238361775875, + 0.009165641851723194, + 0.006732897832989693, + 0.00795756932348013, + 0.007593910209834576, + 0.007904171012341976, + 0.010167223401367664, + 0.013054978102445602, + 0.005900267511606216, + 0.005125674419105053, + 0.006368403322994709, + 0.010367076843976974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.029813595116138458 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635456044226885 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635456044226885, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635456044226885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2584395110607147, + 0.04553045332431793, + 0.032192837446928024, + 0.019285423681139946, + 0.17389576137065887, + 0.1121223047375679, + 0.043429821729660034, + 0.043464016169309616, + 0.045035604387521744, + 0.22660435736179352 + ], + "credit_map": [ + 0.04894166439771652, + 0.05482517555356026, + 0.02430211566388607, + 0.02207152545452118, + 0.040157876908779144, + 0.04132307693362236, + 0.01572902873158455, + 0.011089012958109379, + 0.004310950171202421, + 0.007706827484071255, + 0.004925055895000696, + 0.007504837121814489, + 0.004357537720352411, + 0.006091873627156019, + 0.0073433867655694485, + 0.003289969405159354, + 0.004403259139508009, + 0.005376784130930901, + 0.0034305346198379993, + 0.005964222364127636, + 0.0035882818046957254, + 0.0037860439624637365, + 0.003742590779438615, + 0.0019011121476069093, + 0.0019295528763905168, + 0.0021893682423979044, + 0.001938975416123867, + 0.0022838797885924578, + 0.002831839257851243, + 0.0028562662191689014, + 0.003587684128433466, + 0.0016678544925525784, + 0.020080657675862312, + 0.01972123421728611, + 0.04566328227519989, + 0.02122563309967518, + 0.021356767043471336, + 0.014861922711133957, + 0.01915762946009636, + 0.011828596703708172, + 0.014241773635149002, + 0.011995543725788593, + 0.01455957442522049, + 0.01392404455691576, + 0.016183357685804367, + 0.018992852419614792, + 0.014086119830608368, + 0.008139017969369888, + 0.005523850210011005, + 0.00415258901193738, + 0.007632724475115538, + 0.0037294768262654543, + 0.004859942011535168, + 0.01004235353320837, + 0.004202981013804674, + 0.003285895334556699, + 0.0053578573279082775, + 0.004303573165088892, + 0.007021389901638031, + 0.009296083822846413, + 0.003006624523550272, + 0.005067009478807449, + 0.004503854550421238, + 0.004907611757516861, + 0.007334482856094837, + 0.005174722988158464, + 0.009617934003472328, + 0.005170580931007862, + 0.0035517988726496696, + 0.003982599824666977, + 0.0041167475283145905, + 0.006086726672947407, + 0.029160670936107635, + 0.03141191601753235, + 0.03041435219347477, + 0.026894940063357353, + 0.021372316405177116, + 0.03404299169778824, + 0.029175961390137672, + 0.0241311676800251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.007635456044226885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895100116729736 + ], + "opposite_logprob_tokens": [ + -2.578950881958008 + ], + "logprob_reward": -0.07895100116729736, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895100116729736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49743521213531494, + 0.04774932935833931, + 0.11132414638996124, + 0.02035309188067913, + 0.05344606563448906, + 0.070094533264637, + 0.13105733692646027, + 0.02879074402153492, + 0.018478816375136375, + 0.021270781755447388 + ], + "credit_map": [ + 0.0907333642244339, + 0.025342831388115883, + 0.11472968012094498, + 0.02911452017724514, + 0.0305587500333786, + 0.16737383604049683, + 0.015915529802441597, + 0.023666663095355034, + 0.004494376014918089, + 0.007534163072705269, + 0.0051176827400922775, + 0.00797613337635994, + 0.0038803680799901485, + 0.006801488343626261, + 0.00826045498251915, + 0.003684664610773325, + 0.010003956034779549, + 0.01592235453426838, + 0.007616234943270683, + 0.015092005021870136, + 0.014063689857721329, + 0.012366646900773048, + 0.02550080604851246, + 0.010758452117443085, + 0.002638433827087283, + 0.002855027327314019, + 0.0019578428473323584, + 0.002006123773753643, + 0.002601948566734791, + 0.005139058455824852, + 0.002047402784228325, + 0.0011072532506659627, + 0.005334331188350916, + 0.006204178091138601, + 0.007816134952008724, + 0.00965970754623413, + 0.005289529450237751, + 0.007381797768175602, + 0.006101081147789955, + 0.005659305956214666, + 0.009425519965589046, + 0.006850745994597673, + 0.011648744344711304, + 0.006604148540645838, + 0.0044394670985639095, + 0.018061984330415726, + 0.005612635985016823, + 0.007451291661709547, + 0.015703503042459488, + 0.018300168216228485, + 0.014360795728862286, + 0.013431482017040253, + 0.022107727825641632, + 0.017113767564296722, + 0.012912716716527939, + 0.01712717115879059, + 0.003777448320761323, + 0.0025478117167949677, + 0.00478034932166338, + 0.004443502984941006, + 0.002100013429298997, + 0.0035650618374347687, + 0.00391894718632102, + 0.003657612018287182, + 0.002261742949485779, + 0.0016816714778542519, + 0.0033856043592095375, + 0.0022496692836284637, + 0.0013302663573995233, + 0.0024938220158219337, + 0.002048611408099532, + 0.003027429571375251, + 0.002082815393805504, + 0.003312022890895605, + 0.0028876150026917458, + 0.0033966624177992344, + 0.002578198444098234, + 0.0016674769576638937, + 0.002864543115720153, + 0.0024814484640955925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.07895100116729736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme wide-angle lens, looking upward from the ground, tilts the camera upward in a slow, majestic rhythm, gradually revealing a cosmic spectacle. Against the velvety, deep indigo sky, countless fiery meteors intertwine densely, forming a veritable waterfall of light. These meteors do not streak haphazardly, but rather cascade vertically and unstoppably in a grand, unified rhythm. Each meteor trails a long, shimmering trail of blazing white, icy blue, and fleeting gold. These trails briefly illuminate the faint, distant stars and nebulous dust in the background as they fall. This luminous torrent, the sole source of light for the entire scene, casts dynamic, flowing shadows upon the faintly visible, rolling mountain silhouettes at the bottom of the frame, creating a breathtaking, almost sacred atmosphere of grandeur, imbuing the image with a sense of cosmic tranquility and awe.", + "seed": 24790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_10/index_4264_seed_24790_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": -0.0807080801576376, + "vlm_logprob": -0.0807080801576376, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.0807080801576376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016321372240781784 + ], + "opposite_logprob_tokens": [ + -5.266321182250977 + ], + "logprob_reward": -0.016321372240781784, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.016321372240781784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033726755529642105, + 0.029647154733538628, + 0.6769221425056458, + 0.01966666243970394, + 0.06618257611989975, + 0.046030350029468536, + 0.01729782298207283, + 0.04886889085173607, + 0.011337816715240479, + 0.050319839268922806 + ], + "credit_map": [ + 0.004262183327227831, + 0.0027835906948894262, + 0.003249702975153923, + 0.006748981308192015, + 0.002777558984234929, + 0.004269706550985575, + 0.005924209952354431, + 0.0037108208052814007, + 0.003617549082264304, + 0.0048495009541511536, + 0.002476136665791273, + 0.003210032358765602, + 0.003519400954246521, + 0.004240293055772781, + 0.003967517055571079, + 0.003766726003959775, + 0.06539218127727509, + 0.26021862030029297, + 0.015093470923602581, + 0.01991315372288227, + 0.025672920048236847, + 0.22880373895168304, + 0.011356113478541374, + 0.05047192424535751, + 0.002014265628531575, + 0.002858127234503627, + 0.0018878901610150933, + 0.0028546294197440147, + 0.002987763611599803, + 0.0020471778698265553, + 0.0028196529019623995, + 0.0021971554961055517, + 0.005772717762738466, + 0.010165455751121044, + 0.005132618825882673, + 0.005959151778370142, + 0.005283534526824951, + 0.007270352449268103, + 0.006148862186819315, + 0.02044988051056862, + 0.004385930020362139, + 0.006782574113458395, + 0.005507767200469971, + 0.005457291845232248, + 0.007200433872640133, + 0.0071930973790585995, + 0.00552538875490427, + 0.003977867774665356, + 0.001930265803821385, + 0.0017874211771413684, + 0.0018051810329779983, + 0.002518613822758198, + 0.0016918941400945187, + 0.002280258806422353, + 0.002470509149134159, + 0.002813678467646241, + 0.003592327469959855, + 0.004057710524648428, + 0.008605715818703175, + 0.013673328794538975, + 0.0023339465260505676, + 0.00380844809114933, + 0.004037679173052311, + 0.008759735152125359, + 0.0011525786248967052, + 0.0015219663036987185, + 0.0015682687517255545, + 0.0017841904191300273, + 0.0013225609436631203, + 0.0010167630389332771, + 0.0019326438196003437, + 0.0010388449300080538, + 0.007867071777582169, + 0.00789716187864542, + 0.0054734800942242146, + 0.00871849711984396, + 0.005130574572831392, + 0.005876992829144001, + 0.0037472641561180353, + 0.005608794745057821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.016321372240781784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28136903047561646 + ], + "opposite_logprob_tokens": [ + -1.4063689708709717 + ], + "logprob_reward": -0.28136903047561646, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28136903047561646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03594961017370224, + 0.04443453997373581, + 0.2889888882637024, + 0.03919103741645813, + 0.25430217385292053, + 0.07473355531692505, + 0.026946160942316055, + 0.03526968136429787, + 0.018836496397852898, + 0.1813478022813797 + ], + "credit_map": [ + 0.0047589261084795, + 0.0033153912518173456, + 0.0027889003977179527, + 0.004300592467188835, + 0.003455945523455739, + 0.004702226724475622, + 0.007750715594738722, + 0.004876911174505949, + 0.00504001509398222, + 0.004294624552130699, + 0.004696692805737257, + 0.007737876381725073, + 0.005604909732937813, + 0.006612379103899002, + 0.0060774702578783035, + 0.004370573442429304, + 0.026684312149882317, + 0.10848362743854523, + 0.006802224088460207, + 0.00884808599948883, + 0.012959709390997887, + 0.09709712117910385, + 0.0061753885820508, + 0.021938420832157135, + 0.004913577809929848, + 0.0051245600916445255, + 0.003666413715109229, + 0.005070502869784832, + 0.004037588369101286, + 0.00499066524207592, + 0.003971642814576626, + 0.007416089065372944, + 0.024716665968298912, + 0.027643872424960136, + 0.022830015048384666, + 0.026409577578306198, + 0.02137935906648636, + 0.04043131321668625, + 0.025437571108341217, + 0.06545381993055344, + 0.007594479713588953, + 0.009053485468029976, + 0.00880223698914051, + 0.011670841835439205, + 0.0075637237168848515, + 0.008640659041702747, + 0.013780432753264904, + 0.007627693470567465, + 0.003196662524715066, + 0.0032843579538166523, + 0.00272013689391315, + 0.0033469193149358034, + 0.0032230725046247244, + 0.003056059591472149, + 0.003818897530436516, + 0.004300054628401995, + 0.003603070043027401, + 0.0029023182578384876, + 0.006125905551016331, + 0.00678697694092989, + 0.002925317967310548, + 0.004104017745703459, + 0.0031040343455970287, + 0.005718042142689228, + 0.0025607794523239136, + 0.0025379471480846405, + 0.0022900309413671494, + 0.0021691012661904097, + 0.002362507162615657, + 0.002152402652427554, + 0.002126613399013877, + 0.002637114841490984, + 0.02684224396944046, + 0.0243113674223423, + 0.021120022982358932, + 0.029621202498674393, + 0.021394850686192513, + 0.02304503507912159, + 0.015386959537863731, + 0.019626116380095482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.28136903047561646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302864283323288 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302864283323288, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302864283323288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03814563900232315, + 0.03799615427851677, + 0.22679314017295837, + 0.09158238768577576, + 0.15463854372501373, + 0.21040557324886322, + 0.06470144540071487, + 0.05944802984595299, + 0.030906783416867256, + 0.08538226783275604 + ], + "credit_map": [ + 0.004328908398747444, + 0.004056198988109827, + 0.004200631752610207, + 0.003841052297502756, + 0.005342546384781599, + 0.004556558094918728, + 0.005651714745908976, + 0.006168029736727476, + 0.004496983252465725, + 0.00448768911883235, + 0.0021112849935889244, + 0.0026181910652667284, + 0.003823090111836791, + 0.009528979659080505, + 0.004456035792827606, + 0.006473897956311703, + 0.020479300990700722, + 0.08346925675868988, + 0.006035114638507366, + 0.008166207000613213, + 0.008833086118102074, + 0.07719109207391739, + 0.005445408169180155, + 0.017173685133457184, + 0.011364961974322796, + 0.013255048543214798, + 0.009060949087142944, + 0.01277593057602644, + 0.009559430181980133, + 0.011845413595438004, + 0.009723911061882973, + 0.013996747322380543, + 0.016209343448281288, + 0.020580939948558807, + 0.011735726147890091, + 0.014122938737273216, + 0.015110144391655922, + 0.01708025299012661, + 0.018616126850247383, + 0.04118306562304497, + 0.015554256737232208, + 0.026707328855991364, + 0.01743343472480774, + 0.02537543885409832, + 0.03593507781624794, + 0.028369002044200897, + 0.044953152537345886, + 0.016077879816293716, + 0.005026205908507109, + 0.00997677631676197, + 0.007719412446022034, + 0.012191233225166798, + 0.005621908698230982, + 0.0068658702075481415, + 0.011482805013656616, + 0.005817234981805086, + 0.004920451436191797, + 0.00343199516646564, + 0.016124652698636055, + 0.006904428359121084, + 0.0017324700020253658, + 0.0034103880170732737, + 0.002714406931772828, + 0.020209236070513725, + 0.003977265674620867, + 0.002855788916349411, + 0.004413204733282328, + 0.003737988881766796, + 0.0030819568783044815, + 0.00372297246940434, + 0.00508531229570508, + 0.004032291006296873, + 0.01274419017136097, + 0.010110478848218918, + 0.00993979163467884, + 0.012820389121770859, + 0.013729865662753582, + 0.01051939558237791, + 0.008216350339353085, + 0.00730180786922574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.023302864283323288 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550657615065575 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012550657615065575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550657615065575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05253852531313896, + 0.029674576595425606, + 0.5033515691757202, + 0.05291717126965523, + 0.040512364357709885, + 0.04701768979430199, + 0.1167985126376152, + 0.10207650065422058, + 0.022182144224643707, + 0.03293102607131004 + ], + "credit_map": [ + 0.005405320320278406, + 0.004693150520324707, + 0.006971210706979036, + 0.0048697772435843945, + 0.006856312043964863, + 0.007024019490927458, + 0.00731049245223403, + 0.009408236481249332, + 0.0038093223702162504, + 0.0024900590069592, + 0.0027212826535105705, + 0.003777407808229327, + 0.0033533128444105387, + 0.0060647884383797646, + 0.0034292603377252817, + 0.004029139876365662, + 0.04786184802651405, + 0.19233104586601257, + 0.011379954405128956, + 0.015336072072386742, + 0.019391363486647606, + 0.17039771378040314, + 0.008904945105314255, + 0.03774857148528099, + 0.008055153302848339, + 0.0080904271453619, + 0.005737331695854664, + 0.00810847245156765, + 0.00481951329857111, + 0.005354768596589565, + 0.005495244171470404, + 0.007256255950778723, + 0.0036652404814958572, + 0.0063000950030982494, + 0.005728733725845814, + 0.005639493465423584, + 0.004356399644166231, + 0.004704817198216915, + 0.004859019536525011, + 0.005258559715002775, + 0.0044592805206775665, + 0.005498200189322233, + 0.0072811986319720745, + 0.00576574495062232, + 0.006623413879424334, + 0.0052994959987699986, + 0.0073961117304861546, + 0.004694237839430571, + 0.010491554625332355, + 0.01295188907533884, + 0.009689385071396828, + 0.01931597851216793, + 0.011118992231786251, + 0.014360261149704456, + 0.028353601694107056, + 0.010516832582652569, + 0.008554589934647083, + 0.00681602256372571, + 0.022253453731536865, + 0.022246280685067177, + 0.003895977744832635, + 0.006543158087879419, + 0.00549191003665328, + 0.026275090873241425, + 0.0030169605743139982, + 0.002345975721254945, + 0.0031507485546171665, + 0.0031504989601671696, + 0.0023095300421118736, + 0.0018732852768152952, + 0.0031763743609189987, + 0.003158767241984606, + 0.004623733926564455, + 0.00572060514241457, + 0.004025851376354694, + 0.004782830830663443, + 0.0028285870794206858, + 0.003014540998265147, + 0.003970551770180464, + 0.003964316099882126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.012550657615065575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999647617340088 + ], + "opposite_logprob_tokens": [ + -2.6949963569641113 + ], + "logprob_reward": -0.06999647617340088, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999647617340088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049090683460235596, + 0.08424492925405502, + 0.6229367852210999, + 0.016964493319392204, + 0.04814862087368965, + 0.026728877797722816, + 0.02358962781727314, + 0.09380009025335312, + 0.009910059161484241, + 0.024585723876953125 + ], + "credit_map": [ + 0.005511188413947821, + 0.004082741681486368, + 0.006002907641232014, + 0.007311737164855003, + 0.003924952354282141, + 0.006282177288085222, + 0.009227432310581207, + 0.006747549399733543, + 0.011002031154930592, + 0.007474162615835667, + 0.00774538516998291, + 0.009680170565843582, + 0.0051889363676309586, + 0.022755807265639305, + 0.007248205598443747, + 0.013150235638022423, + 0.059293173253536224, + 0.23844283819198608, + 0.014153460040688515, + 0.01855095662176609, + 0.023214679211378098, + 0.21207427978515625, + 0.010340536944568157, + 0.04686696454882622, + 0.0023833774030208588, + 0.0029447709675878286, + 0.0016828161897137761, + 0.002180693903937936, + 0.0026369362603873014, + 0.0019281480927020311, + 0.0017362528014928102, + 0.001471499097533524, + 0.00464874180033803, + 0.008512703701853752, + 0.006898597814142704, + 0.005956636276096106, + 0.003474593162536621, + 0.00364399841055274, + 0.00391243863850832, + 0.011100913397967815, + 0.00251447013579309, + 0.004005648195743561, + 0.0025963305961340666, + 0.0026357288006693125, + 0.004220954608172178, + 0.0041430918499827385, + 0.004096033982932568, + 0.002516621258109808, + 0.0025524115189909935, + 0.001999733503907919, + 0.003852876601740718, + 0.004437404219061136, + 0.0019227080047130585, + 0.003362012794241309, + 0.0027757908683270216, + 0.002686693100258708, + 0.007628230843693018, + 0.005359983071684837, + 0.025512531399726868, + 0.013121531344950199, + 0.0026075912173837423, + 0.005226461216807365, + 0.0032243391033262014, + 0.031119434162974358, + 0.001302780001424253, + 0.0012636209139600396, + 0.0012899383436888456, + 0.0011790301650762558, + 0.00089485157513991, + 0.0011582417646422982, + 0.0013389470987021923, + 0.0014826501719653606, + 0.003543999744579196, + 0.003125574206933379, + 0.0027873723302036524, + 0.003609665436670184, + 0.0028787788469344378, + 0.0035265551414340734, + 0.0021907552145421505, + 0.002923024585470557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.06999647617340088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme wide-angle lens, looking upward from the ground, tilts the camera upward in a slow, majestic rhythm, gradually revealing a cosmic spectacle. Against the velvety, deep indigo sky, countless fiery meteors intertwine densely, forming a veritable waterfall of light. These meteors do not streak haphazardly, but rather cascade vertically and unstoppably in a grand, unified rhythm. Each meteor trails a long, shimmering trail of blazing white, icy blue, and fleeting gold. These trails briefly illuminate the faint, distant stars and nebulous dust in the background as they fall. This luminous torrent, the sole source of light for the entire scene, casts dynamic, flowing shadows upon the faintly visible, rolling mountain silhouettes at the bottom of the frame, creating a breathtaking, almost sacred atmosphere of grandeur, imbuing the image with a sense of cosmic tranquility and awe.", + "seed": 24791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_10/index_4264_seed_24791_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": -0.06872088937088847, + "vlm_logprob": -0.06872088937088847, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.06872088937088847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007737427484244108 + ], + "opposite_logprob_tokens": [ + -6.007737636566162 + ], + "logprob_reward": -0.007737427484244108, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007737427484244108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12402496486902237, + 0.10699599236249924, + 0.14149877429008484, + 0.0941321924328804, + 0.07824232429265976, + 0.0714021697640419, + 0.16374510526657104, + 0.11269834637641907, + 0.06804206222295761, + 0.03921804949641228 + ], + "credit_map": [ + 0.024925708770751953, + 0.028276512399315834, + 0.009979741647839546, + 0.015541395172476768, + 0.012021619826555252, + 0.009170112200081348, + 0.010041448287665844, + 0.01406843215227127, + 0.01399284414947033, + 0.011907242238521576, + 0.011763876304030418, + 0.011836372315883636, + 0.011987687088549137, + 0.01456572487950325, + 0.012560078874230385, + 0.01838216744363308, + 0.02868313156068325, + 0.020253991708159447, + 0.012329446151852608, + 0.021039828658103943, + 0.01514072809368372, + 0.020344791933894157, + 0.012049632146954536, + 0.011657224036753178, + 0.012372495606541634, + 0.021786801517009735, + 0.007195018231868744, + 0.011097719892859459, + 0.010880086570978165, + 0.013529202900826931, + 0.008737841621041298, + 0.008533031679689884, + 0.00966303888708353, + 0.008897142484784126, + 0.017077412456274033, + 0.01148134283721447, + 0.005984176881611347, + 0.010790691711008549, + 0.005296571180224419, + 0.009051946923136711, + 0.009124798700213432, + 0.011333189904689789, + 0.007933932356536388, + 0.011945368722081184, + 0.006381102837622166, + 0.009703077375888824, + 0.006347123067826033, + 0.008633574470877647, + 0.01586541347205639, + 0.014029748737812042, + 0.024784937500953674, + 0.022635186091065407, + 0.010520214214920998, + 0.03503682464361191, + 0.021957596763968468, + 0.018915191292762756, + 0.01895039528608322, + 0.014339794404804707, + 0.01218035351485014, + 0.013079870492219925, + 0.009923777543008327, + 0.013463443145155907, + 0.014226471073925495, + 0.016534235328435898, + 0.008347004652023315, + 0.008518560789525509, + 0.01088048331439495, + 0.008865759707987309, + 0.008189704269170761, + 0.006914760917425156, + 0.011150157079100609, + 0.005175633355975151, + 0.004254302009940147, + 0.00613748375326395, + 0.004751766100525856, + 0.006002082489430904, + 0.004460691940039396, + 0.0031839224975556135, + 0.006702115293592215, + 0.0037256870418787003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.007737427484244108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813675105571747 + ], + "opposite_logprob_tokens": [ + -1.406367540359497 + ], + "logprob_reward": -0.2813675105571747, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813675105571747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05380550026893616, + 0.1487514078617096, + 0.1241355761885643, + 0.027845436707139015, + 0.047366779297590256, + 0.05365997180342674, + 0.3456854522228241, + 0.13540665805339813, + 0.04632039740681648, + 0.017022835090756416 + ], + "credit_map": [ + 0.0077568949200212955, + 0.0077706086449325085, + 0.0060871560126543045, + 0.0058721206150949, + 0.008270330727100372, + 0.007789420895278454, + 0.005630487110465765, + 0.004628483671694994, + 0.014122119173407555, + 0.015139208175241947, + 0.013996533118188381, + 0.013161483220756054, + 0.03455236926674843, + 0.020801367238163948, + 0.013149069622159004, + 0.023829255253076553, + 0.02210048772394657, + 0.023277323693037033, + 0.010441018268465996, + 0.017578085884451866, + 0.015389600768685341, + 0.017205556854605675, + 0.011174523271620274, + 0.006968993693590164, + 0.0031417300924658775, + 0.0040756831876933575, + 0.0028951375279575586, + 0.0032718784641474485, + 0.0037334123626351357, + 0.00390649912878871, + 0.0029526192229241133, + 0.0038684778846800327, + 0.005011389497667551, + 0.005500013008713722, + 0.00802522525191307, + 0.006395347882062197, + 0.0037819435819983482, + 0.0062890080735087395, + 0.005119489040225744, + 0.007244362961500883, + 0.007345317862927914, + 0.008280249312520027, + 0.00743032107129693, + 0.008369196206331253, + 0.005672437138855457, + 0.007464212365448475, + 0.004192921798676252, + 0.004905319772660732, + 0.030775923281908035, + 0.013526780530810356, + 0.09610263258218765, + 0.02302585169672966, + 0.013913819566369057, + 0.11147014051675797, + 0.01225623395293951, + 0.044614095240831375, + 0.01603175327181816, + 0.03204934671521187, + 0.01621696725487709, + 0.019432445988059044, + 0.012347058393061161, + 0.011989281512796879, + 0.012365687638521194, + 0.014974132180213928, + 0.006499619223177433, + 0.005218431353569031, + 0.012005511671304703, + 0.004819828551262617, + 0.003908113110810518, + 0.0038221608847379684, + 0.0044298539869487286, + 0.005616881884634495, + 0.0013922360958531499, + 0.0015125774079933763, + 0.0032718810252845287, + 0.0021048234775662422, + 0.002309655537828803, + 0.001466005458496511, + 0.0032759581226855516, + 0.0016896985471248627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.2813675105571747 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016074584797024727 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.016074584797024727, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074584797024727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08218319714069366, + 0.05667325109243393, + 0.13973034918308258, + 0.12547896802425385, + 0.050559576600790024, + 0.05618152394890785, + 0.06985162198543549, + 0.28090333938598633, + 0.04878717660903931, + 0.08965102583169937 + ], + "credit_map": [ + 0.01460571400821209, + 0.014760573394596577, + 0.006298184860497713, + 0.010380905121564865, + 0.012814939953386784, + 0.006430611480027437, + 0.009005553089082241, + 0.007886708714067936, + 0.004493082407861948, + 0.0051399231888353825, + 0.007185203954577446, + 0.004973571747541428, + 0.011723923496901989, + 0.00767964543774724, + 0.005554387345910072, + 0.009923510253429413, + 0.027714833617210388, + 0.021666396409273148, + 0.011832513846457005, + 0.021901417523622513, + 0.014513539150357246, + 0.020387833938002586, + 0.01140526868402958, + 0.010308533906936646, + 0.02025691792368889, + 0.02038884162902832, + 0.009572438895702362, + 0.016262304037809372, + 0.016252266243100166, + 0.019476672634482384, + 0.01102529652416706, + 0.012244224548339844, + 0.005869008135050535, + 0.005491504445672035, + 0.007869246415793896, + 0.007035932503640652, + 0.0035835301969200373, + 0.006831105798482895, + 0.0061105722561478615, + 0.007768670096993446, + 0.007863444276154041, + 0.0067481365986168385, + 0.007147091440856457, + 0.008286123163998127, + 0.006701669190078974, + 0.007816601544618607, + 0.006125880405306816, + 0.005492573603987694, + 0.007079167291522026, + 0.0047101289965212345, + 0.018307292833924294, + 0.009263642132282257, + 0.004124405328184366, + 0.010388554073870182, + 0.008369969204068184, + 0.007608460262417793, + 0.036992985755205154, + 0.05398952215909958, + 0.04756169393658638, + 0.05705413222312927, + 0.01901135966181755, + 0.01726399175822735, + 0.019432539120316505, + 0.029597114771604538, + 0.007927191443741322, + 0.009503715671598911, + 0.006210949271917343, + 0.005525656510144472, + 0.005680185277014971, + 0.003304282436147332, + 0.0056754425168037415, + 0.0049597471952438354, + 0.008283674716949463, + 0.009261771105229855, + 0.013400332070887089, + 0.01191339548677206, + 0.01462442334741354, + 0.009156686253845692, + 0.014220474287867546, + 0.008790263906121254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.016074584797024727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644422516226768 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644422516226768, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644422516226768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057847779244184494, + 0.048428621143102646, + 0.056845396757125854, + 0.04744463786482811, + 0.17141763865947723, + 0.21438436210155487, + 0.16854733228683472, + 0.06455141305923462, + 0.07669273763895035, + 0.0938400849699974 + ], + "credit_map": [ + 0.010438724420964718, + 0.010943903587758541, + 0.006917876191437244, + 0.008461369201540947, + 0.006328025832772255, + 0.004926150664687157, + 0.005260372534394264, + 0.004571356810629368, + 0.004687492270022631, + 0.00501810060814023, + 0.007526221685111523, + 0.005780987907201052, + 0.006031945813447237, + 0.006459768861532211, + 0.004778951872140169, + 0.00814515259116888, + 0.008292562328279018, + 0.0070892865769565105, + 0.006833516526967287, + 0.007866496220231056, + 0.0049417875707149506, + 0.00830253679305315, + 0.00655010249465704, + 0.006969110574573278, + 0.006111332681030035, + 0.008482787758111954, + 0.004274897277355194, + 0.0066682808101177216, + 0.005910112056881189, + 0.0071363491006195545, + 0.0033988249488174915, + 0.005462056491523981, + 0.022339601069688797, + 0.015094608999788761, + 0.027014674618840218, + 0.020305098965764046, + 0.015936946496367455, + 0.023062657564878464, + 0.02370322309434414, + 0.023960834369063377, + 0.028904657810926437, + 0.04308090731501579, + 0.020129570737481117, + 0.029861371964216232, + 0.02789401076734066, + 0.031029000878334045, + 0.017298951745033264, + 0.016185881569981575, + 0.016055624932050705, + 0.009171121753752232, + 0.04834446310997009, + 0.01288837194442749, + 0.007230853196233511, + 0.049789536744356155, + 0.006554975640028715, + 0.018512383103370667, + 0.007467996329069138, + 0.011998714879155159, + 0.010394946672022343, + 0.009438419714570045, + 0.0056650685146451, + 0.0048474655486643314, + 0.00658639520406723, + 0.00815240852534771, + 0.0071606808342039585, + 0.007377303671091795, + 0.01818218268454075, + 0.008440800942480564, + 0.007141054142266512, + 0.0055700065568089485, + 0.010129568167030811, + 0.012691142037510872, + 0.007642315234988928, + 0.007803511805832386, + 0.01557512953877449, + 0.011757074855268002, + 0.012128865346312523, + 0.011629932560026646, + 0.016263483092188835, + 0.011039766483008862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.008644422516226768 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the meteors cascading vertically in a dense, unified flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the meteor trails display colors of blazing white, icy blue, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint, rolling mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780501499772072 + ], + "opposite_logprob_tokens": [ + -3.529780387878418 + ], + "logprob_reward": -0.029780501499772072, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780501499772072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09566251188516617, + 0.17857158184051514, + 0.22671295702457428, + 0.1336674988269806, + 0.07624274492263794, + 0.05794818699359894, + 0.06934963166713715, + 0.09251915663480759, + 0.04682764783501625, + 0.02249804325401783 + ], + "credit_map": [ + 0.015675272792577744, + 0.01806586980819702, + 0.009765038266777992, + 0.012153194285929203, + 0.014158113859593868, + 0.010133788920938969, + 0.007801698986440897, + 0.007909531705081463, + 0.01594053953886032, + 0.016704656183719635, + 0.026059040799736977, + 0.02041640877723694, + 0.02891319803893566, + 0.017930081114172935, + 0.018325988203287125, + 0.0342816524207592, + 0.023323602974414825, + 0.01928216964006424, + 0.024762174114584923, + 0.03480970114469528, + 0.017811598256230354, + 0.032506972551345825, + 0.033980317413806915, + 0.040236424654722214, + 0.012993739917874336, + 0.02616371586918831, + 0.014692299999296665, + 0.017079491168260574, + 0.014272594824433327, + 0.017883190885186195, + 0.012978640384972095, + 0.017603835090994835, + 0.005304913502186537, + 0.006035033147782087, + 0.004205001518130302, + 0.006947863847017288, + 0.004209228325635195, + 0.01463338453322649, + 0.008713235147297382, + 0.026194080710411072, + 0.006604914087802172, + 0.005174040328711271, + 0.005245890002697706, + 0.008069013245403767, + 0.005063140764832497, + 0.008938191458582878, + 0.0051957289688289165, + 0.013657266274094582, + 0.005890996661037207, + 0.0046699149534106255, + 0.016856888309121132, + 0.006747353821992874, + 0.004987578373402357, + 0.016813689842820168, + 0.006593361031264067, + 0.006789849605411291, + 0.009694916196167469, + 0.010394373908638954, + 0.009408481419086456, + 0.015450576320290565, + 0.012424588203430176, + 0.011885507963597775, + 0.013036130927503109, + 0.01022457703948021, + 0.006584460847079754, + 0.004878907464444637, + 0.011792805045843124, + 0.005248940549790859, + 0.004107407759875059, + 0.0038902335800230503, + 0.004553769249469042, + 0.005771122872829437, + 0.001756301848217845, + 0.0017906967550516129, + 0.00447717634961009, + 0.0029431984294205904, + 0.002753922715783119, + 0.0024302606470882893, + 0.004031514748930931, + 0.002314972225576639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.029780501499772072 + } + ], + "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/0000125/group_11/index_tensor([4246, 4246])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_11/index_tensor([4246, 4246])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3d223325249b84d444fafc5323a97fc4e651bec2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_11/index_tensor([4246, 4246])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a dynamic first-person perspective, the shot opens with a close-up of hands gripping the worn brown rubber grips of the polished chrome handlebars of a charming ice cream tricycle. The tricycle itself is painted a whimsical soft mint green and features a small brass bell. Suddenly, as the tricycle spins and lifts off the ground, the camera, acting as the rider's eye, experiences a rapid and exhilarating upward swing. In a single, fluid movement, the camera tilts sharply upward, revealing the cobblestone street below as it quickly fades into a richly textured gray-brown pattern. As the tricycle rises into a bright azure sky, its frame fills with fluffy, Ghibli-esque clouds, their edges gleaming with warm golden sunlight. Below, a vast, sun-drenched pastoral landscape stretches to the horizon, a breathtaking display of rolling emerald hills, a field of vibrant wildflowers, and a distant, winding river shimmering in the sunlight. The entire scene is animated in Miyazaki's gentle painterly style, defined by soft-focus details, a rich, warm color palette, and a play of bright light, evoking a profound atmosphere of dreamlike wonder and liberating adventure.", + "seed": 24471, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_11/index_4246_seed_24471_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": -0.3551888033747673, + "vlm_logprob": -0.3551888033747673, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.3551888033747673, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01200050674378872 + ], + "opposite_logprob_tokens": [ + -5.012000560760498 + ], + "logprob_reward": -0.01200050674378872, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01200050674378872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1425885558128357, + 0.07260797172784805, + 0.0858839601278305, + 0.08953860402107239, + 0.030997324734926224, + 0.03853359445929527, + 0.08400899916887283, + 0.10745468735694885, + 0.29356318712234497, + 0.05482310429215431 + ], + "credit_map": [ + 0.018932770937681198, + 0.011288815177977085, + 0.02701534703373909, + 0.01962941884994507, + 0.010439769364893436, + 0.013001946732401848, + 0.016433212906122208, + 0.02584727294743061, + 0.009472906589508057, + 0.004634367302060127, + 0.02017391286790371, + 0.014029729180037975, + 0.004736241418868303, + 0.005780997220426798, + 0.0073085506446659565, + 0.006471262779086828, + 0.007304593920707703, + 0.006266881711781025, + 0.013998775742948055, + 0.022267937660217285, + 0.005888566840440035, + 0.006056192796677351, + 0.006926585920155048, + 0.017174430191516876, + 0.006630886811763048, + 0.012911982834339142, + 0.011929330416023731, + 0.01292937807738781, + 0.006351317744702101, + 0.00817884225398302, + 0.008725298568606377, + 0.021881574764847755, + 0.0038116162177175283, + 0.003691847203299403, + 0.004817567765712738, + 0.004157682415097952, + 0.0029898493085056543, + 0.003925899043679237, + 0.0034794847015291452, + 0.004123377148061991, + 0.004529510624706745, + 0.0039846873842179775, + 0.007370604667812586, + 0.003851196961477399, + 0.004069959744811058, + 0.0033045934978872538, + 0.004293669480830431, + 0.007129370700567961, + 0.008525128476321697, + 0.009243525564670563, + 0.012007229961454868, + 0.008245974779129028, + 0.009592357091605663, + 0.017044322565197945, + 0.009306274354457855, + 0.01004419382661581, + 0.015201875939965248, + 0.01093091070652008, + 0.00855631660670042, + 0.015992604196071625, + 0.012168792076408863, + 0.008601146750152111, + 0.02359507977962494, + 0.012407955713570118, + 0.026577062904834747, + 0.034451380372047424, + 0.029660165309906006, + 0.020173655822873116, + 0.04361419007182121, + 0.06612768024206161, + 0.03694017231464386, + 0.03601887449622154, + 0.005427331663668156, + 0.006653464399278164, + 0.0067829955369234085, + 0.007717962842434645, + 0.008811770007014275, + 0.00539412209764123, + 0.006105971056967974, + 0.007929484359920025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.01200050674378872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054975736886262894 + ], + "opposite_logprob_tokens": [ + -2.929975748062134 + ], + "logprob_reward": -0.054975736886262894, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054975736886262894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13095787167549133, + 0.06988973915576935, + 0.10565701872110367, + 0.11894293874502182, + 0.0659513995051384, + 0.08489446341991425, + 0.14380374550819397, + 0.12033861875534058, + 0.1115364283323288, + 0.048027776181697845 + ], + "credit_map": [ + 0.013506877236068249, + 0.01014956459403038, + 0.02132413536310196, + 0.017291929572820663, + 0.022405685856938362, + 0.010483975522220135, + 0.02006511576473713, + 0.01573059894144535, + 0.012542428448796272, + 0.004176244139671326, + 0.014220359735190868, + 0.011258330196142197, + 0.003623798256739974, + 0.006751639768481255, + 0.005216232500970364, + 0.012100711464881897, + 0.01025158166885376, + 0.008435784839093685, + 0.01905670389533043, + 0.02027679979801178, + 0.006857422646135092, + 0.008999243378639221, + 0.01066374871879816, + 0.021115735173225403, + 0.017768247053027153, + 0.0231227558106184, + 0.015584627166390419, + 0.016893355175852776, + 0.010771394707262516, + 0.01153517048805952, + 0.009419561363756657, + 0.0138478334993124, + 0.008496611379086971, + 0.009103064425289631, + 0.012638118118047714, + 0.007511576171964407, + 0.004396810662001371, + 0.009390086866915226, + 0.005241101142019033, + 0.009174035862088203, + 0.010378822684288025, + 0.008688462898135185, + 0.020315419882535934, + 0.010056569240987301, + 0.006438150070607662, + 0.01011152658611536, + 0.008526249788701534, + 0.010379268787801266, + 0.015267605893313885, + 0.01110592670738697, + 0.035018231719732285, + 0.010608835145831108, + 0.012223830446600914, + 0.0392494834959507, + 0.00897255539894104, + 0.011357282288372517, + 0.024421004578471184, + 0.02158389426767826, + 0.0069198571145534515, + 0.007520342245697975, + 0.018283508718013763, + 0.008832693099975586, + 0.01787625066936016, + 0.014901069924235344, + 0.009214234538376331, + 0.009761422872543335, + 0.016498513519763947, + 0.011673742905259132, + 0.014751475304365158, + 0.022975346073508263, + 0.012923366390168667, + 0.01373833417892456, + 0.004457439295947552, + 0.00615566223859787, + 0.004406773950904608, + 0.004517348483204842, + 0.005273580085486174, + 0.005521393846720457, + 0.008370799012482166, + 0.009324782527983189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.054975736886262894 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7014960050582886 + ], + "opposite_logprob_tokens": [ + -0.2014959752559662 + ], + "logprob_reward": -1.7014960050582886, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7014960050582886, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10824134200811386, + 0.06380921602249146, + 0.07214295864105225, + 0.07528964430093765, + 0.04514339566230774, + 0.07022680342197418, + 0.1935509592294693, + 0.09009488672018051, + 0.1898449808359146, + 0.09165575355291367 + ], + "credit_map": [ + 0.015681616961956024, + 0.009257403202354908, + 0.016454020515084267, + 0.01504646334797144, + 0.009057578630745411, + 0.010143018327653408, + 0.01062091626226902, + 0.02198033593595028, + 0.013119353912770748, + 0.004064377397298813, + 0.012060542590916157, + 0.009015226736664772, + 0.005280390381813049, + 0.007683406583964825, + 0.004761798772960901, + 0.00782412476837635, + 0.006943059153854847, + 0.004838006105273962, + 0.012820296920835972, + 0.020872972905635834, + 0.0031425596680492163, + 0.005065979901701212, + 0.005207662470638752, + 0.01325242593884468, + 0.013157452456653118, + 0.007639449089765549, + 0.011891004629433155, + 0.011920535005629063, + 0.0058878506533801556, + 0.006051286589354277, + 0.008548088371753693, + 0.010193988680839539, + 0.004871905781328678, + 0.005220427643507719, + 0.010411317460238934, + 0.006383943371474743, + 0.002593035576865077, + 0.004230461549013853, + 0.0028196752537041903, + 0.008612634614109993, + 0.009728646837174892, + 0.007650605868548155, + 0.010654840618371964, + 0.008202554658055305, + 0.006230778526514769, + 0.008477015420794487, + 0.008352648466825485, + 0.010929723270237446, + 0.027208320796489716, + 0.019542355090379715, + 0.029504256322979927, + 0.01692330464720726, + 0.020206155255436897, + 0.0310506708920002, + 0.025618435814976692, + 0.023497479036450386, + 0.019208179786801338, + 0.012602249160408974, + 0.005978730041533709, + 0.006460105534642935, + 0.014870918355882168, + 0.008668647147715092, + 0.009418690577149391, + 0.012887375429272652, + 0.016955355182290077, + 0.017321590334177017, + 0.022751282900571823, + 0.015754781663417816, + 0.029469730332493782, + 0.04305531829595566, + 0.021709121763706207, + 0.022827811539173126, + 0.00949295237660408, + 0.008890633471310139, + 0.009293781593441963, + 0.010075455531477928, + 0.008524555712938309, + 0.010201905854046345, + 0.016354378312826157, + 0.018822094425559044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -1.7014960050582886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004650253802537918 + ], + "opposite_logprob_tokens": [ + -5.379650115966797 + ], + "logprob_reward": -0.004650253802537918, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004650253802537918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0822884738445282, + 0.061818692833185196, + 0.09729420393705368, + 0.0543120801448822, + 0.04009367898106575, + 0.05573146417737007, + 0.1893361359834671, + 0.10058393329381943, + 0.18524682521820068, + 0.13329455256462097 + ], + "credit_map": [ + 0.009771606884896755, + 0.0068299793638288975, + 0.017679715529084206, + 0.011593024246394634, + 0.009771476499736309, + 0.006968972738832235, + 0.008203809149563313, + 0.011469895020127296, + 0.012421448715031147, + 0.003700066125020385, + 0.012281865812838078, + 0.009027262218296528, + 0.0048979478888213634, + 0.006733724381774664, + 0.004762876778841019, + 0.007993501611053944, + 0.014689238741993904, + 0.006791685242205858, + 0.016235487535595894, + 0.017902273684740067, + 0.003985767252743244, + 0.0072994534857571125, + 0.006436905357986689, + 0.023953396826982498, + 0.009659898467361927, + 0.005834608338773251, + 0.007720481604337692, + 0.008235028013586998, + 0.004407500848174095, + 0.004415522329509258, + 0.006535226013511419, + 0.007503812201321125, + 0.004789865110069513, + 0.004890759009867907, + 0.007339468691498041, + 0.0054885391145944595, + 0.003323647193610668, + 0.004449144937098026, + 0.0040545086376369, + 0.005757746286690235, + 0.007432620041072369, + 0.0058113522827625275, + 0.008931828662753105, + 0.006683721672743559, + 0.004600634332746267, + 0.007024464663118124, + 0.006537799723446369, + 0.008709041401743889, + 0.021650191396474838, + 0.015764571726322174, + 0.042251188308000565, + 0.011058908887207508, + 0.01780092529952526, + 0.048662617802619934, + 0.01572730764746666, + 0.016420414671301842, + 0.023590048775076866, + 0.013484387658536434, + 0.006120595615357161, + 0.007353294175118208, + 0.012494717724621296, + 0.006595471408218145, + 0.012556992471218109, + 0.018388427793979645, + 0.020751245319843292, + 0.01977388933300972, + 0.0283249169588089, + 0.018984321504831314, + 0.02561447210609913, + 0.03521907702088356, + 0.016643637791275978, + 0.01993526704609394, + 0.009724317118525505, + 0.013717884197831154, + 0.009610630571842194, + 0.012716744095087051, + 0.01525241881608963, + 0.0158211812376976, + 0.024116864427924156, + 0.03233451023697853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.004650253802537918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821514382958412 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821514382958412, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821514382958412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1355096697807312, + 0.07080474495887756, + 0.06080871820449829, + 0.0834212675690651, + 0.029491456225514412, + 0.06360207498073578, + 0.1499984860420227, + 0.11645562946796417, + 0.2252795696258545, + 0.06462839990854263 + ], + "credit_map": [ + 0.02197856456041336, + 0.010179662145674229, + 0.02342757023870945, + 0.017680823802947998, + 0.014517370611429214, + 0.013061048462986946, + 0.012572049163281918, + 0.02209259383380413, + 0.013299627229571342, + 0.004633903503417969, + 0.01150649506598711, + 0.011566226370632648, + 0.007612068671733141, + 0.0077439360320568085, + 0.006593802012503147, + 0.007848696783185005, + 0.009868285618722439, + 0.0051779979839921, + 0.01045442745089531, + 0.010818639770150185, + 0.0033310335129499435, + 0.0060194083489477634, + 0.004785012453794479, + 0.010353918187320232, + 0.011683106422424316, + 0.010710705071687698, + 0.014283529482781887, + 0.0161672942340374, + 0.004300315864384174, + 0.006746017839759588, + 0.005799577571451664, + 0.013730724342167377, + 0.003665592521429062, + 0.0024959908332675695, + 0.005683406256139278, + 0.005613642279058695, + 0.0020667409989982843, + 0.0032553367782384157, + 0.002966972067952156, + 0.0037437784485518932, + 0.007321584969758987, + 0.006937279365956783, + 0.009734947234392166, + 0.007480105385184288, + 0.004775788635015488, + 0.008579739369452, + 0.006930252071470022, + 0.011842377483844757, + 0.019818976521492004, + 0.014537042006850243, + 0.028692565858364105, + 0.010379030369222164, + 0.01598276197910309, + 0.03105233982205391, + 0.01376004982739687, + 0.015775732696056366, + 0.03677602484822273, + 0.022730570286512375, + 0.0065621486864984035, + 0.006519971415400505, + 0.017240429297089577, + 0.008677777834236622, + 0.008272051811218262, + 0.009676658548414707, + 0.03035380318760872, + 0.03202488645911217, + 0.01276264525949955, + 0.02352704107761383, + 0.02737140841782093, + 0.036498695611953735, + 0.029582178220152855, + 0.033158913254737854, + 0.006394903641194105, + 0.007652879226952791, + 0.007681949529796839, + 0.009683949872851372, + 0.008394227363169193, + 0.00782056339085102, + 0.008862404152750969, + 0.008137528784573078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.002821514382958412 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a dynamic first-person perspective, the shot opens with a close-up of hands gripping the worn brown rubber grips of the polished chrome handlebars of a charming ice cream tricycle. The tricycle itself is painted a whimsical soft mint green and features a small brass bell. Suddenly, as the tricycle spins and lifts off the ground, the camera, acting as the rider's eye, experiences a rapid and exhilarating upward swing. In a single, fluid movement, the camera tilts sharply upward, revealing the cobblestone street below as it quickly fades into a richly textured gray-brown pattern. As the tricycle rises into a bright azure sky, its frame fills with fluffy, Ghibli-esque clouds, their edges gleaming with warm golden sunlight. Below, a vast, sun-drenched pastoral landscape stretches to the horizon, a breathtaking display of rolling emerald hills, a field of vibrant wildflowers, and a distant, winding river shimmering in the sunlight. The entire scene is animated in Miyazaki's gentle painterly style, defined by soft-focus details, a rich, warm color palette, and a play of bright light, evoking a profound atmosphere of dreamlike wonder and liberating adventure.", + "seed": 24472, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_11/index_4246_seed_24472_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": -0.17894517940003424, + "vlm_logprob": -0.17894517940003424, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.17894517940003424, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022837018594145775 + ], + "opposite_logprob_tokens": [ + -4.022837162017822 + ], + "logprob_reward": -0.022837018594145775, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.022837018594145775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7757856845855713, + 0.05163506418466568, + 0.057214997708797455, + 0.02365671843290329, + 0.014831854030489922, + 0.022623959928750992, + 0.009370038285851479, + 0.02144838310778141, + 0.014189708046615124, + 0.00924365222454071 + ], + "credit_map": [ + 0.07461737841367722, + 0.058173421770334244, + 0.16723662614822388, + 0.09330902993679047, + 0.09187576919794083, + 0.0806121975183487, + 0.07266510277986526, + 0.13729608058929443, + 0.004616947844624519, + 0.007565145380795002, + 0.008959006518125534, + 0.0073702107183635235, + 0.0065468912944197655, + 0.0048712193965911865, + 0.00646580196917057, + 0.005239835474640131, + 0.004168209619820118, + 0.006525065749883652, + 0.005890709348022938, + 0.010124917142093182, + 0.009901650249958038, + 0.004094097763299942, + 0.008632210083305836, + 0.007878134027123451, + 0.0020625463221222162, + 0.002767255064100027, + 0.00379756442271173, + 0.002305596135556698, + 0.0038558323867619038, + 0.0028045186772942543, + 0.0033518157433718443, + 0.0027115882840007544, + 0.002271207282319665, + 0.0012973668053746223, + 0.0011147281620651484, + 0.001391484634950757, + 0.0018304423429071903, + 0.0017314509022980928, + 0.0020189080387353897, + 0.003176263999193907, + 0.002324832836166024, + 0.0023627944756299257, + 0.0022114322055131197, + 0.004022030625492334, + 0.002589515410363674, + 0.0028975983150303364, + 0.002108152024447918, + 0.004107603337615728, + 0.0014190026558935642, + 0.0012255868641659617, + 0.001553863170556724, + 0.001514560543000698, + 0.0007074360037222505, + 0.0010244334116578102, + 0.0009611663408577442, + 0.0009639878408052027, + 0.0013968892162665725, + 0.002040090737864375, + 0.0019387048669159412, + 0.002744266763329506, + 0.005135750398039818, + 0.002580526750534773, + 0.0013407503720372915, + 0.004271403886377811, + 0.004792391788214445, + 0.0014163684099912643, + 0.0021644628141075373, + 0.0016289438353851438, + 0.0011418514186516404, + 0.0008003807161003351, + 0.0011752323480322957, + 0.0010700749699026346, + 0.0012486856430768967, + 0.001613333006389439, + 0.0010616234503686428, + 0.0011831768788397312, + 0.0015756007051095366, + 0.0010415127035230398, + 0.0008566671167500317, + 0.0006630520219914615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.022837018594145775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812839835882187 + ], + "opposite_logprob_tokens": [ + -3.28812837600708 + ], + "logprob_reward": -0.03812839835882187, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812839835882187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23604892194271088, + 0.10482725501060486, + 0.11797764897346497, + 0.0958239808678627, + 0.11924511939287186, + 0.04779696464538574, + 0.0544002465903759, + 0.152435302734375, + 0.04634937271475792, + 0.025095155462622643 + ], + "credit_map": [ + 0.015526359900832176, + 0.01783132366836071, + 0.05835074186325073, + 0.029325757175683975, + 0.022733652964234352, + 0.01467357762157917, + 0.030716627836227417, + 0.04689089581370354, + 0.01253688707947731, + 0.01068849302828312, + 0.020397547632455826, + 0.012962846085429192, + 0.011673645116388798, + 0.009174744598567486, + 0.011232691816985607, + 0.016160402446985245, + 0.009451862424612045, + 0.013248692266643047, + 0.014206784777343273, + 0.023260490968823433, + 0.01338932104408741, + 0.01114044152200222, + 0.013132148422300816, + 0.020147908478975296, + 0.009329339489340782, + 0.011069987900555134, + 0.014393296092748642, + 0.010223640128970146, + 0.013590151444077492, + 0.012617348693311214, + 0.014527026563882828, + 0.010073188692331314, + 0.01590665802359581, + 0.014762545004487038, + 0.011377137154340744, + 0.010323991067707539, + 0.010505769401788712, + 0.012490483932197094, + 0.01987936533987522, + 0.0239991694688797, + 0.0072334446012973785, + 0.0052103749476373196, + 0.005750305484980345, + 0.006355108227580786, + 0.004543924704194069, + 0.0044729202054440975, + 0.006702763494104147, + 0.007528119720518589, + 0.007839495316147804, + 0.008027540519833565, + 0.00923109333962202, + 0.0063207121565938, + 0.004287108778953552, + 0.006358245387673378, + 0.005701902788132429, + 0.006634149234741926, + 0.012753228656947613, + 0.049524858593940735, + 0.01755235530436039, + 0.020893406122922897, + 0.015290622599422932, + 0.010257995687425137, + 0.007401200942695141, + 0.018761640414595604, + 0.008450076915323734, + 0.007575140334665775, + 0.006774837151169777, + 0.007435518316924572, + 0.006641024257987738, + 0.002824306720867753, + 0.0033002085983753204, + 0.0033482606522738934, + 0.003953918348997831, + 0.002752861240878701, + 0.003068030346184969, + 0.003668352495878935, + 0.0030448706820607185, + 0.0028767164330929518, + 0.0028459655586630106, + 0.0028844387270510197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.03812839835882187 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8260951042175293 + ], + "opposite_logprob_tokens": [ + -0.5760951042175293 + ], + "logprob_reward": -0.8260951042175293, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8260951042175293, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2553868889808655, + 0.09267263114452362, + 0.029278669506311417, + 0.05357665941119194, + 0.04979119449853897, + 0.05687069520354271, + 0.078262560069561, + 0.3117302656173706, + 0.049098506569862366, + 0.02333187870681286 + ], + "credit_map": [ + 0.02559550106525421, + 0.019582008942961693, + 0.05507303774356842, + 0.03102567046880722, + 0.028488898649811745, + 0.026032086461782455, + 0.023414723575115204, + 0.04617495462298393, + 0.009981179609894753, + 0.008018715307116508, + 0.011768530122935772, + 0.00839188788086176, + 0.008366537280380726, + 0.013852287083864212, + 0.010793662630021572, + 0.02149982750415802, + 0.002074564341455698, + 0.003255596850067377, + 0.003716921666637063, + 0.004903683438897133, + 0.004295716527849436, + 0.0025964982341974974, + 0.003154595848172903, + 0.005281094461679459, + 0.0038402569480240345, + 0.005221415776759386, + 0.006625151261687279, + 0.005172130651772022, + 0.009683133102953434, + 0.008797701448202133, + 0.008525318466126919, + 0.005711547564715147, + 0.012790033593773842, + 0.0063081406988203526, + 0.005239373538643122, + 0.004073380492627621, + 0.0038510372396558523, + 0.005336775444447994, + 0.006071818992495537, + 0.006120631936937571, + 0.00628789234906435, + 0.00434426125138998, + 0.005762291140854359, + 0.010010434314608574, + 0.006639019586145878, + 0.0073963673785328865, + 0.006899931933730841, + 0.009530498646199703, + 0.012132047675549984, + 0.009489307180047035, + 0.014645724557340145, + 0.009013374336063862, + 0.006999582052230835, + 0.008966092951595783, + 0.0066092428751289845, + 0.01040718611329794, + 0.020413992926478386, + 0.11730757355690002, + 0.03854099661111832, + 0.031441181898117065, + 0.069582499563694, + 0.010086532682180405, + 0.01252315565943718, + 0.011834351345896721, + 0.009170986711978912, + 0.007431335747241974, + 0.006597030442208052, + 0.007748415227979422, + 0.0062883999198675156, + 0.003096820553764701, + 0.004079396370798349, + 0.0046861255541443825, + 0.0027251781430095434, + 0.0038318356964737177, + 0.0027559781447052956, + 0.0038493038155138493, + 0.004103484563529491, + 0.0019315286772325635, + 0.002449341583997011, + 0.001685229130089283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.8260951042175293 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950947757810354 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950947757810354, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950947757810354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37710538506507874, + 0.09367086738348007, + 0.03949347510933876, + 0.06915456056594849, + 0.04955499619245529, + 0.06031668558716774, + 0.09483477473258972, + 0.11064980179071426, + 0.032619696110486984, + 0.07259970903396606 + ], + "credit_map": [ + 0.030722158029675484, + 0.023316143080592155, + 0.10471032559871674, + 0.044399600476026535, + 0.03254050016403198, + 0.027225973084568977, + 0.02788919024169445, + 0.08630150556564331, + 0.008509505540132523, + 0.012483670376241207, + 0.01307083759456873, + 0.013067963533103466, + 0.013569505885243416, + 0.0069852108135819435, + 0.010424066334962845, + 0.015560104511678219, + 0.002909492701292038, + 0.004287758842110634, + 0.004857099149376154, + 0.005233116913586855, + 0.005588263738900423, + 0.003181485226377845, + 0.007588639389723539, + 0.005847619846463203, + 0.008856994099915028, + 0.008422774262726307, + 0.00921161100268364, + 0.007956845685839653, + 0.007499296683818102, + 0.009301367215812206, + 0.010344238020479679, + 0.007561434525996447, + 0.007993272505700588, + 0.0060493904165923595, + 0.004765211138874292, + 0.004528823774307966, + 0.00615772046148777, + 0.0048971096985042095, + 0.006490591913461685, + 0.008672877214848995, + 0.008577732369303703, + 0.005385400727391243, + 0.007996520958840847, + 0.008990530855953693, + 0.006309683434665203, + 0.00536994356662035, + 0.009774741716682911, + 0.00791213195770979, + 0.012074067257344723, + 0.016242193058133125, + 0.015509992837905884, + 0.012508774176239967, + 0.009296942502260208, + 0.009404324926435947, + 0.011423453688621521, + 0.008375033736228943, + 0.010385083965957165, + 0.029481811448931694, + 0.018350981175899506, + 0.01049699354916811, + 0.020744169130921364, + 0.006834639701992273, + 0.006319727282971144, + 0.008036394603550434, + 0.009013318456709385, + 0.0032460263464599848, + 0.005760796368122101, + 0.003960703033953905, + 0.0026583445724099874, + 0.003056198125705123, + 0.0023178032133728266, + 0.0026065048296004534, + 0.01036038901656866, + 0.00911629106849432, + 0.01289296243339777, + 0.010745951905846596, + 0.007480214349925518, + 0.00760220643132925, + 0.006163352634757757, + 0.00823834165930748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.005950947757810354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017144280718639493 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.0017144280718639493, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017144280718639493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.602601945400238, + 0.05277090147137642, + 0.05081034451723099, + 0.041813258081674576, + 0.015684718266129494, + 0.0381532683968544, + 0.07327881455421448, + 0.05848916992545128, + 0.018252883106470108, + 0.048144664615392685 + ], + "credit_map": [ + 0.05861477181315422, + 0.04763910546898842, + 0.12383595108985901, + 0.07939797639846802, + 0.07226548343896866, + 0.06162805110216141, + 0.05691723898053169, + 0.10230335593223572, + 0.006058901082724333, + 0.006090599577873945, + 0.008544960990548134, + 0.0068592592142522335, + 0.005974321626126766, + 0.0055555920116603374, + 0.005223975051194429, + 0.008463293313980103, + 0.003782502142712474, + 0.00588461896404624, + 0.006346375681459904, + 0.011721337214112282, + 0.006062572821974754, + 0.005403431132435799, + 0.005799039267003536, + 0.005810469854623079, + 0.004508616868406534, + 0.004305331036448479, + 0.007268643472343683, + 0.004087479319423437, + 0.004444872960448265, + 0.006597424391657114, + 0.004746812861412764, + 0.005854074377566576, + 0.003020475385710597, + 0.001991554396227002, + 0.0013740656431764364, + 0.0014910075115039945, + 0.002328579779714346, + 0.0015173287829384208, + 0.0019952491857111454, + 0.001966456649824977, + 0.004680758807808161, + 0.002854629186913371, + 0.0038144541904330254, + 0.006836225278675556, + 0.003525655483826995, + 0.005121305119246244, + 0.005155271850526333, + 0.006164967082440853, + 0.01241548452526331, + 0.010819585993885994, + 0.016198355704545975, + 0.007047147024422884, + 0.00557589391246438, + 0.0075868768617510796, + 0.0057529243640601635, + 0.007882543839514256, + 0.00470383744686842, + 0.009850673377513885, + 0.0070444317534565926, + 0.00934895034879446, + 0.007256546523422003, + 0.00576000427827239, + 0.0030489317141473293, + 0.011475789360702038, + 0.00445307232439518, + 0.001733897952362895, + 0.0026422948576509953, + 0.002445787191390991, + 0.0023441005032509565, + 0.0014080540277063847, + 0.001535783871077001, + 0.0016898917965590954, + 0.005260799080133438, + 0.0067544844932854176, + 0.0069759925827383995, + 0.0063133505173027515, + 0.005698356777429581, + 0.005245789885520935, + 0.007089458871632814, + 0.004806431010365486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0017144280718639493 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a dynamic first-person perspective, the shot opens with a close-up of hands gripping the worn brown rubber grips of the polished chrome handlebars of a charming ice cream tricycle. The tricycle itself is painted a whimsical soft mint green and features a small brass bell. Suddenly, as the tricycle spins and lifts off the ground, the camera, acting as the rider's eye, experiences a rapid and exhilarating upward swing. In a single, fluid movement, the camera tilts sharply upward, revealing the cobblestone street below as it quickly fades into a richly textured gray-brown pattern. As the tricycle rises into a bright azure sky, its frame fills with fluffy, Ghibli-esque clouds, their edges gleaming with warm golden sunlight. Below, a vast, sun-drenched pastoral landscape stretches to the horizon, a breathtaking display of rolling emerald hills, a field of vibrant wildflowers, and a distant, winding river shimmering in the sunlight. The entire scene is animated in Miyazaki's gentle painterly style, defined by soft-focus details, a rich, warm color palette, and a play of bright light, evoking a profound atmosphere of dreamlike wonder and liberating adventure.", + "seed": 24473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_11/index_4246_seed_24473_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": -0.13888791690114885, + "vlm_logprob": -0.13888791690114885, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.13888791690114885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016325123608112335 + ], + "opposite_logprob_tokens": [ + -4.516324996948242 + ], + "logprob_reward": -0.016325123608112335, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.016325123608112335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17269167304039001, + 0.10299551486968994, + 0.0457434244453907, + 0.06554081290960312, + 0.07987800985574722, + 0.3294612467288971, + 0.06110870838165283, + 0.06403256207704544, + 0.05082442983984947, + 0.027723615989089012 + ], + "credit_map": [ + 0.016437288373708725, + 0.014116109348833561, + 0.02671622298657894, + 0.022305717691779137, + 0.015576338395476341, + 0.024082042276859283, + 0.022965053096413612, + 0.03049289993941784, + 0.008242746815085411, + 0.009095036424696445, + 0.024182137101888657, + 0.017217757180333138, + 0.009690206497907639, + 0.009745103307068348, + 0.008514576591551304, + 0.016307955607771873, + 0.004011527635157108, + 0.003037496702745557, + 0.007990452460944653, + 0.005321978125721216, + 0.004148800857365131, + 0.004930043593049049, + 0.006869640666991472, + 0.009433485567569733, + 0.006662335246801376, + 0.00912248995155096, + 0.00702032120898366, + 0.007304871920496225, + 0.008901822380721569, + 0.008743180893361568, + 0.010177549906075, + 0.007608245592564344, + 0.006922082044184208, + 0.013597979210317135, + 0.005814507137984037, + 0.004179562442004681, + 0.021068476140499115, + 0.018262911587953568, + 0.0057809725403785706, + 0.004251523409038782, + 0.021190648898482323, + 0.011869581416249275, + 0.11209670454263687, + 0.01683170534670353, + 0.1192999854683876, + 0.02613285928964615, + 0.008553494699299335, + 0.013486267998814583, + 0.017894703894853592, + 0.009017916396260262, + 0.004235782660543919, + 0.005485848523676395, + 0.005898440256714821, + 0.005504590924829245, + 0.005747663322836161, + 0.007323758210986853, + 0.008716191165149212, + 0.009752445854246616, + 0.00702404510229826, + 0.007123442366719246, + 0.006797619163990021, + 0.007830295711755753, + 0.011226127855479717, + 0.005562388803809881, + 0.0044807856902480125, + 0.006871142890304327, + 0.004515188280493021, + 0.005056426394730806, + 0.009364962577819824, + 0.007052388973534107, + 0.007148057222366333, + 0.006335479207336903, + 0.0021274560131132603, + 0.0033977830316871405, + 0.002566599054262042, + 0.004039168823510408, + 0.003956114407628775, + 0.004205847624689341, + 0.002944223117083311, + 0.004486422054469585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.016325123608112335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308383911848068 + ], + "opposite_logprob_tokens": [ + -3.168083906173706 + ], + "logprob_reward": -0.04308383911848068, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308383911848068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08635561913251877, + 0.02192157320678234, + 0.031046707183122635, + 0.04959036037325859, + 0.04159535467624664, + 0.5692558884620667, + 0.09367308765649796, + 0.050569675862789154, + 0.03916069120168686, + 0.01683100312948227 + ], + "credit_map": [ + 0.010543715208768845, + 0.008868765085935593, + 0.017739111557602882, + 0.007446711882948875, + 0.008237186819314957, + 0.009681503288447857, + 0.007169562857598066, + 0.01666906289756298, + 0.0018560101743787527, + 0.0023812642320990562, + 0.003033318789675832, + 0.0029795088339596987, + 0.0026826963294297457, + 0.003139972221106291, + 0.0026087958831340075, + 0.003240007907152176, + 0.0031536209862679243, + 0.002423904137685895, + 0.004405966494232416, + 0.0033917257096618414, + 0.0027838998939841986, + 0.004316417500376701, + 0.00447375001385808, + 0.006097423378378153, + 0.00817989744246006, + 0.005199444480240345, + 0.0035564915742725134, + 0.004718306940048933, + 0.005241715349256992, + 0.0046554384753108025, + 0.005480504594743252, + 0.012558560818433762, + 0.005023590754717588, + 0.0055314162746071815, + 0.002483450109139085, + 0.00205941591411829, + 0.007364603690803051, + 0.007969635538756847, + 0.007978484034538269, + 0.003184756962582469, + 0.03646707907319069, + 0.021908503025770187, + 0.19115936756134033, + 0.027748558670282364, + 0.21737825870513916, + 0.04155902937054634, + 0.014672806486487389, + 0.018362300470471382, + 0.022044135257601738, + 0.012092484161257744, + 0.007209279574453831, + 0.008009595796465874, + 0.009561608545482159, + 0.006534914951771498, + 0.01122271828353405, + 0.016998352482914925, + 0.0037600130308419466, + 0.006035578902810812, + 0.0076263537630438805, + 0.008048544637858868, + 0.005173584911972284, + 0.005326088517904282, + 0.0068620359525084496, + 0.007737473584711552, + 0.003569274675101042, + 0.005215116776525974, + 0.003962683491408825, + 0.003939371090382338, + 0.0076814075000584126, + 0.005358576308935881, + 0.005233204457908869, + 0.004201056901365519, + 0.0012508644722402096, + 0.002099732169881463, + 0.0020257560536265373, + 0.0025096992030739784, + 0.0024706507101655006, + 0.0023432904854416847, + 0.0023895965423434973, + 0.00174141360912472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.04308383911848068 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327739953994751 + ], + "opposite_logprob_tokens": [ + -0.7577739953994751 + ], + "logprob_reward": -0.6327739953994751, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327739953994751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037505216896533966, + 0.016911199316382408, + 0.011163178831338882, + 0.03478989377617836, + 0.17224188148975372, + 0.5555930733680725, + 0.06214325502514839, + 0.02163318544626236, + 0.07542215287685394, + 0.012596945278346539 + ], + "credit_map": [ + 0.004374803509563208, + 0.002451677806675434, + 0.009065551683306694, + 0.0032990670297294855, + 0.002894079778343439, + 0.004561714828014374, + 0.0035409554839134216, + 0.007317365612834692, + 0.0012548386584967375, + 0.0016611100872978568, + 0.004271121695637703, + 0.002245832234621048, + 0.001709523145109415, + 0.0018148409435525537, + 0.0014258455485105515, + 0.0025280872359871864, + 0.0010007427772507071, + 0.0007952084997668862, + 0.0014056038344278932, + 0.0013170859310775995, + 0.0015114578418433666, + 0.0013760325964540243, + 0.0013581759994849563, + 0.002398871351033449, + 0.005450288765132427, + 0.0023972783237695694, + 0.00677470862865448, + 0.0039139240980148315, + 0.0022464809007942677, + 0.004345288034528494, + 0.003483409294858575, + 0.0061785136349499226, + 0.013249248266220093, + 0.03545703738927841, + 0.010057933628559113, + 0.010482832789421082, + 0.050712183117866516, + 0.03512754663825035, + 0.009133500047028065, + 0.008021597750484943, + 0.03492996469140053, + 0.019798189401626587, + 0.1912073791027069, + 0.027616184204816818, + 0.20666219294071198, + 0.042837053537368774, + 0.013107683509588242, + 0.019434398040175438, + 0.01308351568877697, + 0.008003930561244488, + 0.0037296523805707693, + 0.004833428654819727, + 0.013141180388629436, + 0.00494906771928072, + 0.005961862858384848, + 0.008440617471933365, + 0.0016195349162444472, + 0.004065660294145346, + 0.0031116686295717955, + 0.002065882785245776, + 0.0030270968563854694, + 0.002840222790837288, + 0.0023130010813474655, + 0.0025901184417307377, + 0.007250992115586996, + 0.01076364703476429, + 0.007482116110622883, + 0.006852818187326193, + 0.01674453541636467, + 0.010592804290354252, + 0.007988756522536278, + 0.007746484596282244, + 0.0009529807139188051, + 0.0018270729342475533, + 0.0013029280817136168, + 0.0018645573873072863, + 0.0017652654787525535, + 0.0020190253853797913, + 0.0014121290296316147, + 0.0014529856853187084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.6327739953994751 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001336158369667828 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.001336158369667828, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001336158369667828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057230763137340546, + 0.020653491839766502, + 0.017830295488238335, + 0.0684320405125618, + 0.031097764149308205, + 0.583113431930542, + 0.047021135687828064, + 0.03376828879117966, + 0.09650979191064835, + 0.04434290900826454 + ], + "credit_map": [ + 0.007426141761243343, + 0.004518845584243536, + 0.008224095217883587, + 0.006247668527066708, + 0.005311605986207724, + 0.008426716551184654, + 0.00810514111071825, + 0.00897054560482502, + 0.001896764268167317, + 0.001903743133880198, + 0.0028883805498480797, + 0.0028572981245815754, + 0.001516140066087246, + 0.004822267685085535, + 0.0018157173180952668, + 0.0029531822074204683, + 0.002275114180520177, + 0.001308427075855434, + 0.0018955732230097055, + 0.002150810556486249, + 0.0019261816050857306, + 0.002754757646471262, + 0.0027991076931357384, + 0.002720324555411935, + 0.010654306039214134, + 0.006650714203715324, + 0.009789315052330494, + 0.008513353765010834, + 0.0058335233479738235, + 0.0076441396959125996, + 0.006578571163117886, + 0.012768114916980267, + 0.003976217471063137, + 0.005813617259263992, + 0.0030375116039067507, + 0.002211298793554306, + 0.006314414087682962, + 0.005887768231332302, + 0.0019951972644776106, + 0.0018617369933053851, + 0.03744008392095566, + 0.021217502653598785, + 0.19443729519844055, + 0.030675314366817474, + 0.20848551392555237, + 0.04511764273047447, + 0.019074128940701485, + 0.026665979996323586, + 0.014710727147758007, + 0.007310052402317524, + 0.0032817237079143524, + 0.003530394285917282, + 0.005616011563688517, + 0.0038019351195544004, + 0.00447449367493391, + 0.004295796621590853, + 0.0029149740003049374, + 0.005218246020376682, + 0.004509889055043459, + 0.0044511957094073296, + 0.003585723927244544, + 0.0036774782929569483, + 0.0064750853925943375, + 0.002935698488727212, + 0.008652194403111935, + 0.014259935356676579, + 0.010225612670183182, + 0.009407376870512962, + 0.022467128932476044, + 0.013985885307192802, + 0.009605223312973976, + 0.00790643785148859, + 0.0034055300056934357, + 0.006820801645517349, + 0.005847020074725151, + 0.005929096601903439, + 0.006312768440693617, + 0.0069770654663443565, + 0.004063854925334454, + 0.004986772313714027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.001336158369667828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009204680100083351 + ], + "opposite_logprob_tokens": [ + -7.000920295715332 + ], + "logprob_reward": -0.0009204680100083351, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009204680100083351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09298844635486603, + 0.04030396416783333, + 0.03445851430296898, + 0.0802600085735321, + 0.21795222163200378, + 0.18278059363365173, + 0.10101261734962463, + 0.10618238151073456, + 0.07612962275743484, + 0.06793157011270523 + ], + "credit_map": [ + 0.013990833424031734, + 0.007132056634873152, + 0.01771141216158867, + 0.006600667722523212, + 0.007319191936403513, + 0.011646244674921036, + 0.0067183589562773705, + 0.02186969481408596, + 0.003579002572223544, + 0.0041590528562664986, + 0.009766810573637486, + 0.00488602090626955, + 0.002660052850842476, + 0.0047190370969474316, + 0.004937195219099522, + 0.005596795119345188, + 0.0034895632416009903, + 0.0021175704896450043, + 0.006126371677964926, + 0.0037043059710413218, + 0.002173305954784155, + 0.004230384714901447, + 0.004767392762005329, + 0.007849618792533875, + 0.01299181766808033, + 0.008164350874722004, + 0.009926453232765198, + 0.010136135853827, + 0.008089542388916016, + 0.010467559099197388, + 0.007834824733436108, + 0.012649324722588062, + 0.014230429194867611, + 0.015296242199838161, + 0.009246046654880047, + 0.009944306686520576, + 0.07401768863201141, + 0.07833705097436905, + 0.008674453012645245, + 0.008206022903323174, + 0.019936606287956238, + 0.020752280950546265, + 0.022538146004080772, + 0.024144960567355156, + 0.020142415538430214, + 0.02204456552863121, + 0.021214228123426437, + 0.03200741112232208, + 0.020655108615756035, + 0.012448321096599102, + 0.007796300109475851, + 0.00786539539694786, + 0.017133066430687904, + 0.009400429204106331, + 0.010696981102228165, + 0.01501702331006527, + 0.010239502415060997, + 0.013169306330382824, + 0.011998489499092102, + 0.014559558592736721, + 0.013915789313614368, + 0.012721876613795757, + 0.01801753230392933, + 0.01156033854931593, + 0.00700801145285368, + 0.010075801983475685, + 0.008858231827616692, + 0.008557649329304695, + 0.010068613104522228, + 0.00998261570930481, + 0.009175497107207775, + 0.012403211556375027, + 0.0028295705560594797, + 0.003871653927490115, + 0.007072673179209232, + 0.019773460924625397, + 0.005377121269702911, + 0.012349099852144718, + 0.007225183304399252, + 0.009432812221348286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.0009204680100083351 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a dynamic first-person perspective, the shot opens with a close-up of hands gripping the worn brown rubber grips of the polished chrome handlebars of a charming ice cream tricycle. The tricycle itself is painted a whimsical soft mint green and features a small brass bell. Suddenly, as the tricycle spins and lifts off the ground, the camera, acting as the rider's eye, experiences a rapid and exhilarating upward swing. In a single, fluid movement, the camera tilts sharply upward, revealing the cobblestone street below as it quickly fades into a richly textured gray-brown pattern. As the tricycle rises into a bright azure sky, its frame fills with fluffy, Ghibli-esque clouds, their edges gleaming with warm golden sunlight. Below, a vast, sun-drenched pastoral landscape stretches to the horizon, a breathtaking display of rolling emerald hills, a field of vibrant wildflowers, and a distant, winding river shimmering in the sunlight. The entire scene is animated in Miyazaki's gentle painterly style, defined by soft-focus details, a rich, warm color palette, and a play of bright light, evoking a profound atmosphere of dreamlike wonder and liberating adventure.", + "seed": 24474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_11/index_4246_seed_24474_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": -0.39649243764579295, + "vlm_logprob": -0.39649243764579295, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.39649243764579295, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016644658520817757 + ], + "opposite_logprob_tokens": [ + -4.391644477844238 + ], + "logprob_reward": -0.016644658520817757, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016644658520817757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19370825588703156, + 0.11028039455413818, + 0.1092640608549118, + 0.06550269573926926, + 0.029082175344228745, + 0.0921446830034256, + 0.046768590807914734, + 0.04402093589305878, + 0.04241758584976196, + 0.26681068539619446 + ], + "credit_map": [ + 0.02207246981561184, + 0.014098940417170525, + 0.038989096879959106, + 0.019933708012104034, + 0.022987617179751396, + 0.023762518540024757, + 0.025119222700595856, + 0.02674468792974949, + 0.010993961244821548, + 0.008390570990741253, + 0.015281329862773418, + 0.009630212560296059, + 0.02000398561358452, + 0.01059681922197342, + 0.015186089091002941, + 0.0201974269002676, + 0.009020195342600346, + 0.01270431187003851, + 0.01122764591127634, + 0.014819973148405552, + 0.009855212643742561, + 0.018209027126431465, + 0.018412437289953232, + 0.015015256591141224, + 0.005369597114622593, + 0.006237294990569353, + 0.010483837686479092, + 0.0062193432822823524, + 0.006952277850359678, + 0.009374727495014668, + 0.011180795729160309, + 0.009684818796813488, + 0.003788165980949998, + 0.003446386894211173, + 0.003803249215707183, + 0.0036400682292878628, + 0.00498761422932148, + 0.0034557378385215998, + 0.0031491457484662533, + 0.0028118062764406204, + 0.0055846343748271465, + 0.007300770841538906, + 0.010865408927202225, + 0.003861870151013136, + 0.03034830465912819, + 0.025309493765234947, + 0.003766691079363227, + 0.005107501521706581, + 0.005251838825643063, + 0.004619799554347992, + 0.008030141703784466, + 0.0068936776369810104, + 0.00604563532397151, + 0.005476260557770729, + 0.005975436884909868, + 0.004475798923522234, + 0.0038126225117594004, + 0.005244503263384104, + 0.003824477316811681, + 0.007148946635425091, + 0.009752693586051464, + 0.004791621118783951, + 0.0044196005910634995, + 0.005026473198086023, + 0.006291176658123732, + 0.006127147004008293, + 0.005629313178360462, + 0.0047398838214576244, + 0.0056613232009112835, + 0.005059468559920788, + 0.004193370696157217, + 0.004715902265161276, + 0.11568056046962738, + 0.11766170710325241, + 0.004929347895085812, + 0.005043715704232454, + 0.01295717153698206, + 0.003309624968096614, + 0.004085946828126907, + 0.0031426132190972567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.016644658520817757 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867430776357651 + ], + "opposite_logprob_tokens": [ + -3.0486743450164795 + ], + "logprob_reward": -0.04867430776357651, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867430776357651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09327501058578491, + 0.10564655065536499, + 0.08788957446813583, + 0.05078119412064552, + 0.09824520349502563, + 0.3149009943008423, + 0.08448942005634308, + 0.05541778728365898, + 0.049039725214242935, + 0.06031452864408493 + ], + "credit_map": [ + 0.012537405826151371, + 0.009599943645298481, + 0.012868227437138557, + 0.011888591572642326, + 0.009399320930242538, + 0.010965567082166672, + 0.011928020045161247, + 0.014087938703596592, + 0.010500256903469563, + 0.009246409870684147, + 0.012820198200643063, + 0.010300662368535995, + 0.017916230484843254, + 0.01547573134303093, + 0.014187639579176903, + 0.015199418179690838, + 0.008093834854662418, + 0.006724122911691666, + 0.01013145875185728, + 0.01060858741402626, + 0.006010368932038546, + 0.017889076843857765, + 0.008097300305962563, + 0.020334826782345772, + 0.0049899183213710785, + 0.002927292138338089, + 0.0053987763822078705, + 0.002569471253082156, + 0.004486196208745241, + 0.013784284703433514, + 0.005304758902639151, + 0.011320495046675205, + 0.010662752203643322, + 0.011304831132292747, + 0.01448203343898058, + 0.009447293356060982, + 0.017193911597132683, + 0.01209506019949913, + 0.011825498193502426, + 0.011233818717300892, + 0.014926123432815075, + 0.04898509383201599, + 0.043985314667224884, + 0.011201188899576664, + 0.12516097724437714, + 0.04888565093278885, + 0.008846689946949482, + 0.012909972108900547, + 0.010641192086040974, + 0.009965972043573856, + 0.0105945048853755, + 0.01174774020910263, + 0.01127353310585022, + 0.01052656676620245, + 0.009609511122107506, + 0.010130400769412518, + 0.004499479196965694, + 0.007187770679593086, + 0.005542188882827759, + 0.009500409476459026, + 0.012291663326323032, + 0.006659341510385275, + 0.0053010000847280025, + 0.004435932729393244, + 0.007264405023306608, + 0.006768613588064909, + 0.005758235696703196, + 0.006010161712765694, + 0.00691746361553669, + 0.007628854364156723, + 0.004269239492714405, + 0.004422748927026987, + 0.009676346555352211, + 0.012103450484573841, + 0.005146515090018511, + 0.007793175056576729, + 0.00802143570035696, + 0.008541096933186054, + 0.004820882808417082, + 0.004211620893329382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.04867430776357651 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9102965593338013 + ], + "opposite_logprob_tokens": [ + -0.1602965146303177 + ], + "logprob_reward": -1.9102965593338013, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9102965593338013, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09881862998008728, + 0.046274472028017044, + 0.04812231287360191, + 0.0314386710524559, + 0.09731239080429077, + 0.19040243327617645, + 0.08204265683889389, + 0.1127733513712883, + 0.045737918466329575, + 0.24707713723182678 + ], + "credit_map": [ + 0.011773473583161831, + 0.007633238099515438, + 0.02158603072166443, + 0.011163938790559769, + 0.010582800023257732, + 0.009906488470733166, + 0.012378223240375519, + 0.013794450089335442, + 0.00475563807412982, + 0.0045418329536914825, + 0.005034822504967451, + 0.005899123847484589, + 0.008580918423831463, + 0.0058822943829, + 0.005717828869819641, + 0.005862017162144184, + 0.004363393411040306, + 0.0038316354621201754, + 0.005288663320243359, + 0.00538319768384099, + 0.0038195145316421986, + 0.012471622787415981, + 0.005691481754183769, + 0.007272812537848949, + 0.0031563290394842625, + 0.0022235859651118517, + 0.0044610449112951756, + 0.002351384377107024, + 0.003105978947132826, + 0.008539856411516666, + 0.0026490080635994673, + 0.004951485898345709, + 0.014375814236700535, + 0.01085271779447794, + 0.012481874786317348, + 0.0142497094348073, + 0.010319593362510204, + 0.012599688023328781, + 0.011042783036828041, + 0.011390216648578644, + 0.008651516400277615, + 0.019488828256726265, + 0.020142795518040657, + 0.009353280998766422, + 0.05060575157403946, + 0.05879773199558258, + 0.009815075434744358, + 0.013547463342547417, + 0.006358257494866848, + 0.008417545817792416, + 0.011278457008302212, + 0.012899276800453663, + 0.009654722176492214, + 0.008250103332102299, + 0.014847605489194393, + 0.010336697101593018, + 0.010907886549830437, + 0.021419381722807884, + 0.011386711150407791, + 0.013608588837087154, + 0.019304078072309494, + 0.013587982393801212, + 0.009210790507495403, + 0.01334793958812952, + 0.004967206623405218, + 0.005923894699662924, + 0.005402162671089172, + 0.006078167352825403, + 0.005642433650791645, + 0.006383407860994339, + 0.006196487229317427, + 0.005144163500517607, + 0.0997709184885025, + 0.10103048384189606, + 0.008112456649541855, + 0.008292591199278831, + 0.013272745534777641, + 0.006211052183061838, + 0.0051522827707231045, + 0.005234627518802881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -1.9102965593338013 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647406283766031 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647406283766031, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647406283766031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10428983718156815, + 0.06269047409296036, + 0.06591453403234482, + 0.04770158603787422, + 0.11853763461112976, + 0.2330833077430725, + 0.08676546812057495, + 0.07074539363384247, + 0.05597800016403198, + 0.15429379045963287 + ], + "credit_map": [ + 0.011729595251381397, + 0.006825316231697798, + 0.022190941497683525, + 0.011951802298426628, + 0.01294643897563219, + 0.012241569347679615, + 0.011641252785921097, + 0.01476292498409748, + 0.0072492510080337524, + 0.005157534498721361, + 0.007394676096737385, + 0.006334255449473858, + 0.01146471407264471, + 0.009377382695674896, + 0.006815631873905659, + 0.008897033520042896, + 0.005438676569610834, + 0.004651286639273167, + 0.00612257607281208, + 0.005253256764262915, + 0.004361855797469616, + 0.020460430532693863, + 0.009412326849997044, + 0.010214125737547874, + 0.005029433872550726, + 0.0031991086434572935, + 0.0068543883971869946, + 0.0031480782199651003, + 0.004737036302685738, + 0.013602386228740215, + 0.0038487266283482313, + 0.0072824289090931416, + 0.018030239269137383, + 0.015507034957408905, + 0.014127426780760288, + 0.01645777001976967, + 0.01272545475512743, + 0.014991146512329578, + 0.014186915941536427, + 0.012511644512414932, + 0.01219701673835516, + 0.02439107559621334, + 0.023179009556770325, + 0.012948610819876194, + 0.0570925734937191, + 0.07322363555431366, + 0.012469040229916573, + 0.017582327127456665, + 0.007985022850334644, + 0.01208422239869833, + 0.012113974429666996, + 0.01154929120093584, + 0.0112242940813303, + 0.011207827366888523, + 0.009590202011168003, + 0.011010635644197464, + 0.008499033749103546, + 0.01456369087100029, + 0.007303078658878803, + 0.009621662087738514, + 0.010833924636244774, + 0.007709531579166651, + 0.006071833893656731, + 0.006142644211649895, + 0.007543526589870453, + 0.008330939337611198, + 0.005735826212912798, + 0.006423564627766609, + 0.008146826177835464, + 0.008375589735805988, + 0.006663583684712648, + 0.00475814426317811, + 0.059641510248184204, + 0.06047741323709488, + 0.004257483873516321, + 0.007983911782503128, + 0.008552776649594307, + 0.004885576665401459, + 0.004465104546397924, + 0.004030018113553524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.004647406283766031 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tricycle painted soft mint green with a small brass bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the tricycle spinning and lifting off the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt sharply upward to reveal a cobblestone street below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright azure sky with fluffy clouds and a sun-drenched pastoral landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a soft-focus, painterly animation featuring warm lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021992563270032406 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.0021992563270032406, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021992563270032406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13931213319301605, + 0.07801401615142822, + 0.053686290979385376, + 0.030245399102568626, + 0.06560724228620529, + 0.14301007986068726, + 0.10233598202466965, + 0.12019623070955276, + 0.11043167114257812, + 0.15716104209423065 + ], + "credit_map": [ + 0.016822077333927155, + 0.009416463784873486, + 0.03428156301379204, + 0.014957460574805737, + 0.014095650985836983, + 0.016723426058888435, + 0.01323617808520794, + 0.019779302179813385, + 0.009287738241255283, + 0.006182674318552017, + 0.015109261497855186, + 0.008032209239900112, + 0.010382364504039288, + 0.01232017483562231, + 0.006615710444748402, + 0.010083879344165325, + 0.003819773904979229, + 0.003037070157006383, + 0.0047426470555365086, + 0.004127773456275463, + 0.003251198446378112, + 0.01832226850092411, + 0.007621118798851967, + 0.008764438331127167, + 0.002717245602980256, + 0.002701215445995331, + 0.006380184553563595, + 0.0020745352376252413, + 0.0031904205679893494, + 0.0039046998135745525, + 0.0036440554540604353, + 0.005633039399981499, + 0.009682572446763515, + 0.007013073191046715, + 0.008873144164681435, + 0.008187275379896164, + 0.00780355092138052, + 0.009137424640357494, + 0.009260594844818115, + 0.0056496006436645985, + 0.008081483654677868, + 0.01427419763058424, + 0.014631695114076138, + 0.00829118862748146, + 0.03644343465566635, + 0.04424417391419411, + 0.007468906696885824, + 0.009574981406331062, + 0.008460366167128086, + 0.013162527233362198, + 0.014817160554230213, + 0.01705898530781269, + 0.014524649828672409, + 0.009849616326391697, + 0.013552526943385601, + 0.010910139419138432, + 0.014663657173514366, + 0.027307532727718353, + 0.013788692653179169, + 0.018979204818606377, + 0.01723739132285118, + 0.010174873284995556, + 0.007483767811208963, + 0.010561092756688595, + 0.014626235701143742, + 0.016017619520425797, + 0.014816810376942158, + 0.011352842673659325, + 0.01648934744298458, + 0.018731670454144478, + 0.010583505965769291, + 0.007813626900315285, + 0.060869693756103516, + 0.06264329701662064, + 0.004985883366316557, + 0.009069714695215225, + 0.007937384769320488, + 0.004644240252673626, + 0.003407468553632498, + 0.0036033454816788435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0021992563270032406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_tensor([9467, 9467])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_tensor([9467, 9467])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7c09a837ae8a69206108f8a18bf20d1ef96668d4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_tensor([9467, 9467])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from eye level, the camera remains perfectly still, focused precisely on a row of glass jelly jars on a supermarket shelf. Under bright, even fluorescent light, the cylindrical jars gleam with a clean sheen, their translucent jelly contents shimmering in jewel-like hues—bright ruby red, vibrant sunshine orange, deep grape purple, and vibrant lime green. Neatly arranged on the pristine white metal shelf, the shelf suddenly experiences a gentle, low-frequency vibration, causing the entire row of jars to tremble in an amusing manner. Rather than violently shaking, they jiggle in a steady, springy motion, while the jelly contents within seem to come alive, dancing, swaying, and swaying in a slightly delayed, comical rhythm. The glass jars clink with subtle, crisp sounds, and light dances across their swaying surfaces, creating dynamic highlights. The entire image exudes a wondrous and playful atmosphere, as if these inanimate objects, amidst an otherwise mundane scene, have been imbued with a fleeting, joyful vitality.", + "seed": 12197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_9467_seed_12197_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": -0.34168585487641395, + "vlm_logprob": -0.34168585487641395, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.34168585487641395, + "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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039383888244628906 + ], + "opposite_logprob_tokens": [ + -4.164383888244629 + ], + "logprob_reward": -0.039383888244628906, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.039383888244628906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16771331429481506, + 0.06397219002246857, + 0.038451820611953735, + 0.046227067708969116, + 0.03082895092666149, + 0.07836057990789413, + 0.19923503696918488, + 0.19576507806777954, + 0.12296058237552643, + 0.05648539960384369 + ], + "credit_map": [ + 0.015114136040210724, + 0.01216958835721016, + 0.024997733533382416, + 0.03580573573708534, + 0.015112052671611309, + 0.01946941949427128, + 0.02811242640018463, + 0.016932211816310883, + 0.008185843005776405, + 0.004910218995064497, + 0.010855727829039097, + 0.01969842053949833, + 0.005403838120400906, + 0.003769373521208763, + 0.007683334872126579, + 0.0034654310438781977, + 0.004151522181928158, + 0.003934643231332302, + 0.006048036273568869, + 0.007623583078384399, + 0.004568258300423622, + 0.002966321771964431, + 0.0062577142380177975, + 0.00290174363180995, + 0.005287104286253452, + 0.00591671047732234, + 0.006742320489138365, + 0.007008624263107777, + 0.005524317733943462, + 0.00518630538135767, + 0.006405384745448828, + 0.00415630079805851, + 0.0023375165183097124, + 0.0032497455831617117, + 0.00419591274112463, + 0.004291274584829807, + 0.004638900980353355, + 0.002733649220317602, + 0.004745768383145332, + 0.0046361833810806274, + 0.019308703020215034, + 0.00804020557552576, + 0.015583612024784088, + 0.008430136367678642, + 0.006313050165772438, + 0.00753785902634263, + 0.005265503190457821, + 0.00788151565939188, + 0.04143349826335907, + 0.028043698519468307, + 0.02616964280605316, + 0.03113607130944729, + 0.01812523417174816, + 0.0180803295224905, + 0.022372474893927574, + 0.013874094001948833, + 0.016875285655260086, + 0.021525483578443527, + 0.011257198639214039, + 0.012665051966905594, + 0.022514384239912033, + 0.024064207449555397, + 0.054571833461523056, + 0.03229163587093353, + 0.01387447863817215, + 0.015733111649751663, + 0.010484224185347557, + 0.017819002270698547, + 0.013541493564844131, + 0.013083688914775848, + 0.020743021741509438, + 0.017681555822491646, + 0.012981866486370564, + 0.00573959294706583, + 0.0061758290976285934, + 0.006418919190764427, + 0.00706122163683176, + 0.00517438305541873, + 0.006336594000458717, + 0.006596991326659918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.039383888244628906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.6006066799163818 + ], + "opposite_logprob_tokens": [ + -0.22560662031173706 + ], + "logprob_reward": -1.6006066799163818, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6006066799163818, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14964056015014648, + 0.04555322229862213, + 0.08412042260169983, + 0.05480135977268219, + 0.2450857162475586, + 0.10325244069099426, + 0.07084890455007553, + 0.04952552542090416, + 0.17680969834327698, + 0.020362215116620064 + ], + "credit_map": [ + 0.019008051604032516, + 0.02129191718995571, + 0.022710833698511124, + 0.028297925367951393, + 0.017086666077375412, + 0.009824090637266636, + 0.019852740690112114, + 0.011568338610231876, + 0.005816163029521704, + 0.006576563697308302, + 0.007835998199880123, + 0.008001366630196571, + 0.0039026776794344187, + 0.00401643430814147, + 0.005366061814129353, + 0.004037961363792419, + 0.010046581737697124, + 0.007221311796456575, + 0.01616642437875271, + 0.017082521691918373, + 0.010755996219813824, + 0.007546223234385252, + 0.009989622980356216, + 0.005311743821948767, + 0.0062319315038621426, + 0.0076772733591496944, + 0.006550971884280443, + 0.008628206327557564, + 0.005932438652962446, + 0.006843114737421274, + 0.007574685383588076, + 0.005362736526876688, + 0.022892653942108154, + 0.015778206288814545, + 0.03493553772568703, + 0.015108463354408741, + 0.01654268428683281, + 0.034406404942274094, + 0.041794661432504654, + 0.06362708657979965, + 0.024950983002781868, + 0.010898342356085777, + 0.015601729042828083, + 0.01067518349736929, + 0.008440482430160046, + 0.010222818702459335, + 0.009624236263334751, + 0.012838666327297688, + 0.010722054168581963, + 0.012817352078855038, + 0.009738530032336712, + 0.012185023166239262, + 0.007639283314347267, + 0.006038425490260124, + 0.008226080797612667, + 0.003482154104858637, + 0.00382427591830492, + 0.004058613907545805, + 0.0046106441877782345, + 0.004742965567857027, + 0.0053671482019126415, + 0.0058761113323271275, + 0.013308585621416569, + 0.007737182546406984, + 0.010746031999588013, + 0.009501456283032894, + 0.013307086192071438, + 0.009478607214987278, + 0.014847156591713428, + 0.018814021721482277, + 0.055011995136737823, + 0.045103345066308975, + 0.004442834760993719, + 0.0029481349047273397, + 0.002636462915688753, + 0.002396149793639779, + 0.0018374163191765547, + 0.001706329290755093, + 0.0018858910771086812, + 0.0025089962873607874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -1.6006066799163818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016117166727781296 + ], + "opposite_logprob_tokens": [ + -4.141117095947266 + ], + "logprob_reward": -0.016117166727781296, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016117166727781296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11467209458351135, + 0.056712742894887924, + 0.06795139610767365, + 0.03692144155502319, + 0.20268630981445312, + 0.07880109548568726, + 0.14154012501239777, + 0.1664651781320572, + 0.09734179824590683, + 0.03690779581665993 + ], + "credit_map": [ + 0.0212056003510952, + 0.01216946728527546, + 0.013795106671750546, + 0.015466137789189816, + 0.01259618904441595, + 0.01412357110530138, + 0.013967897742986679, + 0.011348123662173748, + 0.009051434695720673, + 0.0067319669760763645, + 0.00776006281375885, + 0.0064940424636006355, + 0.005093852523714304, + 0.007966773584485054, + 0.006601961329579353, + 0.007012652698904276, + 0.010553553700447083, + 0.0054006427526474, + 0.011999652720987797, + 0.012169718742370605, + 0.005863350350409746, + 0.008050528354942799, + 0.006691536866128445, + 0.00722241448238492, + 0.0031593679450452328, + 0.0035378937609493732, + 0.004813224542886019, + 0.005097975954413414, + 0.004009269177913666, + 0.006731865927577019, + 0.004042981192469597, + 0.005528861191123724, + 0.020847687497735023, + 0.014848887920379639, + 0.02807515673339367, + 0.014618647284805775, + 0.018246080726385117, + 0.026186848059296608, + 0.03246046602725983, + 0.04740254208445549, + 0.011482887901365757, + 0.00592406839132309, + 0.016245435923337936, + 0.007746316958218813, + 0.004970758222043514, + 0.008999436162412167, + 0.007179115898907185, + 0.016253074631094933, + 0.033080022782087326, + 0.018851038068532944, + 0.019839493557810783, + 0.02173108235001564, + 0.011764745227992535, + 0.011060815304517746, + 0.01504108589142561, + 0.010171833448112011, + 0.01558338850736618, + 0.018901608884334564, + 0.011101369746029377, + 0.013480683788657188, + 0.01853739283978939, + 0.019422467797994614, + 0.04384699463844299, + 0.025591274723410606, + 0.008474759757518768, + 0.009328720159828663, + 0.0077035254798829556, + 0.010472223162651062, + 0.0080120749771595, + 0.010277634486556053, + 0.024236125871539116, + 0.018836740404367447, + 0.0064964983612298965, + 0.0039423368871212006, + 0.003921940457075834, + 0.005041867028921843, + 0.004167432431131601, + 0.004127044230699539, + 0.004753791727125645, + 0.004456881433725357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.016117166727781296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003642236115410924 + ], + "opposite_logprob_tokens": [ + -5.6286420822143555 + ], + "logprob_reward": -0.003642236115410924, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003642236115410924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1875624805688858, + 0.0762215256690979, + 0.07500393688678741, + 0.05186547338962555, + 0.18741047382354736, + 0.05942137539386749, + 0.10600034892559052, + 0.06184583902359009, + 0.14895178377628326, + 0.045716773718595505 + ], + "credit_map": [ + 0.029834654182195663, + 0.02628500387072563, + 0.026311183348298073, + 0.030592532828450203, + 0.021356042474508286, + 0.01472276821732521, + 0.025062283501029015, + 0.013398000039160252, + 0.011730375699698925, + 0.00904008187353611, + 0.010373443365097046, + 0.0094639016315341, + 0.006599482614547014, + 0.00935574434697628, + 0.00960880983620882, + 0.01004969235509634, + 0.007424202747642994, + 0.00658905366435647, + 0.01737409085035324, + 0.016074035316705704, + 0.007856519892811775, + 0.00586704770103097, + 0.008493522182106972, + 0.005325463600456715, + 0.006487028207629919, + 0.007917605340480804, + 0.0067502399906516075, + 0.00801524892449379, + 0.005688253324478865, + 0.0047670574858784676, + 0.007794697303324938, + 0.0044453442096710205, + 0.015152335166931152, + 0.014622118324041367, + 0.02695642225444317, + 0.01687193103134632, + 0.016426431015133858, + 0.024856192991137505, + 0.02727239765226841, + 0.045252636075019836, + 0.00719604454934597, + 0.0038460884243249893, + 0.009401305578649044, + 0.006968979723751545, + 0.005259423051029444, + 0.006839768961071968, + 0.010193483904004097, + 0.009716280736029148, + 0.01575876772403717, + 0.008664273656904697, + 0.016036411747336388, + 0.011721223592758179, + 0.013020944781601429, + 0.009576351381838322, + 0.020279111340641975, + 0.010943267494440079, + 0.004809584002941847, + 0.012333431281149387, + 0.005239903926849365, + 0.01075681671500206, + 0.006171160377562046, + 0.008550465106964111, + 0.006767838727682829, + 0.007216636557132006, + 0.011376825161278248, + 0.01028556190431118, + 0.010958325117826462, + 0.011792520061135292, + 0.013353826478123665, + 0.014361346140503883, + 0.042489346116781235, + 0.03433404490351677, + 0.009305175393819809, + 0.005131884478032589, + 0.0053537702187895775, + 0.005567819811403751, + 0.005741541273891926, + 0.004422889556735754, + 0.004938533995300531, + 0.005255160387605429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.003642236115410924 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048679303377866745 + ], + "opposite_logprob_tokens": [ + -3.0486793518066406 + ], + "logprob_reward": -0.048679303377866745, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048679303377866745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2428545504808426, + 0.04181336611509323, + 0.08266684412956238, + 0.11451037228107452, + 0.108608677983284, + 0.10363035649061203, + 0.11358291655778885, + 0.07905280590057373, + 0.07029194384813309, + 0.04298815503716469 + ], + "credit_map": [ + 0.032086096704006195, + 0.033428966999053955, + 0.02711779996752739, + 0.048871323466300964, + 0.01975417137145996, + 0.02662941999733448, + 0.03215768188238144, + 0.02280910685658455, + 0.006763990502804518, + 0.004556034225970507, + 0.007047969847917557, + 0.005548498127609491, + 0.004591669887304306, + 0.003126879222691059, + 0.0057790386490523815, + 0.004399286117404699, + 0.016264384612441063, + 0.006148479878902435, + 0.012021149508655071, + 0.010902222245931625, + 0.007942602969706059, + 0.009431903250515461, + 0.009388677775859833, + 0.01056742761284113, + 0.011310562491416931, + 0.011713800020515919, + 0.01432905625551939, + 0.017271127551794052, + 0.012949502095580101, + 0.017382796853780746, + 0.016003042459487915, + 0.013550491072237492, + 0.006462707184255123, + 0.0071486495435237885, + 0.015634814277291298, + 0.00985635630786419, + 0.00881248526275158, + 0.01626942865550518, + 0.01814408414065838, + 0.02628014236688614, + 0.01577111892402172, + 0.008042781613767147, + 0.018032178282737732, + 0.012444100342690945, + 0.010701477527618408, + 0.009775198996067047, + 0.010526766069233418, + 0.018336733803153038, + 0.03041413053870201, + 0.01532756071537733, + 0.016342919319868088, + 0.016761964187026024, + 0.01058935932815075, + 0.008352883160114288, + 0.008934528566896915, + 0.006859573535621166, + 0.00644666375592351, + 0.008293483406305313, + 0.006440639961510897, + 0.009018694050610065, + 0.008212472312152386, + 0.01040683127939701, + 0.017680397257208824, + 0.012553628534078598, + 0.007750560529530048, + 0.0071402196772396564, + 0.0042544021271169186, + 0.008011381141841412, + 0.0073455325327813625, + 0.008067663758993149, + 0.014740795828402042, + 0.01298139151185751, + 0.008940203115344048, + 0.004450469743460417, + 0.00535595091059804, + 0.005047865677624941, + 0.00440393527969718, + 0.0049076806753873825, + 0.004732504021376371, + 0.00514954561367631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.048679303377866745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from eye level, the camera remains perfectly still, focused precisely on a row of glass jelly jars on a supermarket shelf. Under bright, even fluorescent light, the cylindrical jars gleam with a clean sheen, their translucent jelly contents shimmering in jewel-like hues—bright ruby red, vibrant sunshine orange, deep grape purple, and vibrant lime green. Neatly arranged on the pristine white metal shelf, the shelf suddenly experiences a gentle, low-frequency vibration, causing the entire row of jars to tremble in an amusing manner. Rather than violently shaking, they jiggle in a steady, springy motion, while the jelly contents within seem to come alive, dancing, swaying, and swaying in a slightly delayed, comical rhythm. The glass jars clink with subtle, crisp sounds, and light dances across their swaying surfaces, creating dynamic highlights. The entire image exudes a wondrous and playful atmosphere, as if these inanimate objects, amidst an otherwise mundane scene, have been imbued with a fleeting, joyful vitality.", + "seed": 12198, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_9467_seed_12198_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": -0.3992384792305529, + "vlm_logprob": -0.3992384792305529, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.3992384792305529, + "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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09135591238737106 + ], + "opposite_logprob_tokens": [ + -3.591355800628662 + ], + "logprob_reward": -0.09135591238737106, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.09135591238737106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19548556208610535, + 0.12907719612121582, + 0.13792602717876434, + 0.038784973323345184, + 0.08944710344076157, + 0.07271604984998703, + 0.08949204534292221, + 0.08237512409687042, + 0.09494967013597488, + 0.06974626332521439 + ], + "credit_map": [ + 0.020430641248822212, + 0.019346032291650772, + 0.026841191574931145, + 0.04055754840373993, + 0.013726504519581795, + 0.015244275331497192, + 0.02527015283703804, + 0.03406919911503792, + 0.016543133184313774, + 0.022923141717910767, + 0.02000686153769493, + 0.01635221391916275, + 0.015994368121027946, + 0.009696204215288162, + 0.017580369487404823, + 0.009980879724025726, + 0.01388067938387394, + 0.020905425772070885, + 0.020311860367655754, + 0.02222369983792305, + 0.01567627303302288, + 0.011424945667386055, + 0.020810337737202644, + 0.01269279234111309, + 0.0032133879140019417, + 0.005335658323019743, + 0.005947188939899206, + 0.007820526137948036, + 0.003999245353043079, + 0.003110842779278755, + 0.005369598511606455, + 0.003988522570580244, + 0.006903765723109245, + 0.0074442788027226925, + 0.01204231008887291, + 0.012399619445204735, + 0.010938230901956558, + 0.011845262721180916, + 0.015957556664943695, + 0.011916072107851505, + 0.009569671005010605, + 0.008450213819742203, + 0.01232595182955265, + 0.012843899428844452, + 0.006897216662764549, + 0.007151030469685793, + 0.008989857509732246, + 0.00648819375783205, + 0.012356458231806755, + 0.010631898418068886, + 0.012834805063903332, + 0.019492004066705704, + 0.0062097301706671715, + 0.006220042239874601, + 0.014553453773260117, + 0.007193649187684059, + 0.008833900094032288, + 0.011359471827745438, + 0.009958996437489986, + 0.015505153685808182, + 0.0073403771966695786, + 0.006316016428172588, + 0.015066133812069893, + 0.007995063439011574, + 0.013877603225409985, + 0.017096124589443207, + 0.012500676326453686, + 0.015983562916517258, + 0.008000032976269722, + 0.005366761237382889, + 0.012202942743897438, + 0.009921960532665253, + 0.008129878900945187, + 0.010383019223809242, + 0.00923804473131895, + 0.010292376391589642, + 0.0044336263090372086, + 0.007037122268229723, + 0.01091059111058712, + 0.009321593679487705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.09135591238737106 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.7015856504440308 + ], + "opposite_logprob_tokens": [ + -0.20158563554286957 + ], + "logprob_reward": -1.7015856504440308, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015856504440308, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17365369200706482, + 0.16524472832679749, + 0.05855106934905052, + 0.09473646432161331, + 0.1735984832048416, + 0.0686674416065216, + 0.10526246577501297, + 0.07184232026338577, + 0.05267413333058357, + 0.03576919808983803 + ], + "credit_map": [ + 0.031904611736536026, + 0.02055777981877327, + 0.022911781445145607, + 0.0281381718814373, + 0.01896996609866619, + 0.014308826997876167, + 0.023662732914090157, + 0.013199825771152973, + 0.027338646352291107, + 0.027742981910705566, + 0.02422518841922283, + 0.026272689923644066, + 0.013981358148157597, + 0.011262045241892338, + 0.02121000550687313, + 0.013211811892688274, + 0.006632446311414242, + 0.009578481316566467, + 0.00823253858834505, + 0.012553063221275806, + 0.005995134823024273, + 0.003346645273268223, + 0.008128692395985126, + 0.004084068816155195, + 0.00766531890258193, + 0.016290390864014626, + 0.01238624844700098, + 0.012416630052030087, + 0.01412244699895382, + 0.010207857936620712, + 0.011500283144414425, + 0.010147293098270893, + 0.013768462464213371, + 0.016449060291051865, + 0.024028709158301353, + 0.032434262335300446, + 0.01956252194941044, + 0.014667356386780739, + 0.03437997028231621, + 0.018308131024241447, + 0.011490649543702602, + 0.00959285069257021, + 0.008540734648704529, + 0.011038255877792835, + 0.007991134189069271, + 0.0074925790540874004, + 0.005404809024184942, + 0.007116427645087242, + 0.010656551457941532, + 0.011091158725321293, + 0.014712944626808167, + 0.027502516284585, + 0.00923993345350027, + 0.004493654239922762, + 0.01944722607731819, + 0.008118478581309319, + 0.008979086764156818, + 0.007085332181304693, + 0.008321905508637428, + 0.01060920674353838, + 0.00482393940910697, + 0.008115062490105629, + 0.015297223813831806, + 0.008610565215349197, + 0.006876481231302023, + 0.006494377739727497, + 0.0065469215624034405, + 0.0069979457184672356, + 0.004411212168633938, + 0.006363338325172663, + 0.007269945461302996, + 0.007713909726589918, + 0.005555544514209032, + 0.006450348068028688, + 0.004230491816997528, + 0.005152724217623472, + 0.0034148001577705145, + 0.003320238785818219, + 0.003471031552180648, + 0.00417401734739542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -1.7015856504440308 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160471573472023 + ], + "opposite_logprob_tokens": [ + -1.9104715585708618 + ], + "logprob_reward": -0.160471573472023, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160471573472023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4164397418498993, + 0.13470664620399475, + 0.1115715503692627, + 0.043054427951574326, + 0.06249183788895607, + 0.07564494013786316, + 0.04244384542107582, + 0.04150085523724556, + 0.023564254865050316, + 0.04858190566301346 + ], + "credit_map": [ + 0.071654312312603, + 0.06210561469197273, + 0.04998330771923065, + 0.0592888742685318, + 0.029745401814579964, + 0.03545580804347992, + 0.05295881628990173, + 0.05524758994579315, + 0.026321282610297203, + 0.019356386736035347, + 0.01964408904314041, + 0.02132374607026577, + 0.009419641457498074, + 0.011558984406292439, + 0.014755114912986755, + 0.0123274065554142, + 0.011828817427158356, + 0.014312944374978542, + 0.01854262501001358, + 0.02647126279771328, + 0.010171589441597462, + 0.007549135945737362, + 0.014764034189283848, + 0.007931137457489967, + 0.0027862144634127617, + 0.003573054913431406, + 0.007192445453256369, + 0.010491077788174152, + 0.0042247409000992775, + 0.003611002117395401, + 0.007547197863459587, + 0.0036286916583776474, + 0.0057242801412940025, + 0.008900093846023083, + 0.008892644196748734, + 0.012329389341175556, + 0.0058281878009438515, + 0.004210808780044317, + 0.01207654271274805, + 0.004529888741672039, + 0.011438201181590557, + 0.009338383562862873, + 0.010186711326241493, + 0.01059078611433506, + 0.006775727029889822, + 0.007629718165844679, + 0.011098410934209824, + 0.008586997166275978, + 0.0040096715092659, + 0.003528064815327525, + 0.006866211537271738, + 0.01000217255204916, + 0.0029030577279627323, + 0.0016971389995887876, + 0.009901761077344418, + 0.0035357715096324682, + 0.004928542301058769, + 0.0042778002098202705, + 0.00532873347401619, + 0.0069853696040809155, + 0.004240692593157291, + 0.003116017673164606, + 0.008990872651338577, + 0.0036328271962702274, + 0.004084083251655102, + 0.0026120806578546762, + 0.003847981570288539, + 0.0031460607424378395, + 0.0020435391925275326, + 0.002816979307681322, + 0.0027578084263950586, + 0.002255719155073166, + 0.004612566903233528, + 0.006605517119169235, + 0.007167714182287455, + 0.008897796273231506, + 0.003224933287128806, + 0.005952340550720692, + 0.00687370914965868, + 0.005247325170785189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.160471573472023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004658204037696123 + ], + "opposite_logprob_tokens": [ + -5.379658222198486 + ], + "logprob_reward": -0.004658204037696123, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004658204037696123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21716903150081635, + 0.0937725156545639, + 0.182076096534729, + 0.08489987999200821, + 0.045487180352211, + 0.062007393687963486, + 0.11184430867433548, + 0.11469783633947372, + 0.03441856801509857, + 0.05362715944647789 + ], + "credit_map": [ + 0.025321239605545998, + 0.021206241101026535, + 0.02745690941810608, + 0.027315275743603706, + 0.01997806504368782, + 0.025851260870695114, + 0.025137020274996758, + 0.04490301385521889, + 0.008590678684413433, + 0.017154991626739502, + 0.013515496626496315, + 0.01411039661616087, + 0.010198083706200123, + 0.007334941532462835, + 0.012774543836712837, + 0.010093380697071552, + 0.019622227177023888, + 0.020752834156155586, + 0.03039412572979927, + 0.050955630838871, + 0.013845481909811497, + 0.009659107774496078, + 0.022424744442105293, + 0.014421946369111538, + 0.008571626618504524, + 0.014034140855073929, + 0.010396962985396385, + 0.01102882344275713, + 0.010846409015357494, + 0.010077676735818386, + 0.010627098381519318, + 0.009317150339484215, + 0.004183207638561726, + 0.004773357883095741, + 0.006044432055205107, + 0.008123168721795082, + 0.006105564534664154, + 0.003830946981906891, + 0.008519759401679039, + 0.003906743135303259, + 0.010667254216969013, + 0.008143053390085697, + 0.007818291895091534, + 0.009369836188852787, + 0.006549544166773558, + 0.006486745551228523, + 0.007289809640496969, + 0.005682851653546095, + 0.010542302392423153, + 0.011122367344796658, + 0.016580943018198013, + 0.027915433049201965, + 0.009276378899812698, + 0.005292485002428293, + 0.022425562143325806, + 0.00868883728981018, + 0.016091370955109596, + 0.013577105477452278, + 0.014761162921786308, + 0.01771945133805275, + 0.011473463848233223, + 0.009242208674550056, + 0.017908355221152306, + 0.013924725353717804, + 0.0070540946908295155, + 0.006524678319692612, + 0.00465433718636632, + 0.004788792226463556, + 0.0030941064469516277, + 0.0021553889382630587, + 0.0033688528928905725, + 0.0027783168479800224, + 0.009603208862245083, + 0.007065494544804096, + 0.00832389947026968, + 0.008516418747603893, + 0.0044525302946567535, + 0.003767306450754404, + 0.008038689382374287, + 0.003859611228108406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.004658204037696123 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381210558116436 + ], + "opposite_logprob_tokens": [ + -3.288120985031128 + ], + "logprob_reward": -0.0381210558116436, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381210558116436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21738503873348236, + 0.16966968774795532, + 0.07601193338632584, + 0.15092410147190094, + 0.057215914130210876, + 0.06773531436920166, + 0.11480743438005447, + 0.03845907002687454, + 0.06431753933429718, + 0.04347389563918114 + ], + "credit_map": [ + 0.04416855797171593, + 0.02665264531970024, + 0.02852657623589039, + 0.025404512882232666, + 0.024776503443717957, + 0.022774407640099525, + 0.024019869044423103, + 0.021061968058347702, + 0.018156927078962326, + 0.03131839260458946, + 0.024221014231443405, + 0.028525644913315773, + 0.014486029744148254, + 0.013515117578208447, + 0.020438391715288162, + 0.019008157774806023, + 0.013616771437227726, + 0.009477120824158192, + 0.01268483605235815, + 0.009738096036016941, + 0.008387153036892414, + 0.007099770940840244, + 0.008288698270916939, + 0.00671948678791523, + 0.013526967726647854, + 0.020255940034985542, + 0.019954994320869446, + 0.02069864422082901, + 0.022369621321558952, + 0.017928248271346092, + 0.018542703241109848, + 0.017646964639425278, + 0.005992074962705374, + 0.0065269675105810165, + 0.00852338969707489, + 0.011251519434154034, + 0.005640024319291115, + 0.0049257599748671055, + 0.00917425099760294, + 0.005181930959224701, + 0.009454526007175446, + 0.008188764564692974, + 0.009365368634462357, + 0.011482813395559788, + 0.006637671496719122, + 0.0064276657067239285, + 0.008832764811813831, + 0.007345744874328375, + 0.011806327849626541, + 0.013242866843938828, + 0.01585272327065468, + 0.0284232497215271, + 0.00909337680786848, + 0.005189326126128435, + 0.023299727588891983, + 0.007899823598563671, + 0.007911729626357555, + 0.005246100947260857, + 0.005766377318650484, + 0.005048554390668869, + 0.0034608275163918734, + 0.003124774433672428, + 0.004668240901082754, + 0.0032324683852493763, + 0.007372473366558552, + 0.008671604096889496, + 0.009156673215329647, + 0.00968380831182003, + 0.007175073493272066, + 0.005889474879950285, + 0.011216560378670692, + 0.0051518757827579975, + 0.005626068916171789, + 0.006748486310243607, + 0.0060814921744167805, + 0.007015332113951445, + 0.0031336965039372444, + 0.003643751610070467, + 0.006777419708669186, + 0.004447649233043194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0381210558116436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from eye level, the camera remains perfectly still, focused precisely on a row of glass jelly jars on a supermarket shelf. Under bright, even fluorescent light, the cylindrical jars gleam with a clean sheen, their translucent jelly contents shimmering in jewel-like hues—bright ruby red, vibrant sunshine orange, deep grape purple, and vibrant lime green. Neatly arranged on the pristine white metal shelf, the shelf suddenly experiences a gentle, low-frequency vibration, causing the entire row of jars to tremble in an amusing manner. Rather than violently shaking, they jiggle in a steady, springy motion, while the jelly contents within seem to come alive, dancing, swaying, and swaying in a slightly delayed, comical rhythm. The glass jars clink with subtle, crisp sounds, and light dances across their swaying surfaces, creating dynamic highlights. The entire image exudes a wondrous and playful atmosphere, as if these inanimate objects, amidst an otherwise mundane scene, have been imbued with a fleeting, joyful vitality.", + "seed": 12199, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_9467_seed_12199_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": -0.19723742676433176, + "vlm_logprob": -0.19723742676433176, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.19723742676433176, + "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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031671639531850815 + ], + "opposite_logprob_tokens": [ + -4.531671524047852 + ], + "logprob_reward": -0.031671639531850815, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.031671639531850815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3466251790523529, + 0.0676795095205307, + 0.1172049418091774, + 0.0699128806591034, + 0.06421113759279251, + 0.07717106491327286, + 0.07975572347640991, + 0.040676504373550415, + 0.07993924617767334, + 0.05682380497455597 + ], + "credit_map": [ + 0.028091685846447945, + 0.03778912127017975, + 0.07237353175878525, + 0.05750298872590065, + 0.033959902822971344, + 0.026620008051395416, + 0.034460313618183136, + 0.05582761764526367, + 0.0058487774804234505, + 0.010082022286951542, + 0.008563654497265816, + 0.008604591712355614, + 0.010168715380132198, + 0.005139682907611132, + 0.009328065440058708, + 0.009943992830812931, + 0.007088697515428066, + 0.011552931740880013, + 0.02773444354534149, + 0.014324219897389412, + 0.009274694137275219, + 0.008035451173782349, + 0.012544423341751099, + 0.026650061830878258, + 0.00774475559592247, + 0.007897373288869858, + 0.007875092327594757, + 0.010953162796795368, + 0.007049794774502516, + 0.01182190515100956, + 0.0070762732066214085, + 0.009494515135884285, + 0.005464582238346338, + 0.006073806434869766, + 0.007225072011351585, + 0.008927308954298496, + 0.005450222175568342, + 0.009196179918944836, + 0.008949347771704197, + 0.012924611568450928, + 0.006926779635250568, + 0.006187894847244024, + 0.01099399384111166, + 0.010063462890684605, + 0.006345138885080814, + 0.008527372032403946, + 0.01234497968107462, + 0.01578143984079361, + 0.008809411898255348, + 0.011404530145227909, + 0.013148115016520023, + 0.011237476952373981, + 0.005187325645238161, + 0.010203085839748383, + 0.008048121817409992, + 0.01171764824539423, + 0.006292800884693861, + 0.004422847647219896, + 0.006003298796713352, + 0.006898918654769659, + 0.0029639839194715023, + 0.0038078781217336655, + 0.003880927339196205, + 0.006405843421816826, + 0.006025754846632481, + 0.008055075071752071, + 0.015114347450435162, + 0.011736704036593437, + 0.00653317105025053, + 0.0077642714604735374, + 0.013533554039895535, + 0.01117636263370514, + 0.008273868821561337, + 0.00780841801315546, + 0.0072944024577736855, + 0.008690941147506237, + 0.004648164846003056, + 0.005196987651288509, + 0.007280467078089714, + 0.007630551233887672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.031671639531850815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8983903527259827 + ], + "opposite_logprob_tokens": [ + -0.5233903527259827 + ], + "logprob_reward": -0.8983903527259827, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983903527259827, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3515298068523407, + 0.1763213872909546, + 0.0945901945233345, + 0.10092920064926147, + 0.06182819604873657, + 0.07581229507923126, + 0.03875749558210373, + 0.03227327764034271, + 0.031030874699354172, + 0.03692732751369476 + ], + "credit_map": [ + 0.03226282820105553, + 0.04738449305295944, + 0.061826154589653015, + 0.04492581635713577, + 0.045170292258262634, + 0.02418772503733635, + 0.034590691328048706, + 0.06118176877498627, + 0.012327266857028008, + 0.00935666635632515, + 0.05238530784845352, + 0.020127471536397934, + 0.00946805439889431, + 0.00938524678349495, + 0.012485438957810402, + 0.05078592896461487, + 0.007731768302619457, + 0.01626620441675186, + 0.014794208109378815, + 0.014682337641716003, + 0.006696013733744621, + 0.010220127180218697, + 0.007594074122607708, + 0.016605444252490997, + 0.0074299718253314495, + 0.009920332580804825, + 0.015577694401144981, + 0.017747746780514717, + 0.008499175310134888, + 0.0166641753166914, + 0.008057382889091969, + 0.017032712697982788, + 0.006761371158063412, + 0.004119842313230038, + 0.01647002622485161, + 0.00892406702041626, + 0.003750184550881386, + 0.004794172011315823, + 0.006848377175629139, + 0.010160152800381184, + 0.005120248068124056, + 0.00759228365495801, + 0.021112110465765, + 0.011204794980585575, + 0.00496742594987154, + 0.005518087185919285, + 0.005832033231854439, + 0.014465298503637314, + 0.006137341260910034, + 0.006291981320828199, + 0.006048950832337141, + 0.006214262917637825, + 0.002714097499847412, + 0.0046202945522964, + 0.0030194949358701706, + 0.003711068071424961, + 0.0027823878917843103, + 0.005080951377749443, + 0.005251884460449219, + 0.005347642116248608, + 0.002910243347287178, + 0.004034944344311953, + 0.0031969163101166487, + 0.003668305929750204, + 0.0025708922185003757, + 0.002446817234158516, + 0.006401835009455681, + 0.004668544512242079, + 0.0022803638130426407, + 0.00450508389621973, + 0.0038139522075653076, + 0.0043433839455246925, + 0.0033565463963896036, + 0.0029596949461847544, + 0.010032343678176403, + 0.007993869483470917, + 0.00214935839176178, + 0.0029695327393710613, + 0.0025840969756245613, + 0.004881878383457661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.8983903527259827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099145747721195 + ], + "opposite_logprob_tokens": [ + -4.511099338531494 + ], + "logprob_reward": -0.011099145747721195, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099145747721195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46495118737220764, + 0.12017755210399628, + 0.10403579473495483, + 0.04896184802055359, + 0.040668584406375885, + 0.0439891554415226, + 0.04786539077758789, + 0.030545702204108238, + 0.04627908766269684, + 0.05252572149038315 + ], + "credit_map": [ + 0.04808678478002548, + 0.07191598415374756, + 0.061644215136766434, + 0.0760052502155304, + 0.05613790452480316, + 0.031011121347546577, + 0.05111505463719368, + 0.06903482973575592, + 0.009492621757090092, + 0.010124336928129196, + 0.031053559854626656, + 0.01741478592157364, + 0.011518662795424461, + 0.006759434938430786, + 0.00967316422611475, + 0.024140983819961548, + 0.007401496637612581, + 0.00956681091338396, + 0.027946410700678825, + 0.011724166572093964, + 0.007267410401254892, + 0.007550728041678667, + 0.00911552831530571, + 0.023463230580091476, + 0.0055179051123559475, + 0.004336232785135508, + 0.00831840094178915, + 0.005907564889639616, + 0.004719733260571957, + 0.008234195411205292, + 0.005667283199727535, + 0.006260529160499573, + 0.0034578333143144846, + 0.003938878886401653, + 0.0062224287539720535, + 0.005993301514536142, + 0.0034440739545971155, + 0.005375630687922239, + 0.006902920547872782, + 0.005333513021469116, + 0.009424381889402866, + 0.0058944206684827805, + 0.006261697970330715, + 0.005326990038156509, + 0.003377234796062112, + 0.004793178755789995, + 0.003450976451858878, + 0.005460269283503294, + 0.006452945061028004, + 0.00595227861776948, + 0.008903812617063522, + 0.00625570397824049, + 0.0033316027838736773, + 0.0029008151032030582, + 0.005413949489593506, + 0.008654280565679073, + 0.0027741361409425735, + 0.003688405267894268, + 0.004762264434248209, + 0.005086100660264492, + 0.002306907204911113, + 0.002643723739311099, + 0.003811111208051443, + 0.00547304842621088, + 0.0035473343450576067, + 0.003560068551450968, + 0.007950390689074993, + 0.005405296105891466, + 0.0037648596335202456, + 0.007810818497091532, + 0.006197238340973854, + 0.008043074980378151, + 0.005619590170681477, + 0.004255100153386593, + 0.010897853411734104, + 0.009725801646709442, + 0.0032988786697387695, + 0.00532264169305563, + 0.006292594596743584, + 0.00711325416341424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011099145747721195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019530047429725528 + ], + "opposite_logprob_tokens": [ + -6.251953125 + ], + "logprob_reward": -0.0019530047429725528, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019530047429725528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34540510177612305, + 0.18078981339931488, + 0.08794578164815903, + 0.08804464340209961, + 0.07853840291500092, + 0.07044274359941483, + 0.04233572632074356, + 0.03491848334670067, + 0.02712075784802437, + 0.044458549469709396 + ], + "credit_map": [ + 0.035901736468076706, + 0.040351249277591705, + 0.04523862153291702, + 0.06487513333559036, + 0.040075477212667465, + 0.03497578576207161, + 0.04040505737066269, + 0.043582070618867874, + 0.016051387414336205, + 0.018260976299643517, + 0.04010102525353432, + 0.021003995090723038, + 0.012588848359882832, + 0.010283750481903553, + 0.021366087719798088, + 0.04113375023007393, + 0.0059999399818480015, + 0.003951597027480602, + 0.0235524233430624, + 0.010265486314892769, + 0.006818725261837244, + 0.006793622858822346, + 0.008336218073964119, + 0.022227775305509567, + 0.012419898062944412, + 0.010591907426714897, + 0.009377827867865562, + 0.009762819856405258, + 0.008563373237848282, + 0.01587352156639099, + 0.0105514507740736, + 0.01090384740382433, + 0.006923388689756393, + 0.004449771251529455, + 0.021672386676073074, + 0.010440309531986713, + 0.004290458746254444, + 0.005969259422272444, + 0.006946943234652281, + 0.017845887690782547, + 0.010989378206431866, + 0.006488812621682882, + 0.011097433045506477, + 0.008516988717019558, + 0.005615762900561094, + 0.009888073429465294, + 0.008773189969360828, + 0.009073106572031975, + 0.005822156555950642, + 0.005069998558610678, + 0.0073480624705553055, + 0.006351540796458721, + 0.0033191645052284002, + 0.003674060106277466, + 0.0050039226189255714, + 0.00574682280421257, + 0.003966695629060268, + 0.003654325380921364, + 0.006234090309590101, + 0.006063822191208601, + 0.002512982813641429, + 0.003808800131082535, + 0.0036818869411945343, + 0.004995881114155054, + 0.0018279396463185549, + 0.002464313292875886, + 0.005376212298870087, + 0.004425423219799995, + 0.0020466626156121492, + 0.003428815398365259, + 0.003983203321695328, + 0.003568191546946764, + 0.003452511038631201, + 0.0036694156005978584, + 0.010462040081620216, + 0.009347268380224705, + 0.0023891322780400515, + 0.003270306158810854, + 0.005989490542560816, + 0.005878389347344637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0019530047429725528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307299107313156 + ], + "opposite_logprob_tokens": [ + -3.1680729389190674 + ], + "logprob_reward": -0.04307299107313156, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307299107313156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45317384600639343, + 0.13881520926952362, + 0.0677788034081459, + 0.06479663401842117, + 0.04192567989230156, + 0.04937699809670448, + 0.04098337143659592, + 0.04844202846288681, + 0.05622777342796326, + 0.03847967088222504 + ], + "credit_map": [ + 0.04026227444410324, + 0.033762093633413315, + 0.10309205949306488, + 0.05825445428490639, + 0.05108058825135231, + 0.029656480997800827, + 0.04284338653087616, + 0.09422250092029572, + 0.011213921010494232, + 0.01766273006796837, + 0.027210591360926628, + 0.015404339879751205, + 0.016552424058318138, + 0.010253282263875008, + 0.011382619850337505, + 0.02913529984652996, + 0.005431333556771278, + 0.009981066919863224, + 0.014088046737015247, + 0.009926613420248032, + 0.005934660788625479, + 0.005600780714303255, + 0.005154716316610575, + 0.011661585420370102, + 0.006017385981976986, + 0.0063067167066037655, + 0.011755309998989105, + 0.012994250282645226, + 0.004868245217949152, + 0.006355782505124807, + 0.006909193936735392, + 0.009589754976332188, + 0.004658317659050226, + 0.004003865644335747, + 0.006418398581445217, + 0.008920733816921711, + 0.00278150150552392, + 0.004736647941172123, + 0.0043868618085980415, + 0.006019351538270712, + 0.004999550525099039, + 0.0053475177846848965, + 0.010413836687803268, + 0.007143008057028055, + 0.004978000186383724, + 0.003748980350792408, + 0.004399054683744907, + 0.008347048424184322, + 0.002894448582082987, + 0.004728624131530523, + 0.008985703811049461, + 0.006483781151473522, + 0.003095362801104784, + 0.0050153047777712345, + 0.0029909603763371706, + 0.006789186038076878, + 0.0042594317346811295, + 0.0056819855235517025, + 0.010263080708682537, + 0.008149881847202778, + 0.003735897596925497, + 0.003981195390224457, + 0.006405472755432129, + 0.005965082440525293, + 0.005007496103644371, + 0.007378414273262024, + 0.01040960755199194, + 0.011259367689490318, + 0.004317831713706255, + 0.005490340292453766, + 0.006472340319305658, + 0.005892377812415361, + 0.003695115214213729, + 0.003351387567818165, + 0.009053871966898441, + 0.008509158156812191, + 0.002263149945065379, + 0.0029979918617755175, + 0.003756537102162838, + 0.004852461162954569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04307299107313156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from eye level, the camera remains perfectly still, focused precisely on a row of glass jelly jars on a supermarket shelf. Under bright, even fluorescent light, the cylindrical jars gleam with a clean sheen, their translucent jelly contents shimmering in jewel-like hues—bright ruby red, vibrant sunshine orange, deep grape purple, and vibrant lime green. Neatly arranged on the pristine white metal shelf, the shelf suddenly experiences a gentle, low-frequency vibration, causing the entire row of jars to tremble in an amusing manner. Rather than violently shaking, they jiggle in a steady, springy motion, while the jelly contents within seem to come alive, dancing, swaying, and swaying in a slightly delayed, comical rhythm. The glass jars clink with subtle, crisp sounds, and light dances across their swaying surfaces, creating dynamic highlights. The entire image exudes a wondrous and playful atmosphere, as if these inanimate objects, amidst an otherwise mundane scene, have been imbued with a fleeting, joyful vitality.", + "seed": 12200, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_9467_seed_12200_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": -0.23618760267272593, + "vlm_logprob": -0.23618760267272593, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.23618760267272593, + "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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044562775641679764 + ], + "opposite_logprob_tokens": [ + -4.044562816619873 + ], + "logprob_reward": -0.044562775641679764, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.044562775641679764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.095465749502182, + 0.057896509766578674, + 0.3909990191459656, + 0.04445900022983551, + 0.08671893179416656, + 0.14057035744190216, + 0.09004175662994385, + 0.03188866376876831, + 0.04050213843584061, + 0.02145792357623577 + ], + "credit_map": [ + 0.008218619972467422, + 0.011299753561615944, + 0.014795572496950626, + 0.011479995213449001, + 0.01045592874288559, + 0.009749331511557102, + 0.01363640557974577, + 0.01583014242351055, + 0.0041958848014473915, + 0.005919843912124634, + 0.009598283097147942, + 0.007644142024219036, + 0.0027072830125689507, + 0.003209077985957265, + 0.0173722505569458, + 0.007249742746353149, + 0.018116021528840065, + 0.028219616040587425, + 0.12068693339824677, + 0.024215031415224075, + 0.00999924074858427, + 0.004646352492272854, + 0.1673882007598877, + 0.017727604135870934, + 0.004425326827913523, + 0.004434444010257721, + 0.006294550374150276, + 0.005957977380603552, + 0.0038342175539582968, + 0.004164376761764288, + 0.0069633410312235355, + 0.008384764194488525, + 0.0073434943333268166, + 0.00990955252200365, + 0.00928796362131834, + 0.010195615701377392, + 0.008652234449982643, + 0.010101770050823689, + 0.011090888641774654, + 0.020137406885623932, + 0.01214224100112915, + 0.0157712921500206, + 0.015045681037008762, + 0.014716639183461666, + 0.017944158986210823, + 0.016613002866506577, + 0.019810331985354424, + 0.02852700464427471, + 0.008477497845888138, + 0.009853369556367397, + 0.013570958748459816, + 0.011630973778665066, + 0.006651102565228939, + 0.007757515646517277, + 0.011907543055713177, + 0.020192787051200867, + 0.0031511755660176277, + 0.0035919565707445145, + 0.004583856090903282, + 0.00539941294118762, + 0.003585624508559704, + 0.0020280880853533745, + 0.005195016041398048, + 0.00435353210195899, + 0.0033485747408121824, + 0.0044421665370464325, + 0.0057160817086696625, + 0.006505316123366356, + 0.005520843900740147, + 0.002742413431406021, + 0.005953092128038406, + 0.0062736463733017445, + 0.0019613264594227076, + 0.0026890940498560667, + 0.003595365211367607, + 0.0025399639271199703, + 0.002147183520719409, + 0.0031007551588118076, + 0.00250915065407753, + 0.0029150808695703745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.044562775641679764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0539579391479492 + ], + "opposite_logprob_tokens": [ + -0.4289579391479492 + ], + "logprob_reward": -1.0539579391479492, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539579391479492, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11847047507762909, + 0.08787597715854645, + 0.372446209192276, + 0.07072027027606964, + 0.12702691555023193, + 0.06734982132911682, + 0.033241983503103256, + 0.04739080369472504, + 0.0383208803832531, + 0.03715664893388748 + ], + "credit_map": [ + 0.009114066138863564, + 0.010923493653535843, + 0.0178206916898489, + 0.015138167887926102, + 0.00975281372666359, + 0.013916634023189545, + 0.016472570598125458, + 0.02533203922212124, + 0.0067225792445242405, + 0.009038359858095646, + 0.015826083719730377, + 0.011435502208769321, + 0.005462726578116417, + 0.004078544210642576, + 0.02728366106748581, + 0.008028521202504635, + 0.017530450597405434, + 0.027832986786961555, + 0.1107560396194458, + 0.025820104405283928, + 0.01087182480841875, + 0.007570665795356035, + 0.1560247391462326, + 0.016039414331316948, + 0.00629435246810317, + 0.008823882788419724, + 0.01295938529074192, + 0.009210151620209217, + 0.005844856612384319, + 0.005365246906876564, + 0.01383502222597599, + 0.008387366309762001, + 0.0106873894110322, + 0.01497007068246603, + 0.024133313447237015, + 0.02201722003519535, + 0.008902824483811855, + 0.011625226587057114, + 0.017490733414888382, + 0.01720014028251171, + 0.007832099683582783, + 0.011746266856789589, + 0.010518231429159641, + 0.009304202161729336, + 0.00565328961238265, + 0.003413086524233222, + 0.01159158069640398, + 0.007291064132004976, + 0.004193695727735758, + 0.004233071580529213, + 0.005242710001766682, + 0.006143009290099144, + 0.0023569187615066767, + 0.0016482799546793103, + 0.006453040521591902, + 0.002971257083117962, + 0.0037961031775921583, + 0.005616314243525267, + 0.0067708478309214115, + 0.00965515710413456, + 0.006633140612393618, + 0.002748904749751091, + 0.008130471222102642, + 0.004039863124489784, + 0.0047708419151604176, + 0.0047069089487195015, + 0.005232499446719885, + 0.005786551628261805, + 0.006595477927476168, + 0.0022290160413831472, + 0.0057015977799892426, + 0.003297986462712288, + 0.0042055812664330006, + 0.0048509398475289345, + 0.004912746604532003, + 0.004048492759466171, + 0.0045238337479531765, + 0.002777683548629284, + 0.005319391377270222, + 0.006517978850752115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -1.0539579391479492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332417666912079 + ], + "opposite_logprob_tokens": [ + -3.7733242511749268 + ], + "logprob_reward": -0.02332417666912079, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332417666912079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16398723423480988, + 0.12527815997600555, + 0.24745599925518036, + 0.05393383651971817, + 0.056847572326660156, + 0.13934354484081268, + 0.10176625102758408, + 0.055122010409832, + 0.029639899730682373, + 0.02662549912929535 + ], + "credit_map": [ + 0.014027384109795094, + 0.021217046305537224, + 0.024361377581954002, + 0.018013672903180122, + 0.013715192675590515, + 0.016044871881604195, + 0.03362831473350525, + 0.02297937497496605, + 0.009158647619187832, + 0.012132206931710243, + 0.0190594382584095, + 0.017866041511297226, + 0.011347595602273941, + 0.007610772270709276, + 0.03134046867489815, + 0.01676298677921295, + 0.012092194519937038, + 0.01750103384256363, + 0.07956946641206741, + 0.014765051193535328, + 0.006423287093639374, + 0.005596376024186611, + 0.09971978515386581, + 0.011788807809352875, + 0.004012300632894039, + 0.0060376254841685295, + 0.0070388298481702805, + 0.005600995849817991, + 0.005007951054722071, + 0.004842400085180998, + 0.008716562762856483, + 0.012677174992859364, + 0.005291569046676159, + 0.00715071614831686, + 0.012604842893779278, + 0.007921932265162468, + 0.0042615607380867004, + 0.005587025545537472, + 0.007235911674797535, + 0.006794012151658535, + 0.01274283230304718, + 0.015194790437817574, + 0.02149948664009571, + 0.01652088761329651, + 0.01752476766705513, + 0.015123475342988968, + 0.019710775464773178, + 0.02102653682231903, + 0.011544196866452694, + 0.013054469600319862, + 0.014386400580406189, + 0.014222472906112671, + 0.00879814475774765, + 0.007448799442499876, + 0.013727680779993534, + 0.01858408935368061, + 0.003941775299608707, + 0.00703837163746357, + 0.00857362151145935, + 0.01207590289413929, + 0.0061646513640880585, + 0.0029406154062598944, + 0.008582926355302334, + 0.005804144311696291, + 0.003269418841227889, + 0.0037716247607022524, + 0.0035763231571763754, + 0.00552326301112771, + 0.0036182699259370565, + 0.001679726759903133, + 0.004183310084044933, + 0.004017961211502552, + 0.0022243924904614687, + 0.002803585259243846, + 0.004031426273286343, + 0.0044691599905490875, + 0.002471972955390811, + 0.002116990741342306, + 0.004726080223917961, + 0.003781890030950308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.02332417666912079 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004115920979529619 + ], + "opposite_logprob_tokens": [ + -5.504116058349609 + ], + "logprob_reward": -0.004115920979529619, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004115920979529619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1639176607131958, + 0.13933128118515015, + 0.28828316926956177, + 0.0889715701341629, + 0.05739666149020195, + 0.04357580840587616, + 0.04713764041662216, + 0.03351045772433281, + 0.04301053658127785, + 0.09486515820026398 + ], + "credit_map": [ + 0.01353511307388544, + 0.018861429765820503, + 0.023001182824373245, + 0.02590109221637249, + 0.014025280252099037, + 0.015764445066452026, + 0.02580060437321663, + 0.0270285215228796, + 0.01353969145566225, + 0.016475260257720947, + 0.02532116323709488, + 0.02603851445019245, + 0.012740641832351685, + 0.0070845456793904305, + 0.02478468418121338, + 0.013346773572266102, + 0.013105673715472221, + 0.021219350397586823, + 0.08588942885398865, + 0.0180220864713192, + 0.009401609189808369, + 0.006603121757507324, + 0.12044800817966461, + 0.013593918643891811, + 0.008003564551472664, + 0.009738583117723465, + 0.02350137196481228, + 0.011478720232844353, + 0.00481446785852313, + 0.005594122223556042, + 0.013490610755980015, + 0.012350130826234818, + 0.004365043248981237, + 0.0069947149604558945, + 0.012430657632648945, + 0.00968743022531271, + 0.003865084145218134, + 0.004330649971961975, + 0.007598462514579296, + 0.008124614134430885, + 0.0038395544979721308, + 0.00565717788413167, + 0.005549614317715168, + 0.004605842288583517, + 0.005745380651205778, + 0.005719261709600687, + 0.007044261321425438, + 0.005414717830717564, + 0.0030400182586163282, + 0.004102242179214954, + 0.009766793809831142, + 0.005977225955575705, + 0.00314907543361187, + 0.0036371194291859865, + 0.006638022605329752, + 0.01082714181393385, + 0.0029226671904325485, + 0.005001472774893045, + 0.004854531027376652, + 0.006621606647968292, + 0.0040252297185361385, + 0.0017455954803153872, + 0.005370059981942177, + 0.00296929688192904, + 0.0036145371850579977, + 0.005098973400890827, + 0.0055910819210112095, + 0.0057447925209999084, + 0.005456076934933662, + 0.0029036463238298893, + 0.00696623744443059, + 0.007635188288986683, + 0.009743615984916687, + 0.0156555138528347, + 0.012599092908203602, + 0.012942063622176647, + 0.010232258588075638, + 0.006532837636768818, + 0.016412001103162766, + 0.010747764259576797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004115920979529619 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 row of glass jelly jars on a supermarket shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still at eye level without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the jars illuminated by bright, even fluorescent light showing a clean sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jars contain translucent jelly in ruby red, orange, purple, and lime green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jars and their contents jiggle in a steady, springy, comical motion due to a gentle shelf vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497720092535019 + ], + "opposite_logprob_tokens": [ + -2.9299771785736084 + ], + "logprob_reward": -0.05497720092535019, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497720092535019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11659029126167297, + 0.1854623705148697, + 0.1909303218126297, + 0.07063057273626328, + 0.1476399004459381, + 0.11016599833965302, + 0.048516564071178436, + 0.039149295538663864, + 0.04876944050192833, + 0.0421452559530735 + ], + "credit_map": [ + 0.008943620137870312, + 0.013228045776486397, + 0.016106508672237396, + 0.016319405287504196, + 0.013067097403109074, + 0.0164865180850029, + 0.017925959080457687, + 0.014513139612972736, + 0.018444163724780083, + 0.018398329615592957, + 0.028782591223716736, + 0.02026188187301159, + 0.017055569216609, + 0.008669632486999035, + 0.05869525671005249, + 0.015154932625591755, + 0.009179689921438694, + 0.013633711263537407, + 0.06034565716981888, + 0.01212799921631813, + 0.004909793846309185, + 0.0034792497754096985, + 0.07676028460264206, + 0.010493933223187923, + 0.007149239536374807, + 0.009307777509093285, + 0.009720640257000923, + 0.00793951004743576, + 0.008041299879550934, + 0.007144155912101269, + 0.008576144464313984, + 0.012751804664731026, + 0.014637697488069534, + 0.01693759299814701, + 0.024277130141854286, + 0.02885354496538639, + 0.01045215129852295, + 0.008871219120919704, + 0.029774120077490807, + 0.013836435042321682, + 0.007821345701813698, + 0.014503991231322289, + 0.028029726818203926, + 0.01708015613257885, + 0.00868978165090084, + 0.007733000908046961, + 0.011876003816723824, + 0.014431997202336788, + 0.0038244081661105156, + 0.0045082056894898415, + 0.007192439399659634, + 0.006148320157080889, + 0.0024740672670304775, + 0.004134520888328552, + 0.006571611389517784, + 0.013662990182638168, + 0.002639689715579152, + 0.004273861180990934, + 0.0052008614875376225, + 0.006733437068760395, + 0.0037316579837352037, + 0.002865847898647189, + 0.007902669720351696, + 0.005801270715892315, + 0.005292900837957859, + 0.006162632256746292, + 0.007191953249275684, + 0.008173609152436256, + 0.005040776915848255, + 0.0033705595415085554, + 0.007736008148640394, + 0.005800999701023102, + 0.003639685222879052, + 0.00425815861672163, + 0.005986334290355444, + 0.005806364584714174, + 0.003956332802772522, + 0.0043286425061523914, + 0.007642808835953474, + 0.006526929792016745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.05497720092535019 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_14/index_tensor([501, 501])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_14/index_tensor([501, 501])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e24bff3d2e4973f8fbe57adfb8645980cae1853e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_14/index_tensor([501, 501])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a remarkably steady first-person perspective, the camera pans in a slow, smooth clockwise orbit, circling an old-fashioned rattan chair, resting in the center of a vast, empty room. A beam of warm afternoon sunlight slants in from an off-screen window, casting a soft golden halo around the chair's outline and long, sharply defined shadows across the polished wooden floor. Airborne dust particles are clearly visible in the beam. As the camera pans, light flows over every surface of the rattan chair, vividly highlighting its age-worn texture with hyper-realistic rendering. Some of the rattan strips on the back are worn and broken, revealing their coarse fibers. The armrests, once sleek and glossy from years of wear, have lost their original varnish, revealing a lighter wood tone beneath. A fine layer of dust accumulates in the seams of the seat's weaving, visible in the light. The whole picture presents an extreme realism style. The texture of the rattan and the grain of the wood seem to be within reach, creating a quiet, nostalgic and highly contemplative atmosphere.", + "seed": 19687, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_14/index_501_seed_19687_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": -0.2696475928649306, + "vlm_logprob": -0.2696475928649306, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.2696475928649306, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2535516023635864 + ], + "opposite_logprob_tokens": [ + -1.6285516023635864 + ], + "logprob_reward": -0.2535516023635864, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2535516023635864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21083252131938934, + 0.07828817516565323, + 0.06822343170642853, + 0.09657703340053558, + 0.09164483845233917, + 0.049042709171772, + 0.12292074412107468, + 0.1480158567428589, + 0.0883115604519844, + 0.04614318162202835 + ], + "credit_map": [ + 0.020748062059283257, + 0.044706862419843674, + 0.03676533326506615, + 0.024809038266539574, + 0.017651187255978584, + 0.026792144402861595, + 0.022376051172614098, + 0.016983838751912117, + 0.009491308592259884, + 0.011679746210575104, + 0.008405204862356186, + 0.009303957223892212, + 0.012839133851230145, + 0.007863164879381657, + 0.009222430177032948, + 0.009483237750828266, + 0.007044743746519089, + 0.004987002816051245, + 0.011862926185131073, + 0.003979853820055723, + 0.005189947318285704, + 0.009880047291517258, + 0.007163008209317923, + 0.018115904182195663, + 0.013335205614566803, + 0.014713921584188938, + 0.01461261697113514, + 0.006207966711372137, + 0.008499851450324059, + 0.015947962179780006, + 0.0077941277995705605, + 0.015465379692614079, + 0.01389866042882204, + 0.015425711870193481, + 0.01636294461786747, + 0.007778210565447807, + 0.00649194186553359, + 0.014056436717510223, + 0.007710434962064028, + 0.009920487180352211, + 0.006740668322890997, + 0.006405292544513941, + 0.006574741564691067, + 0.004037258215248585, + 0.006348523776978254, + 0.007468144875019789, + 0.004390187561511993, + 0.0070778909139335155, + 0.012741616927087307, + 0.017281055450439453, + 0.016436507925391197, + 0.012971808202564716, + 0.011697320267558098, + 0.011547700501978397, + 0.020596351474523544, + 0.01964837871491909, + 0.02290414087474346, + 0.026865437626838684, + 0.020703520625829697, + 0.013403641059994698, + 0.017806880176067352, + 0.01412754226475954, + 0.016105718910694122, + 0.016098974272608757, + 0.010527140460908413, + 0.014773314818739891, + 0.008218780159950256, + 0.013482554815709591, + 0.012068995274603367, + 0.007978479377925396, + 0.012577265501022339, + 0.008685034699738026, + 0.006144074257463217, + 0.005995286628603935, + 0.0043238927610218525, + 0.0067090620286762714, + 0.006896717939525843, + 0.006635338068008423, + 0.005467779003083706, + 0.003971031401306391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.2535516023635864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006774434354156256 + ], + "opposite_logprob_tokens": [ + -5.006774425506592 + ], + "logprob_reward": -0.006774434354156256, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006774434354156256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2038610577583313, + 0.09221343696117401, + 0.07749678194522858, + 0.10357251763343811, + 0.0584799163043499, + 0.067452572286129, + 0.13120277225971222, + 0.0940416008234024, + 0.12430630624294281, + 0.04737303778529167 + ], + "credit_map": [ + 0.021582987159490585, + 0.026731954887509346, + 0.05358593910932541, + 0.017752856016159058, + 0.015627717599272728, + 0.03259648382663727, + 0.0185089111328125, + 0.017474208027124405, + 0.012801658362150192, + 0.020459484308958054, + 0.0099832434207201, + 0.008712236769497395, + 0.01220344752073288, + 0.009990082122385502, + 0.007907314226031303, + 0.010155973955988884, + 0.010158680379390717, + 0.009460562840104103, + 0.01005830429494381, + 0.004483298864215612, + 0.005701006855815649, + 0.012801955454051495, + 0.0058982884511351585, + 0.018934687599539757, + 0.016294114291667938, + 0.014189659617841244, + 0.015287572517991066, + 0.005719591397792101, + 0.0069174873642623425, + 0.01274862140417099, + 0.01331367902457714, + 0.019101792946457863, + 0.007206595037132502, + 0.009303438477218151, + 0.00907599925994873, + 0.006085888482630253, + 0.005051984451711178, + 0.007302822079509497, + 0.007552590221166611, + 0.006900598295032978, + 0.008137635886669159, + 0.010283452458679676, + 0.009514207020401955, + 0.006888650357723236, + 0.008598968386650085, + 0.009486770257353783, + 0.0053961873054504395, + 0.00914670154452324, + 0.013269982300698757, + 0.017375009134411812, + 0.013807905837893486, + 0.01653132773935795, + 0.011987930163741112, + 0.008071242831647396, + 0.027574168518185616, + 0.02258521318435669, + 0.013670669868588448, + 0.016761451959609985, + 0.014916121028363705, + 0.008384617045521736, + 0.011478916741907597, + 0.009799791499972343, + 0.009299218654632568, + 0.009730814024806023, + 0.012754908762872219, + 0.021933773532509804, + 0.014457390643656254, + 0.017488233745098114, + 0.019051123410463333, + 0.010898490436375141, + 0.017131388187408447, + 0.010590999387204647, + 0.004293360747396946, + 0.004513937048614025, + 0.005661450792104006, + 0.00727928476408124, + 0.00659899041056633, + 0.006599364336580038, + 0.005967315286397934, + 0.006459333933889866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.006774434354156256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639005314558744 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007639005314558744, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639005314558744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3594921827316284, + 0.04753836244344711, + 0.1535109132528305, + 0.07116915285587311, + 0.050854071974754333, + 0.08343353122472763, + 0.045949239283800125, + 0.07637326419353485, + 0.06930401921272278, + 0.04237526282668114 + ], + "credit_map": [ + 0.04886169731616974, + 0.04916072264313698, + 0.06878986954689026, + 0.040013909339904785, + 0.032673005014657974, + 0.05147486925125122, + 0.03177107870578766, + 0.0367470346391201, + 0.004886717069894075, + 0.007036016322672367, + 0.0068981098011136055, + 0.005722847301512957, + 0.006007002200931311, + 0.0045264046639204025, + 0.004574593622237444, + 0.00788667518645525, + 0.020383255556225777, + 0.01814926229417324, + 0.027622075751423836, + 0.007967546582221985, + 0.011282682418823242, + 0.017203452065587044, + 0.012157469987869263, + 0.03874516487121582, + 0.005612064152956009, + 0.00977863185107708, + 0.009976455941796303, + 0.005934900604188442, + 0.008626366034150124, + 0.01335209421813488, + 0.006766404956579208, + 0.011122235096991062, + 0.0059609124436974525, + 0.007946118712425232, + 0.009051908738911152, + 0.004971121437847614, + 0.00465004239231348, + 0.00966678187251091, + 0.0040954663418233395, + 0.00451172050088644, + 0.007094731554389, + 0.017896950244903564, + 0.021271828562021255, + 0.006244488060474396, + 0.006567258853465319, + 0.00836010742932558, + 0.007576842326670885, + 0.008421317674219608, + 0.006131470203399658, + 0.007309535518288612, + 0.006292307283729315, + 0.0047385054640471935, + 0.0030411414336413145, + 0.0045667593367397785, + 0.007391524501144886, + 0.0064779953099787235, + 0.008790092542767525, + 0.012839813716709614, + 0.013966596685349941, + 0.009208228439092636, + 0.010468712076544762, + 0.0056902640499174595, + 0.006579689681529999, + 0.008829868398606777, + 0.006753003224730492, + 0.009603307582437992, + 0.00625296775251627, + 0.00930104311555624, + 0.012312476523220539, + 0.006369836162775755, + 0.010380735620856285, + 0.008330648764967918, + 0.0056752655655145645, + 0.004719652235507965, + 0.004611379001289606, + 0.005092797800898552, + 0.007110347039997578, + 0.00518671004101634, + 0.00427657924592495, + 0.005702530033886433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.007639005314558744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636326290667057 + ], + "opposite_logprob_tokens": [ + -3.6513633728027344 + ], + "logprob_reward": -0.02636326290667057, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636326290667057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09578393399715424, + 0.07882732897996902, + 0.297611802816391, + 0.11944952607154846, + 0.04091266915202141, + 0.07177390903234482, + 0.08241426944732666, + 0.10274013131856918, + 0.053857091814279556, + 0.05662933364510536 + ], + "credit_map": [ + 0.013293465599417686, + 0.014225211925804615, + 0.010769186541438103, + 0.011388765648007393, + 0.010183786042034626, + 0.009366045705974102, + 0.01065351814031601, + 0.01590396650135517, + 0.00865914486348629, + 0.018584031611680984, + 0.008629792369902134, + 0.007968086749315262, + 0.010150651447474957, + 0.007036979775875807, + 0.007624438032507896, + 0.010174218565225601, + 0.04156562313437462, + 0.036319028586149216, + 0.04721662029623985, + 0.017003417015075684, + 0.02717924490571022, + 0.0287475623190403, + 0.029008524492383003, + 0.07057181000709534, + 0.01024646032601595, + 0.016836974769830704, + 0.017055273056030273, + 0.010422039777040482, + 0.01162960659712553, + 0.019335318356752396, + 0.011536811478435993, + 0.022387050092220306, + 0.0048154364340007305, + 0.006924130953848362, + 0.005837212782353163, + 0.0038865343667566776, + 0.004083526786416769, + 0.0044821095652878284, + 0.005625024903565645, + 0.005258695688098669, + 0.006113885436207056, + 0.013189615681767464, + 0.013480234891176224, + 0.005959704983979464, + 0.00612670136615634, + 0.007744619157165289, + 0.009621145203709602, + 0.009538008831441402, + 0.008521267212927341, + 0.0109702805057168, + 0.011638814583420753, + 0.00906692910939455, + 0.0073636602610349655, + 0.008747805841267109, + 0.01316049788147211, + 0.012945019640028477, + 0.014460298232734203, + 0.013395686633884907, + 0.019808918237686157, + 0.00982500147074461, + 0.013787634670734406, + 0.009126231074333191, + 0.01158367283642292, + 0.010752696543931961, + 0.008036421611905098, + 0.007718109525740147, + 0.0068891881965100765, + 0.00832759216427803, + 0.00803734827786684, + 0.004036634229123592, + 0.005677361041307449, + 0.005134438630193472, + 0.00967707484960556, + 0.007574711926281452, + 0.0055414228700101376, + 0.00673246756196022, + 0.009071498177945614, + 0.006315171252936125, + 0.006470399908721447, + 0.005246592685580254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.02636326290667057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0539096593856812 + ], + "opposite_logprob_tokens": [ + -0.42890965938568115 + ], + "logprob_reward": -1.0539096593856812, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539096593856812, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08708610385656357, + 0.11191046983003616, + 0.2124144583940506, + 0.2089076042175293, + 0.04970933124423027, + 0.04087430238723755, + 0.0606890469789505, + 0.13040320575237274, + 0.0723661556839943, + 0.025639338418841362 + ], + "credit_map": [ + 0.011586704291403294, + 0.019161120057106018, + 0.014014125801622868, + 0.009559296071529388, + 0.00799970980733633, + 0.010656826198101044, + 0.0067609460093081, + 0.007347372826188803, + 0.013398099690675735, + 0.022566886618733406, + 0.01282310951501131, + 0.01194716989994049, + 0.014512646943330765, + 0.010914746671915054, + 0.011259247548878193, + 0.014488555490970612, + 0.028653306886553764, + 0.02364569716155529, + 0.03701270371675491, + 0.0102148512378335, + 0.01644877716898918, + 0.024005785584449768, + 0.0160744097083807, + 0.0563589371740818, + 0.024783214554190636, + 0.024741079658269882, + 0.030759669840335846, + 0.013438558205962181, + 0.019567938521504402, + 0.04187280684709549, + 0.021004963666200638, + 0.03273935988545418, + 0.005670398473739624, + 0.006581870373338461, + 0.0079307546839118, + 0.0035746300127357244, + 0.006003936752676964, + 0.011537895537912846, + 0.004111443646252155, + 0.0042984033934772015, + 0.005653840024024248, + 0.006360245402902365, + 0.0075880675576627254, + 0.003410153556615114, + 0.004356557037681341, + 0.005455810111016035, + 0.004184577614068985, + 0.0038650508504360914, + 0.007689973339438438, + 0.008536379784345627, + 0.00861715991050005, + 0.006701601203531027, + 0.005727340467274189, + 0.005024991929531097, + 0.009727424010634422, + 0.008664180524647236, + 0.023685500025749207, + 0.016660401597619057, + 0.026621637865900993, + 0.012585376389324665, + 0.01274279598146677, + 0.011042715050280094, + 0.01287183165550232, + 0.014192955568432808, + 0.009782235138118267, + 0.011701825074851513, + 0.010311438702046871, + 0.009849843569099903, + 0.011239934712648392, + 0.005517071112990379, + 0.008114849217236042, + 0.005848961882293224, + 0.004352210555225611, + 0.0038885052781552076, + 0.0021967613138258457, + 0.0032891284208744764, + 0.003549078945070505, + 0.002772754989564419, + 0.003089107805863023, + 0.0025017925072461367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -1.0539096593856812 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "From a remarkably steady first-person perspective, the camera pans in a slow, smooth clockwise orbit, circling an old-fashioned rattan chair, resting in the center of a vast, empty room. A beam of warm afternoon sunlight slants in from an off-screen window, casting a soft golden halo around the chair's outline and long, sharply defined shadows across the polished wooden floor. Airborne dust particles are clearly visible in the beam. As the camera pans, light flows over every surface of the rattan chair, vividly highlighting its age-worn texture with hyper-realistic rendering. Some of the rattan strips on the back are worn and broken, revealing their coarse fibers. The armrests, once sleek and glossy from years of wear, have lost their original varnish, revealing a lighter wood tone beneath. A fine layer of dust accumulates in the seams of the seat's weaving, visible in the light. The whole picture presents an extreme realism style. The texture of the rattan and the grain of the wood seem to be within reach, creating a quiet, nostalgic and highly contemplative atmosphere.", + "seed": 19688, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_14/index_501_seed_19688_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": -0.31408572383224964, + "vlm_logprob": -0.31408572383224964, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.31408572383224964, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28043872117996216 + ], + "opposite_logprob_tokens": [ + -1.5304386615753174 + ], + "logprob_reward": -0.28043872117996216, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.28043872117996216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07416463643312454, + 0.034950755536556244, + 0.024306727573275566, + 0.19678540527820587, + 0.07743623852729797, + 0.10206588357686996, + 0.03361562639474869, + 0.22353540360927582, + 0.1869419515132904, + 0.04619735851883888 + ], + "credit_map": [ + 0.009124928154051304, + 0.010197623632848263, + 0.011604498140513897, + 0.006301053334027529, + 0.00799628160893917, + 0.0090389559045434, + 0.010123031213879585, + 0.00977826677262783, + 0.006444309372454882, + 0.0038472511805593967, + 0.0077121080830693245, + 0.002493423642590642, + 0.0035978234373033047, + 0.00459353718906641, + 0.0024177294690161943, + 0.003844573860988021, + 0.0030308810528367758, + 0.0028403454925864935, + 0.003487854264676571, + 0.002301811007782817, + 0.003701048204675317, + 0.002909149741753936, + 0.004301808308809996, + 0.001733829965814948, + 0.016718724742531776, + 0.02075939252972603, + 0.04152686148881912, + 0.020149042829871178, + 0.02427074871957302, + 0.028053270652890205, + 0.02683921717107296, + 0.018468137830495834, + 0.010810760781168938, + 0.007870149798691273, + 0.018616443499922752, + 0.006135658826678991, + 0.010626505129039288, + 0.009637986309826374, + 0.0063795894384384155, + 0.007359139621257782, + 0.00758365960791707, + 0.019787339493632317, + 0.017750222235918045, + 0.009219503030180931, + 0.013515453785657883, + 0.016177194193005562, + 0.010538791306316853, + 0.0074937124736607075, + 0.004860597662627697, + 0.005492453929036856, + 0.00474312761798501, + 0.0038786716759204865, + 0.0045660557225346565, + 0.004873054102063179, + 0.002499836264178157, + 0.0027018291875720024, + 0.040579553693532944, + 0.020061513409018517, + 0.02742377109825611, + 0.009202971123158932, + 0.01820646971464157, + 0.03218856081366539, + 0.02173498459160328, + 0.054137568920850754, + 0.04151695594191551, + 0.019278325140476227, + 0.010529179126024246, + 0.00674500223249197, + 0.011897269636392593, + 0.07666335254907608, + 0.00735717685893178, + 0.012954694218933582, + 0.007546859793365002, + 0.004124491009861231, + 0.003307004924863577, + 0.0034813317470252514, + 0.0037051094695925713, + 0.011987193487584591, + 0.0033470855560153723, + 0.008698281832039356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.28043872117996216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004664136562496424 + ], + "opposite_logprob_tokens": [ + -5.379663944244385 + ], + "logprob_reward": -0.004664136562496424, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004664136562496424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15729276835918427, + 0.024115566164255142, + 0.05059955269098282, + 0.20710889995098114, + 0.056007735431194305, + 0.11020945012569427, + 0.06521735340356827, + 0.17234206199645996, + 0.0871611088514328, + 0.06994548439979553 + ], + "credit_map": [ + 0.028803713619709015, + 0.015095147304236889, + 0.0295699592679739, + 0.008732757531106472, + 0.01764157973229885, + 0.01934262178838253, + 0.01844714768230915, + 0.019659852609038353, + 0.00470726378262043, + 0.0040057054720819, + 0.0037113719154149294, + 0.001975125866010785, + 0.003210952505469322, + 0.0027306298725306988, + 0.0022140666842460632, + 0.001560449250973761, + 0.008961609564721584, + 0.0052034673281013966, + 0.01025191880762577, + 0.0033533796668052673, + 0.008131342940032482, + 0.0071493638679385185, + 0.0041463859379291534, + 0.0034020827151834965, + 0.018682794645428658, + 0.021426904946565628, + 0.04449672996997833, + 0.02067217230796814, + 0.025384249165654182, + 0.03039482794702053, + 0.027355216443538666, + 0.018696004524827003, + 0.005364503711462021, + 0.006318565923720598, + 0.013171263039112091, + 0.005555589683353901, + 0.007914024405181408, + 0.006640053354203701, + 0.005403286311775446, + 0.005640449933707714, + 0.009809495881199837, + 0.017676740884780884, + 0.015009352937340736, + 0.012443825602531433, + 0.015415706671774387, + 0.0169037077575922, + 0.012717202305793762, + 0.010233419947326183, + 0.007272873539477587, + 0.011623261496424675, + 0.00771320378407836, + 0.009077400900423527, + 0.007073770742863417, + 0.009588751941919327, + 0.007169662043452263, + 0.00569842429831624, + 0.03313577175140381, + 0.016699740663170815, + 0.020573053508996964, + 0.0067480518482625484, + 0.015286101028323174, + 0.02487810328602791, + 0.015719041228294373, + 0.03930221125483513, + 0.009600069373846054, + 0.014881359413266182, + 0.009171715006232262, + 0.00904091726988554, + 0.013416026718914509, + 0.013285856693983078, + 0.008747715502977371, + 0.009017447009682655, + 0.010661964304745197, + 0.0053459578193724155, + 0.006847269833087921, + 0.0044624884612858295, + 0.004609155934303999, + 0.015395169146358967, + 0.005678801331669092, + 0.01694467104971409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.004664136562496424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674400432035327 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.00674400432035327, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674400432035327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128468319773674, + 0.0682397335767746, + 0.03726508840918541, + 0.025096746161580086, + 0.06166396662592888, + 0.07206013798713684, + 0.06906706094741821, + 0.2795381546020508, + 0.19646048545837402, + 0.062140338122844696 + ], + "credit_map": [ + 0.02008172497153282, + 0.020500315353274345, + 0.018672410398721695, + 0.009650735184550285, + 0.01741383969783783, + 0.013485989533364773, + 0.014807986095547676, + 0.013855313882231712, + 0.015429100953042507, + 0.007226888556033373, + 0.010503188706934452, + 0.004642974119633436, + 0.011520447209477425, + 0.008384994231164455, + 0.005479131359606981, + 0.005053004249930382, + 0.00705262366682291, + 0.005461329594254494, + 0.004345563240349293, + 0.0031262764241546392, + 0.004768481478095055, + 0.004540863446891308, + 0.004375038668513298, + 0.0035949095617979765, + 0.0022361166775226593, + 0.0021838464308530092, + 0.006392292678356171, + 0.0017385795945301652, + 0.0030979544389992952, + 0.003283575177192688, + 0.003165705129504204, + 0.00299867638386786, + 0.008301318623125553, + 0.005080760456621647, + 0.017380911856889725, + 0.005476407241076231, + 0.007881510071456432, + 0.009319642558693886, + 0.004414078313857317, + 0.003809335408732295, + 0.008105478249490261, + 0.005827188957482576, + 0.014547382481396198, + 0.006400264333933592, + 0.007135320454835892, + 0.01256479136645794, + 0.0069871023297309875, + 0.010492607951164246, + 0.005743531044572592, + 0.009080039337277412, + 0.011631689965724945, + 0.00638862419873476, + 0.0098712258040905, + 0.011607972905039787, + 0.005206533707678318, + 0.009537439793348312, + 0.05411004647612572, + 0.02631501480937004, + 0.0337517186999321, + 0.01102334912866354, + 0.02350122109055519, + 0.040267497301101685, + 0.025449352338910103, + 0.06511994451284409, + 0.04227984696626663, + 0.019789056852459908, + 0.011552685871720314, + 0.007417828310281038, + 0.012361071072518826, + 0.07980724424123764, + 0.008218224160373211, + 0.015034534968435764, + 0.009766234084963799, + 0.01120163407176733, + 0.006864628288894892, + 0.006666216067969799, + 0.0107098538428545, + 0.005492129363119602, + 0.005182585213333368, + 0.006257057189941406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.00674400432035327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054962534457445145 + ], + "opposite_logprob_tokens": [ + -2.929962635040283 + ], + "logprob_reward": -0.054962534457445145, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054962534457445145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13383594155311584, + 0.06968223303556442, + 0.08241027593612671, + 0.06613301485776901, + 0.037227239459753036, + 0.06048426777124405, + 0.03593698889017105, + 0.2891418933868408, + 0.08111248165369034, + 0.14403559267520905 + ], + "credit_map": [ + 0.025225140154361725, + 0.013323470950126648, + 0.02394607663154602, + 0.008540943264961243, + 0.01896214298903942, + 0.01633162796497345, + 0.01355762779712677, + 0.013948907144367695, + 0.007092136424034834, + 0.013606053777039051, + 0.008648662827908993, + 0.007554884999990463, + 0.012000955641269684, + 0.008132598362863064, + 0.007977931760251522, + 0.004669006913900375, + 0.013888480141758919, + 0.008797923102974892, + 0.01535097323358059, + 0.007199665065854788, + 0.010907931253314018, + 0.010918446816504002, + 0.008367346599698067, + 0.006979505531489849, + 0.006771983113139868, + 0.006591420620679855, + 0.01396167278289795, + 0.006273829843848944, + 0.008611922152340412, + 0.011093288660049438, + 0.008399627171456814, + 0.004429264925420284, + 0.0032927230931818485, + 0.0055415746755898, + 0.005188527051359415, + 0.003904387354850769, + 0.005151460878551006, + 0.005898699164390564, + 0.004724659025669098, + 0.0035252058878540993, + 0.005382584407925606, + 0.011003134772181511, + 0.007147310767322779, + 0.006516552530229092, + 0.008642172440886497, + 0.009048035368323326, + 0.0061334893107414246, + 0.006610988173633814, + 0.005261058919131756, + 0.0051645259372889996, + 0.0057628583163022995, + 0.0035829623229801655, + 0.0038409121334552765, + 0.006682397797703743, + 0.0032144421711564064, + 0.0024278301279991865, + 0.05861625820398331, + 0.027018794789910316, + 0.03382844105362892, + 0.012036596424877644, + 0.023165393620729446, + 0.04237882047891617, + 0.026161473244428635, + 0.06593611091375351, + 0.01614314690232277, + 0.008371135219931602, + 0.004990838933736086, + 0.003400384681299329, + 0.0060353525914251804, + 0.028950413689017296, + 0.004183893091976643, + 0.009037312120199203, + 0.01950572431087494, + 0.01839076355099678, + 0.012516926042735577, + 0.012314027175307274, + 0.019392212852835655, + 0.01993251033127308, + 0.010450161062180996, + 0.03153328597545624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.054962534457445145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.2236192226409912 + ], + "opposite_logprob_tokens": [ + -0.3486192226409912 + ], + "logprob_reward": -1.2236192226409912, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236192226409912, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24399369955062866, + 0.05154483765363693, + 0.03611389175057411, + 0.05294214189052582, + 0.07027165591716766, + 0.03474261984229088, + 0.03811005875468254, + 0.07966666668653488, + 0.18581846356391907, + 0.20679594576358795 + ], + "credit_map": [ + 0.04333893582224846, + 0.022716516628861427, + 0.04933299496769905, + 0.013557753525674343, + 0.025692177936434746, + 0.03197525441646576, + 0.02093229815363884, + 0.036447763442993164, + 0.009525143541395664, + 0.005119480192661285, + 0.011352229863405228, + 0.0035546242725104094, + 0.005540726240724325, + 0.006304264068603516, + 0.005094853229820728, + 0.005053517408668995, + 0.005533520597964525, + 0.005289341788738966, + 0.004871796816587448, + 0.002848452189937234, + 0.004766877740621567, + 0.004942384548485279, + 0.0037841848097741604, + 0.004077336750924587, + 0.00499174976721406, + 0.004760469309985638, + 0.01122208870947361, + 0.004850390367209911, + 0.006048643030226231, + 0.008542874827980995, + 0.007118936628103256, + 0.005406991112977266, + 0.0061079529114067554, + 0.010157031007111073, + 0.014691529795527458, + 0.006696437019854784, + 0.010255992412567139, + 0.008790969848632812, + 0.0068680900149047375, + 0.006703651510179043, + 0.0032938742078840733, + 0.005162525922060013, + 0.006608267314732075, + 0.0032746312208473682, + 0.0041037667542696, + 0.005914003122597933, + 0.0029843891970813274, + 0.00340116024017334, + 0.004824018571525812, + 0.0040416778065264225, + 0.0061891404911875725, + 0.004355581011623144, + 0.004389344248920679, + 0.005745404865592718, + 0.003984479699283838, + 0.004580413457006216, + 0.013153521344065666, + 0.00684344070032239, + 0.009416446089744568, + 0.005357970483601093, + 0.006412052549421787, + 0.011859527789056301, + 0.007025038357824087, + 0.019598668441176414, + 0.03930722177028656, + 0.019409116357564926, + 0.011738675646483898, + 0.0077194832265377045, + 0.01345000695437193, + 0.07034656405448914, + 0.008533284068107605, + 0.015314106829464436, + 0.033449772745370865, + 0.016515402123332024, + 0.018528616055846214, + 0.015034574083983898, + 0.014944704249501228, + 0.04523370414972305, + 0.015655463561415672, + 0.047433700412511826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -1.2236192226409912 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "From a remarkably steady first-person perspective, the camera pans in a slow, smooth clockwise orbit, circling an old-fashioned rattan chair, resting in the center of a vast, empty room. A beam of warm afternoon sunlight slants in from an off-screen window, casting a soft golden halo around the chair's outline and long, sharply defined shadows across the polished wooden floor. Airborne dust particles are clearly visible in the beam. As the camera pans, light flows over every surface of the rattan chair, vividly highlighting its age-worn texture with hyper-realistic rendering. Some of the rattan strips on the back are worn and broken, revealing their coarse fibers. The armrests, once sleek and glossy from years of wear, have lost their original varnish, revealing a lighter wood tone beneath. A fine layer of dust accumulates in the seams of the seat's weaving, visible in the light. The whole picture presents an extreme realism style. The texture of the rattan and the grain of the wood seem to be within reach, creating a quiet, nostalgic and highly contemplative atmosphere.", + "seed": 19689, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_14/index_501_seed_19689_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": -0.2576404111459851, + "vlm_logprob": -0.2576404111459851, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.2576404111459851, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3375943601131439 + ], + "opposite_logprob_tokens": [ + -1.3375943899154663 + ], + "logprob_reward": -0.3375943601131439, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3375943601131439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09722467511892319, + 0.04967828840017319, + 0.0816817358136177, + 0.06803563982248306, + 0.07495655864477158, + 0.327913373708725, + 0.07855967432260513, + 0.10779852420091629, + 0.054588254541158676, + 0.05956330895423889 + ], + "credit_map": [ + 0.012998306192457676, + 0.013595063239336014, + 0.011015268042683601, + 0.012255830690264702, + 0.011950879357755184, + 0.011534142307937145, + 0.013152728788554668, + 0.010722463950514793, + 0.0038993193302303553, + 0.008748416788876057, + 0.00809259433299303, + 0.0025356861297041178, + 0.004753890447318554, + 0.003957665991038084, + 0.0032374574802815914, + 0.014453255571424961, + 0.010647776536643505, + 0.011187420226633549, + 0.014518776908516884, + 0.0074918814934790134, + 0.013027526438236237, + 0.0074553475715219975, + 0.006027742754667997, + 0.011325264349579811, + 0.005971813108772039, + 0.009602135978639126, + 0.01259563583880663, + 0.006864458788186312, + 0.012427953071892262, + 0.008074031211435795, + 0.004341144114732742, + 0.008158460259437561, + 0.012749876827001572, + 0.00521516427397728, + 0.003303146455436945, + 0.027947688475251198, + 0.005976627115160227, + 0.004542001057416201, + 0.004535744898021221, + 0.010686303488910198, + 0.0692610815167427, + 0.11452600359916687, + 0.013609444722533226, + 0.021925874054431915, + 0.02620660699903965, + 0.040612656623125076, + 0.022133843973279, + 0.01963784359395504, + 0.009924390353262424, + 0.01586073264479637, + 0.005556685850024223, + 0.00956946425139904, + 0.009888536296784878, + 0.011177705600857735, + 0.008835173211991787, + 0.00774699030444026, + 0.016783930361270905, + 0.012859185226261616, + 0.022002633661031723, + 0.009807872585952282, + 0.013917719013988972, + 0.014602088369429111, + 0.009916558861732483, + 0.007908547297120094, + 0.0053120204247534275, + 0.004369910340756178, + 0.004682912491261959, + 0.012627577409148216, + 0.005745945032685995, + 0.008741485886275768, + 0.0054814680479466915, + 0.007626937236636877, + 0.00827002339065075, + 0.007197835016995668, + 0.006895578000694513, + 0.006885151844471693, + 0.007272201590240002, + 0.008759234100580215, + 0.006488373503088951, + 0.007794910576194525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.3375943601131439 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009815759025514126 + ], + "opposite_logprob_tokens": [ + -4.634815692901611 + ], + "logprob_reward": -0.009815759025514126, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009815759025514126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1332203596830368, + 0.06787078082561493, + 0.26360008120536804, + 0.04547605291008949, + 0.045723773539066315, + 0.15603797137737274, + 0.09832979738712311, + 0.04252462461590767, + 0.061403416097164154, + 0.08581313490867615 + ], + "credit_map": [ + 0.018521875143051147, + 0.015440339222550392, + 0.015746617689728737, + 0.017299000173807144, + 0.021696247160434723, + 0.014655142091214657, + 0.013062212616205215, + 0.016798943281173706, + 0.009770162403583527, + 0.014554044231772423, + 0.006025549955666065, + 0.009324670769274235, + 0.008140109479427338, + 0.006722167134284973, + 0.007251587696373463, + 0.006082496140152216, + 0.03321612626314163, + 0.03342948853969574, + 0.04048158600926399, + 0.021577375009655952, + 0.0622660256922245, + 0.017653290182352066, + 0.01998511143028736, + 0.03499109670519829, + 0.003466276917606592, + 0.007205336820334196, + 0.008698907680809498, + 0.005442207213491201, + 0.006530610844492912, + 0.0050908359698951244, + 0.004371791146695614, + 0.004670090042054653, + 0.005478749983012676, + 0.016460705548524857, + 0.00391509709879756, + 0.004081792663782835, + 0.0040627699345350266, + 0.0037631094455718994, + 0.0027687603142112494, + 0.005192792508751154, + 0.02499384805560112, + 0.04808909073472023, + 0.008838904090225697, + 0.015219165943562984, + 0.013878746889531612, + 0.021158039569854736, + 0.011461525224149227, + 0.01239864807575941, + 0.010035396553575993, + 0.014643054455518723, + 0.008596502244472504, + 0.011941627599298954, + 0.013334841467440128, + 0.014611025340855122, + 0.0138966990634799, + 0.01127066183835268, + 0.004903730005025864, + 0.006207407917827368, + 0.0033791125752031803, + 0.0061861006543040276, + 0.005356380250304937, + 0.0071877469308674335, + 0.0059143188409507275, + 0.003389830933883786, + 0.004457672126591206, + 0.004783222917467356, + 0.004266100469976664, + 0.0193566232919693, + 0.00440117297694087, + 0.008855827152729034, + 0.00745781883597374, + 0.007824983447790146, + 0.008405446074903011, + 0.00942659005522728, + 0.010783360339701176, + 0.00907410029321909, + 0.01419584546238184, + 0.014147430658340454, + 0.009978462010622025, + 0.009801914915442467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.009815759025514126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607470214366913 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.01607470214366913, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607470214366913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.139076367020607, + 0.09285851567983627, + 0.22437182068824768, + 0.11692743003368378, + 0.11298944801092148, + 0.15066303312778473, + 0.028434831649065018, + 0.02225266955792904, + 0.058554623275995255, + 0.0538712702691555 + ], + "credit_map": [ + 0.022576401010155678, + 0.018180668354034424, + 0.016047339886426926, + 0.01552759762853384, + 0.02486603334546089, + 0.01460558082908392, + 0.011216320097446442, + 0.016056431457400322, + 0.017107253894209862, + 0.021774617955088615, + 0.007357372436672449, + 0.007622988428920507, + 0.011379559524357319, + 0.008981985040009022, + 0.007046929094940424, + 0.011587805114686489, + 0.027119910344481468, + 0.03914708271622658, + 0.026880528777837753, + 0.019122518599033356, + 0.04828110709786415, + 0.0173768550157547, + 0.014278184622526169, + 0.03216564282774925, + 0.010533444583415985, + 0.01915682852268219, + 0.02005479484796524, + 0.0132142324000597, + 0.017384707927703857, + 0.013967631384730339, + 0.009554838761687279, + 0.013060947880148888, + 0.024363795295357704, + 0.010075446218252182, + 0.00477774441242218, + 0.02757387049496174, + 0.005191253963857889, + 0.016881603747606277, + 0.005424410104751587, + 0.018701326102018356, + 0.027632249519228935, + 0.05442406237125397, + 0.006450172048062086, + 0.0116692204028368, + 0.010579789057374, + 0.020564023405313492, + 0.01074548065662384, + 0.008598046377301216, + 0.003670871490612626, + 0.006447131279855967, + 0.0014364083763211966, + 0.004003217443823814, + 0.0024614257272332907, + 0.0046859439462423325, + 0.0035756153520196676, + 0.002154218265786767, + 0.0025246376171708107, + 0.0034322082065045834, + 0.0016053911531344056, + 0.002659827470779419, + 0.0023731410037726164, + 0.0050421240739524364, + 0.0027528617065399885, + 0.0018624811200425029, + 0.009630205109715462, + 0.005660468712449074, + 0.003610161365941167, + 0.009227992035448551, + 0.005684086587280035, + 0.010263411328196526, + 0.006344733294099569, + 0.00813356600701809, + 0.004924776963889599, + 0.005258812103420496, + 0.005202966276556253, + 0.00682647293433547, + 0.006476893089711666, + 0.010396498255431652, + 0.007862028665840626, + 0.006922823842614889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.01607470214366913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361869648098946 + ], + "opposite_logprob_tokens": [ + -3.6513619422912598 + ], + "logprob_reward": -0.026361869648098946, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361869648098946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11017593741416931, + 0.15893442928791046, + 0.07122956216335297, + 0.0673089250922203, + 0.07056643068790436, + 0.07566072046756744, + 0.05272648110985756, + 0.05124270170927048, + 0.18367119133472443, + 0.1584835797548294 + ], + "credit_map": [ + 0.017110740765929222, + 0.00902936514467001, + 0.012433024123311043, + 0.014982319436967373, + 0.014738347381353378, + 0.012279377318918705, + 0.012435905635356903, + 0.017166852951049805, + 0.014181618578732014, + 0.011865105479955673, + 0.037897855043411255, + 0.008935248479247093, + 0.015832586213946342, + 0.00925651378929615, + 0.007655913010239601, + 0.053309597074985504, + 0.010728543624281883, + 0.008287807926535606, + 0.013536863960325718, + 0.006956874392926693, + 0.011453315615653992, + 0.006733965594321489, + 0.005542403552681208, + 0.007989790290594101, + 0.008130875416100025, + 0.011907491832971573, + 0.009441886097192764, + 0.006912467535585165, + 0.011258058249950409, + 0.006600518245249987, + 0.005687902215868235, + 0.007369723170995712, + 0.01178919430822134, + 0.0076199788600206375, + 0.0033971292432397604, + 0.014482160098850727, + 0.006032317411154509, + 0.010169042274355888, + 0.0036961839068681, + 0.013380428776144981, + 0.010175046510994434, + 0.012668722309172153, + 0.00490012438967824, + 0.014005957171320915, + 0.009816902689635754, + 0.008749961853027344, + 0.009735757485032082, + 0.005608249455690384, + 0.00721420394256711, + 0.013739200308918953, + 0.002817838918417692, + 0.006071922369301319, + 0.005786517169326544, + 0.008054186590015888, + 0.005405227188020945, + 0.0036373839247971773, + 0.005967207718640566, + 0.006850103847682476, + 0.003714857157319784, + 0.005894793197512627, + 0.007224321831017733, + 0.00979867298156023, + 0.006713067647069693, + 0.005079683382064104, + 0.01670169271528721, + 0.01561484020203352, + 0.010391703806817532, + 0.05088033527135849, + 0.017606982961297035, + 0.029260849580168724, + 0.02030201070010662, + 0.022912779822945595, + 0.02088012918829918, + 0.015887504443526268, + 0.01230668555945158, + 0.023116232827305794, + 0.013246548362076283, + 0.025311607867479324, + 0.022485030815005302, + 0.025249838829040527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.026361869648098946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8983553647994995 + ], + "opposite_logprob_tokens": [ + -0.5233553647994995 + ], + "logprob_reward": -0.8983553647994995, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983553647994995, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15352749824523926, + 0.03806913271546364, + 0.06602919101715088, + 0.013275562785565853, + 0.10270214825868607, + 0.3790556490421295, + 0.026494013145565987, + 0.034884363412857056, + 0.13203448057174683, + 0.05392793193459511 + ], + "credit_map": [ + 0.025699982419610023, + 0.020574932917952538, + 0.01891939714550972, + 0.018633222207427025, + 0.024845639243721962, + 0.018247246742248535, + 0.013785078190267086, + 0.012821987271308899, + 0.005179549567401409, + 0.006416982971131802, + 0.007567433640360832, + 0.0031853874679654837, + 0.003771504620090127, + 0.0037635122425854206, + 0.002595716854557395, + 0.005589041858911514, + 0.010582788847386837, + 0.007677673827856779, + 0.01091547030955553, + 0.004653459414839745, + 0.014218133874237537, + 0.005694995168596506, + 0.003117281710729003, + 0.009169384837150574, + 0.0015447646146640182, + 0.002622688887640834, + 0.001613559783436358, + 0.001088800490833819, + 0.0022859591990709305, + 0.0016069627599790692, + 0.0011141475988551974, + 0.001398678170517087, + 0.023472553119063377, + 0.006293673999607563, + 0.004121553152799606, + 0.025822030380368233, + 0.0045665898360311985, + 0.017521515488624573, + 0.005952736828476191, + 0.014951479621231556, + 0.07471875846385956, + 0.1339026540517807, + 0.015998248010873795, + 0.029732119292020798, + 0.030537758022546768, + 0.04608429595828056, + 0.026256874203681946, + 0.021824941039085388, + 0.0028055477887392044, + 0.0036570215597748756, + 0.00221149786375463, + 0.0031467473600059748, + 0.004498997237533331, + 0.003397156484425068, + 0.004298886749893427, + 0.0024781569372862577, + 0.004470708314329386, + 0.004631424322724342, + 0.004525855183601379, + 0.00375681696459651, + 0.005280947778373957, + 0.005615937057882547, + 0.0037646950222551823, + 0.002837975975126028, + 0.01100954320281744, + 0.009944843128323555, + 0.006990282330662012, + 0.042593102902173996, + 0.009775565937161446, + 0.020821111276745796, + 0.014163336716592312, + 0.01673668995499611, + 0.0058386800810694695, + 0.005459451116621494, + 0.005700905807316303, + 0.007745520677417517, + 0.005172876175493002, + 0.01068094652146101, + 0.005447624251246452, + 0.007881924510002136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.8983553647994995 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "From a remarkably steady first-person perspective, the camera pans in a slow, smooth clockwise orbit, circling an old-fashioned rattan chair, resting in the center of a vast, empty room. A beam of warm afternoon sunlight slants in from an off-screen window, casting a soft golden halo around the chair's outline and long, sharply defined shadows across the polished wooden floor. Airborne dust particles are clearly visible in the beam. As the camera pans, light flows over every surface of the rattan chair, vividly highlighting its age-worn texture with hyper-realistic rendering. Some of the rattan strips on the back are worn and broken, revealing their coarse fibers. The armrests, once sleek and glossy from years of wear, have lost their original varnish, revealing a lighter wood tone beneath. A fine layer of dust accumulates in the seams of the seat's weaving, visible in the light. The whole picture presents an extreme realism style. The texture of the rattan and the grain of the wood seem to be within reach, creating a quiet, nostalgic and highly contemplative atmosphere.", + "seed": 19690, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_14/index_501_seed_19690_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": -0.297171538695693, + "vlm_logprob": -0.297171538695693, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.297171538695693, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2914409339427948 + ], + "opposite_logprob_tokens": [ + -1.5414409637451172 + ], + "logprob_reward": -0.2914409339427948, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2914409339427948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10475770384073257, + 0.08024092763662338, + 0.025097323581576347, + 0.04991845041513443, + 0.029405856505036354, + 0.2628146708011627, + 0.08001650869846344, + 0.03495391085743904, + 0.24996928870677948, + 0.08282525092363358 + ], + "credit_map": [ + 0.022910693660378456, + 0.014541517943143845, + 0.011225184425711632, + 0.013824227266013622, + 0.011275251396000385, + 0.009114171378314495, + 0.01286719273775816, + 0.008999481797218323, + 0.017789924517273903, + 0.010656910017132759, + 0.009300099685788155, + 0.011249287985265255, + 0.008312893100082874, + 0.00814455933868885, + 0.005714769009500742, + 0.00907249003648758, + 0.004417479503899813, + 0.003763550193980336, + 0.002404259517788887, + 0.003889152081683278, + 0.0021409604232758284, + 0.0026922933757305145, + 0.0032152202911674976, + 0.0025744105223566294, + 0.00836257729679346, + 0.005848631728440523, + 0.0076654404401779175, + 0.006908323150128126, + 0.006044560577720404, + 0.006153981667011976, + 0.003982383757829666, + 0.00495256157591939, + 0.0049695586785674095, + 0.005328561645001173, + 0.003270304761826992, + 0.003263830905780196, + 0.0037305792793631554, + 0.0038606487214565277, + 0.002307133749127388, + 0.002675242256373167, + 0.030039293691515923, + 0.039765726774930954, + 0.030687982216477394, + 0.027673548087477684, + 0.021497931331396103, + 0.0439179353415966, + 0.017376450821757317, + 0.051855817437171936, + 0.007840396836400032, + 0.00733299320563674, + 0.008816744200885296, + 0.006496313959360123, + 0.007847982458770275, + 0.013007603585720062, + 0.007579646538943052, + 0.02109483629465103, + 0.004544840659946203, + 0.004291119985282421, + 0.004586982540786266, + 0.0032943657133728266, + 0.0032141455449163914, + 0.005189078859984875, + 0.005117855034768581, + 0.004715525545179844, + 0.04155448451638222, + 0.03271815925836563, + 0.03121567890048027, + 0.05753852799534798, + 0.018700208514928818, + 0.020788822323083878, + 0.021988103166222572, + 0.0254653412848711, + 0.008520971983671188, + 0.01173418015241623, + 0.01173394825309515, + 0.008061443455517292, + 0.0103566600009799, + 0.010177806951105595, + 0.009936708956956863, + 0.012303538620471954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.2914409339427948 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01257090549916029 + ], + "opposite_logprob_tokens": [ + -4.387570858001709 + ], + "logprob_reward": -0.01257090549916029, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01257090549916029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14847712218761444, + 0.0900544822216034, + 0.07017500698566437, + 0.04302565008401871, + 0.09235069155693054, + 0.05797508358955383, + 0.10252957791090012, + 0.043116070330142975, + 0.07274676859378815, + 0.27954959869384766 + ], + "credit_map": [ + 0.04322318732738495, + 0.022793911397457123, + 0.017017362639307976, + 0.017489159479737282, + 0.012960674241185188, + 0.010513050481677055, + 0.014199674129486084, + 0.010280094109475613, + 0.025593608617782593, + 0.014771275222301483, + 0.007900098338723183, + 0.010129187256097794, + 0.006799763068556786, + 0.008444695733487606, + 0.007070925552397966, + 0.009344917722046375, + 0.01957188919186592, + 0.009823815897107124, + 0.009769570082426071, + 0.007949655875563622, + 0.005812973249703646, + 0.006251242011785507, + 0.005217578262090683, + 0.0057782684452831745, + 0.008395970799028873, + 0.005661708302795887, + 0.0056406245566904545, + 0.004614527337253094, + 0.0044694687239825726, + 0.005669906735420227, + 0.00403542909771204, + 0.004538011737167835, + 0.017920099198818207, + 0.010747892782092094, + 0.020080694928765297, + 0.008559185080230236, + 0.01172164548188448, + 0.010030067525804043, + 0.004766169469803572, + 0.008524928241968155, + 0.006915846839547157, + 0.00855387281626463, + 0.007590160239487886, + 0.007248361129313707, + 0.0061129070818424225, + 0.009729346260428429, + 0.004207610618323088, + 0.007616971153765917, + 0.012328436598181725, + 0.011421436443924904, + 0.010649784468114376, + 0.009435176849365234, + 0.010753774084150791, + 0.020295603200793266, + 0.01183375809341669, + 0.01581159606575966, + 0.004984201863408089, + 0.003951210994273424, + 0.005765607580542564, + 0.00541852880269289, + 0.004596120677888393, + 0.008493655361235142, + 0.00478822598233819, + 0.005118511617183685, + 0.01242897566407919, + 0.008619512431323528, + 0.009853865951299667, + 0.019348420202732086, + 0.005025659222155809, + 0.005583606194704771, + 0.005743735935539007, + 0.006142988335341215, + 0.027956364676356316, + 0.035806335508823395, + 0.03867533057928085, + 0.03157015144824982, + 0.026242809370160103, + 0.03806864097714424, + 0.02213374897837639, + 0.05909621715545654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.01257090549916029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108358334749937 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.01108358334749937, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108358334749937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10075704008340836, + 0.24143829941749573, + 0.01833382062613964, + 0.06738398224115372, + 0.01814304292201996, + 0.08180343359708786, + 0.02626638300716877, + 0.030484091490507126, + 0.3574397563934326, + 0.05795017257332802 + ], + "credit_map": [ + 0.025623850524425507, + 0.012364706955850124, + 0.01192709244787693, + 0.016613252460956573, + 0.011922929435968399, + 0.006445464212447405, + 0.009069276973605156, + 0.00679045869037509, + 0.04602304473519325, + 0.040334612131118774, + 0.025861725211143494, + 0.04916710779070854, + 0.022460173815488815, + 0.020520158112049103, + 0.018935471773147583, + 0.018136009573936462, + 0.005085496697574854, + 0.002738509327173233, + 0.0017571249045431614, + 0.0027258119080215693, + 0.001237582415342331, + 0.002075809985399246, + 0.0013125501573085785, + 0.0014009351143613458, + 0.01611943542957306, + 0.009407415986061096, + 0.0072386902756989, + 0.009160223416984081, + 0.006524825468659401, + 0.00904804840683937, + 0.00440967408940196, + 0.005475665908306837, + 0.0035994057543575764, + 0.0017399924108758569, + 0.0018052986124530435, + 0.001905188080854714, + 0.0021292120218276978, + 0.0031989202834665775, + 0.001612414256669581, + 0.0021526108030229807, + 0.010105925612151623, + 0.013111443258821964, + 0.011143087409436703, + 0.010183172300457954, + 0.006195091642439365, + 0.013373651541769505, + 0.004989097826182842, + 0.012701957486569881, + 0.002567106857895851, + 0.003900941926985979, + 0.0027863874565809965, + 0.002656474243849516, + 0.0021181758493185043, + 0.004210617858916521, + 0.002530129626393318, + 0.00549654895439744, + 0.0035078551154583693, + 0.003956898115575314, + 0.003821487305685878, + 0.0025169372092932463, + 0.00308783445507288, + 0.0058231474831700325, + 0.0030838530510663986, + 0.004686078522354364, + 0.06060270220041275, + 0.04816349968314171, + 0.0422496497631073, + 0.07163269072771072, + 0.02891639620065689, + 0.03191646561026573, + 0.033558424562215805, + 0.04039992392063141, + 0.0064180586487054825, + 0.00728380773216486, + 0.005818212404847145, + 0.0075856843031942844, + 0.004834191873669624, + 0.006727639120072126, + 0.006418739911168814, + 0.012863840907812119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.01108358334749937 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371343016624451 + ], + "opposite_logprob_tokens": [ + -3.4087133407592773 + ], + "logprob_reward": -0.03371343016624451, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371343016624451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20146530866622925, + 0.10544571280479431, + 0.0669165849685669, + 0.12245989590883255, + 0.09216344356536865, + 0.04736475273966789, + 0.10460352152585983, + 0.07134626060724258, + 0.026579950004816055, + 0.16165465116500854 + ], + "credit_map": [ + 0.04924186319112778, + 0.022703249007463455, + 0.025340283289551735, + 0.022036883980035782, + 0.021731359884142876, + 0.017853619530797005, + 0.02512275241315365, + 0.017435289919376373, + 0.02529030293226242, + 0.016586437821388245, + 0.011582553386688232, + 0.020366927608847618, + 0.009102771990001202, + 0.008340833708643913, + 0.00783224031329155, + 0.006343637593090534, + 0.014166848734021187, + 0.010536198504269123, + 0.006808237172663212, + 0.013372299261391163, + 0.004796704743057489, + 0.006462274119257927, + 0.006057053338736296, + 0.004716965835541487, + 0.021486055105924606, + 0.019448895007371902, + 0.015269429422914982, + 0.016494199633598328, + 0.010709905065596104, + 0.014009756967425346, + 0.010965215973556042, + 0.014076435007154942, + 0.019528497010469437, + 0.012410708703100681, + 0.016779104247689247, + 0.00729473726823926, + 0.010253269225358963, + 0.012202584184706211, + 0.004680512007325888, + 0.009014030918478966, + 0.008254652842879295, + 0.005337115377187729, + 0.007296274416148663, + 0.005874961614608765, + 0.003668065881356597, + 0.007226044312119484, + 0.0032558636739850044, + 0.006451773457229137, + 0.012577390298247337, + 0.01405654102563858, + 0.013673948124051094, + 0.010832871310412884, + 0.011515002697706223, + 0.01827799156308174, + 0.0114012835547328, + 0.012268486432731152, + 0.0062229083850979805, + 0.007655241526663303, + 0.011070137843489647, + 0.007941313087940216, + 0.009314495138823986, + 0.010328064672648907, + 0.010654325596988201, + 0.008159777149558067, + 0.002305530710145831, + 0.004221378825604916, + 0.003170119831338525, + 0.00420197332277894, + 0.0023456658236682415, + 0.0027591651305556297, + 0.004005064256489277, + 0.003571052337065339, + 0.01639403961598873, + 0.02097914181649685, + 0.020730867981910706, + 0.019492875784635544, + 0.019817456603050232, + 0.01649862341582775, + 0.018755802884697914, + 0.02898584119975567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.03371343016624451 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and remarkably steady?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a slow, smooth clockwise orbit around an old-fashioned rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a golden halo and long shadows on a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are airborne dust particles clearly visible in the light beam, with visible wear, broken strips, and dust in the rattan chair's seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in an extreme realism style with hyper-realistic textures of the rattan and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.1370488405227661 + ], + "opposite_logprob_tokens": [ + -0.3870488107204437 + ], + "logprob_reward": -1.1370488405227661, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370488405227661, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22035536170005798, + 0.08576325327157974, + 0.06435942649841309, + 0.2030985951423645, + 0.025956932455301285, + 0.06415162980556488, + 0.08132290840148926, + 0.07681216299533844, + 0.10469955205917358, + 0.07348014414310455 + ], + "credit_map": [ + 0.053642380982637405, + 0.03528820350766182, + 0.022612670436501503, + 0.028564246371388435, + 0.024946685880422592, + 0.016351884230971336, + 0.020459134131669998, + 0.01849016174674034, + 0.012679310515522957, + 0.012425447814166546, + 0.01026246976107359, + 0.018521850928664207, + 0.010600459761917591, + 0.008281925693154335, + 0.007567967288196087, + 0.005423821043223143, + 0.01160397008061409, + 0.007823440246284008, + 0.008646819740533829, + 0.008337010629475117, + 0.005595046561211348, + 0.006383497733622789, + 0.007437645457684994, + 0.008531996048986912, + 0.05181315541267395, + 0.02927597612142563, + 0.02452450804412365, + 0.024348655715584755, + 0.02022172138094902, + 0.021903127431869507, + 0.012967108748853207, + 0.01804434508085251, + 0.004705168306827545, + 0.0038185063749551773, + 0.003291748696938157, + 0.0026622526347637177, + 0.0030015285592526197, + 0.003494597040116787, + 0.0019359465222805738, + 0.0030471864156425, + 0.00785928312689066, + 0.008439346216619015, + 0.007477172184735537, + 0.010599266737699509, + 0.005282259080559015, + 0.008668828755617142, + 0.005357309710234404, + 0.010468159802258015, + 0.00937903393059969, + 0.008106665685772896, + 0.009600689634680748, + 0.00536912027746439, + 0.007723173126578331, + 0.013764023780822754, + 0.007511553354561329, + 0.019868653267621994, + 0.008640549145638943, + 0.006567663513123989, + 0.01053859107196331, + 0.009055692702531815, + 0.009295269846916199, + 0.015688059851527214, + 0.008444116450846195, + 0.008582220412790775, + 0.01787271909415722, + 0.014336349442601204, + 0.011955690570175648, + 0.020856261253356934, + 0.008198649622499943, + 0.008567500859498978, + 0.009636221453547478, + 0.013276162557303905, + 0.00848905649036169, + 0.010100633837282658, + 0.008446571417152882, + 0.010663003660738468, + 0.00873530376702547, + 0.007230789866298437, + 0.011000207625329494, + 0.008814570493996143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -1.1370488405227661 + } + ], + "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/0000125/group_15/index_tensor([1733, 1733])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_15/index_tensor([1733, 1733])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ad8182438eb44387a649ec77f399c30ddaa65e54 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_15/index_tensor([1733, 1733])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot, slightly overlooking, slowly zooms out, bringing us into a bedroom in a fresh, softly styled Huizhou-style building. The first rays of morning's soft, golden sunlight filter through an intricately carved wooden window lattice, casting a dappled, warm light across the room. Motes of dust seem to dance quietly in the light. The camera focuses on a rustic, carved wooden bed. A young woman, dressed in a simple white cotton nightgown, lies sprawled on her side, her long, dark hair casually scattered like silk on the soft, white cotton pillow. A ray of light falls upon her delicate face, casting a warm glow over her skin. Her long, curled eyelashes begin to flutter slightly. Then, ever so slowly, her eyelids open, revealing a pair of brown eyes, initially hazy with sleep, but quickly becoming clear and bright, silently reflecting the light from the window pane. She narrowed her eyes slightly, as if silently feeling the tranquility. The whole awakening process was slow and leisurely, and the picture was filled with an atmosphere of tranquility and peace.", + "seed": 23493, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_15/index_1733_seed_23493_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": -0.2797658054623753, + "vlm_logprob": -0.2797658054623753, + "vlm_yes_no_margin": 2.974999988079071, + "vlm_token_credit": -0.2797658054623753, + "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 inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.11144223064184189 + ], + "opposite_logprob_tokens": [ + -2.7364423274993896 + ], + "logprob_reward": -0.11144223064184189, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.11144223064184189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13893768191337585, + 0.18160909414291382, + 0.1400345265865326, + 0.16020600497722626, + 0.06651999801397324, + 0.04761036857962608, + 0.11084046959877014, + 0.033481210470199585, + 0.0804748386144638, + 0.04028576985001564 + ], + "credit_map": [ + 0.011672881431877613, + 0.014567441307008266, + 0.01850414276123047, + 0.02780729904770851, + 0.014958578161895275, + 0.018730403855443, + 0.01960383728146553, + 0.013093098066747189, + 0.020163822919130325, + 0.024924393743276596, + 0.022571751847863197, + 0.027758995071053505, + 0.027036527171730995, + 0.02882690727710724, + 0.016428742557764053, + 0.013897934928536415, + 0.010658316314220428, + 0.015380816534161568, + 0.03366479277610779, + 0.011486238799989223, + 0.0322718620300293, + 0.007749513722956181, + 0.014511273242533207, + 0.014311714097857475, + 0.01818479783833027, + 0.02357650361955166, + 0.018588876351714134, + 0.013394289650022984, + 0.015054842457175255, + 0.039638057351112366, + 0.01567094400525093, + 0.016097694635391235, + 0.00833851657807827, + 0.007815453223884106, + 0.008766058832406998, + 0.009590895846486092, + 0.004629047121852636, + 0.006442319136112928, + 0.009755735285580158, + 0.011181969195604324, + 0.00474370364099741, + 0.005528988782316446, + 0.005105160642415285, + 0.006895159371197224, + 0.005753546487540007, + 0.0051990351639688015, + 0.007495995610952377, + 0.006888780277222395, + 0.007918287999927998, + 0.01155002973973751, + 0.01686418615281582, + 0.017125587910413742, + 0.010584705509245396, + 0.00824816059321165, + 0.03127951920032501, + 0.00726999482139945, + 0.003227334702387452, + 0.0019394087139517069, + 0.0035156963858753443, + 0.005865972023457289, + 0.0024036585818976164, + 0.0029288872610777617, + 0.0045896489173173904, + 0.00901060365140438, + 0.005388506222516298, + 0.00591420428827405, + 0.010540653020143509, + 0.01692054606974125, + 0.004579464439302683, + 0.004886416252702475, + 0.009560630656778812, + 0.022684423252940178, + 0.0021099168807268143, + 0.0010390921961516142, + 0.0027247017715126276, + 0.0027802588883787394, + 0.0017656575655564666, + 0.0018681121291592717, + 0.013889221474528313, + 0.014108807779848576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.11144223064184189 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.002532015787437558 + ], + "opposite_logprob_tokens": [ + -6.002532005310059 + ], + "logprob_reward": -0.002532015787437558, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002532015787437558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09388719499111176, + 0.11772484332323074, + 0.06573818624019623, + 0.11526764184236526, + 0.0318472646176815, + 0.07724016159772873, + 0.07403154671192169, + 0.2133353054523468, + 0.105096735060215, + 0.10583110898733139 + ], + "credit_map": [ + 0.010567059740424156, + 0.013032212853431702, + 0.010329670272767544, + 0.009628023020923138, + 0.011850365437567234, + 0.01678680256009102, + 0.014239976182579994, + 0.007453096564859152, + 0.013533166609704494, + 0.01436447910964489, + 0.011930721811950207, + 0.0208992101252079, + 0.01468109991401434, + 0.01636485382914543, + 0.013679471798241138, + 0.012271856889128685, + 0.008598189800977707, + 0.007109966594725847, + 0.009634620510041714, + 0.008517266251146793, + 0.005243134684860706, + 0.004714094568043947, + 0.009319642558693886, + 0.0126012759283185, + 0.011864801868796349, + 0.014789117500185966, + 0.013113247230648994, + 0.014345947653055191, + 0.01140980701893568, + 0.023916075006127357, + 0.012766227126121521, + 0.013062425889074802, + 0.0032332714181393385, + 0.002827640622854233, + 0.004682208877056837, + 0.003839674172922969, + 0.0027210586704313755, + 0.0024503893218934536, + 0.007805821020156145, + 0.004287202842533588, + 0.007110261358320713, + 0.007013489026576281, + 0.00688214460387826, + 0.011686577461659908, + 0.012770825996994972, + 0.009890671819448471, + 0.014118043705821037, + 0.007768148556351662, + 0.00656370772048831, + 0.0037922384217381477, + 0.008042631670832634, + 0.00807932112365961, + 0.005392125342041254, + 0.006967965979129076, + 0.022480487823486328, + 0.012713080272078514, + 0.015360632911324501, + 0.010977495461702347, + 0.024916158989071846, + 0.06847947090864182, + 0.01431445311754942, + 0.017038021236658096, + 0.04517129808664322, + 0.017077792435884476, + 0.01190861128270626, + 0.008545535616576672, + 0.017983848229050636, + 0.01409193780273199, + 0.01168791949748993, + 0.013844828121364117, + 0.016145631670951843, + 0.010888427495956421, + 0.008229275234043598, + 0.004929207731038332, + 0.009659644216299057, + 0.01831836998462677, + 0.0052650319412350655, + 0.006361362524330616, + 0.018166691064834595, + 0.034901537001132965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.002532015787437558 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.0181918703019619 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181918703019619, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181918703019619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07312008738517761, + 0.0637022852897644, + 0.03885851800441742, + 0.03526805713772774, + 0.027161410078406334, + 0.13651029765605927, + 0.14513666927814484, + 0.04736281558871269, + 0.3670010268688202, + 0.06587883830070496 + ], + "credit_map": [ + 0.009843043982982635, + 0.009358315728604794, + 0.010934190824627876, + 0.00952161941677332, + 0.008462787605822086, + 0.010825513862073421, + 0.008340494707226753, + 0.005834117531776428, + 0.0062849922105669975, + 0.007787689100950956, + 0.009863266721367836, + 0.010342992842197418, + 0.008386071771383286, + 0.009137379005551338, + 0.006347498390823603, + 0.005552393849939108, + 0.002610772615298629, + 0.005655974615365267, + 0.0075531513430178165, + 0.0031726183369755745, + 0.009614520706236362, + 0.002810824429616332, + 0.004514692816883326, + 0.0029259633738547564, + 0.0024931428488343954, + 0.0024961638264358044, + 0.004827183671295643, + 0.0077345059253275394, + 0.003288933075964451, + 0.0023197431582957506, + 0.007522183936089277, + 0.004586201626807451, + 0.0026590251363813877, + 0.0026362386997789145, + 0.003903825767338276, + 0.002879393519833684, + 0.0022898223251104355, + 0.0029625247698277235, + 0.007166733965277672, + 0.0026638456620275974, + 0.014877068810164928, + 0.014779971912503242, + 0.025194451212882996, + 0.0192889217287302, + 0.011385270394384861, + 0.014706136658787727, + 0.01941482350230217, + 0.016863655298948288, + 0.011053239926695824, + 0.011151718907058239, + 0.01988316886126995, + 0.022467903792858124, + 0.010197078809142113, + 0.010898814536631107, + 0.0391954742372036, + 0.020289268344640732, + 0.0028692055493593216, + 0.004000989254564047, + 0.0057202475145459175, + 0.013048609718680382, + 0.0042950003407895565, + 0.002958089578896761, + 0.01020751055330038, + 0.004263162612915039, + 0.06415051966905594, + 0.03325615078210831, + 0.06877287477254868, + 0.042277030646800995, + 0.02714574709534645, + 0.030855506658554077, + 0.0584484301507473, + 0.04209475219249725, + 0.0029813945293426514, + 0.0020825613755732775, + 0.00493262056261301, + 0.006929819937795401, + 0.0054990374483168125, + 0.0031197560019791126, + 0.021536385640501976, + 0.018797267228364944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.0181918703019619 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.04305894672870636 + ], + "opposite_logprob_tokens": [ + -3.1680588722229004 + ], + "logprob_reward": -0.04305894672870636, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305894672870636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06506043672561646, + 0.0560535304248333, + 0.08347249776124954, + 0.11539030075073242, + 0.06600996106863022, + 0.11553622782230377, + 0.07177092134952545, + 0.16232101619243622, + 0.08912794291973114, + 0.1752571314573288 + ], + "credit_map": [ + 0.004912594333291054, + 0.006363828666508198, + 0.005748758092522621, + 0.009941382333636284, + 0.008001222275197506, + 0.008716167882084846, + 0.015461890958249569, + 0.0059145912528038025, + 0.0053918142803013325, + 0.007136122789233923, + 0.0059132822789251804, + 0.008076492697000504, + 0.009311257861554623, + 0.008024025708436966, + 0.007365216035395861, + 0.004835318308323622, + 0.007058087736368179, + 0.009940115734934807, + 0.014101456850767136, + 0.010316045954823494, + 0.010073039680719376, + 0.009642571210861206, + 0.010250129736959934, + 0.012091049924492836, + 0.011283153668045998, + 0.015665531158447266, + 0.014179970137774944, + 0.009978480637073517, + 0.011500528082251549, + 0.02853548154234886, + 0.01248276885598898, + 0.011764387600123882, + 0.006691130809485912, + 0.005924252327531576, + 0.007966972887516022, + 0.008173681795597076, + 0.004805287811905146, + 0.007619301788508892, + 0.014918290078639984, + 0.009911040775477886, + 0.011055013164877892, + 0.00929785892367363, + 0.02212134189903736, + 0.021443650126457214, + 0.00760437548160553, + 0.010904183611273766, + 0.013657734729349613, + 0.019452080130577087, + 0.008717846125364304, + 0.00716618075966835, + 0.011117440648376942, + 0.011923864483833313, + 0.007435469422489405, + 0.00709898304194212, + 0.009482334367930889, + 0.00882880762219429, + 0.010284765623509884, + 0.007069761864840984, + 0.022486593574285507, + 0.060396719723939896, + 0.008336459286510944, + 0.006367791909724474, + 0.03862081095576286, + 0.008758108131587505, + 0.013563711196184158, + 0.007173158694058657, + 0.00957010593265295, + 0.01984640769660473, + 0.006533687934279442, + 0.006723236758261919, + 0.012407522648572922, + 0.013310112990438938, + 0.009463869035243988, + 0.0069929431192576885, + 0.010083607397973537, + 0.011303444392979145, + 0.00803473312407732, + 0.0069516184739768505, + 0.05392281338572502, + 0.06850410997867584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.04305894672870636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -1.2236039638519287 + ], + "opposite_logprob_tokens": [ + -0.34860390424728394 + ], + "logprob_reward": -1.2236039638519287, + "yes_no_margin_reward": -0.8750000596046448, + "token_credit_reward": -1.2236039638519287, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06006603315472603, + 0.0632048025727272, + 0.11379486322402954, + 0.15505912899971008, + 0.15770943462848663, + 0.09035900980234146, + 0.04905354976654053, + 0.18518605828285217, + 0.10679005831480026, + 0.018777040764689445 + ], + "credit_map": [ + 0.006390179507434368, + 0.005831355694681406, + 0.00619658874347806, + 0.007700981106609106, + 0.007634522393345833, + 0.010264793410897255, + 0.01059858687222004, + 0.005449020769447088, + 0.00480868574231863, + 0.008519422262907028, + 0.0077019259333610535, + 0.015229415148496628, + 0.006889257580041885, + 0.004371936433017254, + 0.00938757136464119, + 0.006296588107943535, + 0.012520494870841503, + 0.024520309641957283, + 0.012619292363524437, + 0.011102664284408092, + 0.01618841476738453, + 0.013782386668026447, + 0.01000128872692585, + 0.013060013763606548, + 0.011064392514526844, + 0.01795588806271553, + 0.020396864041686058, + 0.02300863154232502, + 0.014471974223852158, + 0.02876068837940693, + 0.022860825061798096, + 0.016539858654141426, + 0.01281081885099411, + 0.006992573384195566, + 0.018867669627070427, + 0.017800381407141685, + 0.008876185864210129, + 0.011513599194586277, + 0.04522816836833954, + 0.03562003746628761, + 0.008459151722490788, + 0.007916711270809174, + 0.015440733171999454, + 0.014418658800423145, + 0.008592134341597557, + 0.009913022629916668, + 0.012050618417561054, + 0.013567970134317875, + 0.008681518957018852, + 0.007380266673862934, + 0.005892029497772455, + 0.0059936936013400555, + 0.0063127693720161915, + 0.004315092694014311, + 0.005136779043823481, + 0.005341398995369673, + 0.013106674887239933, + 0.00786193273961544, + 0.023711243644356728, + 0.06458523869514465, + 0.009845645166933537, + 0.009110781364142895, + 0.04607611522078514, + 0.010888436809182167, + 0.010445646941661835, + 0.00789697002619505, + 0.019344422966241837, + 0.015035645104944706, + 0.00941495131701231, + 0.006258807145059109, + 0.016666345298290253, + 0.021727275103330612, + 0.0009077226859517395, + 0.001024399884045124, + 0.001323451055213809, + 0.002248608972877264, + 0.0010673176730051637, + 0.000726575031876564, + 0.006272762548178434, + 0.005206204019486904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -1.2236039638519287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium close-up shot, slightly overlooking, slowly zooms out, bringing us into a bedroom in a fresh, softly styled Huizhou-style building. The first rays of morning's soft, golden sunlight filter through an intricately carved wooden window lattice, casting a dappled, warm light across the room. Motes of dust seem to dance quietly in the light. The camera focuses on a rustic, carved wooden bed. A young woman, dressed in a simple white cotton nightgown, lies sprawled on her side, her long, dark hair casually scattered like silk on the soft, white cotton pillow. A ray of light falls upon her delicate face, casting a warm glow over her skin. Her long, curled eyelashes begin to flutter slightly. Then, ever so slowly, her eyelids open, revealing a pair of brown eyes, initially hazy with sleep, but quickly becoming clear and bright, silently reflecting the light from the window pane. She narrowed her eyes slightly, as if silently feeling the tranquility. The whole awakening process was slow and leisurely, and the picture was filled with an atmosphere of tranquility and peace.", + "seed": 23494, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_15/index_1733_seed_23494_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": -0.19361906889826058, + "vlm_logprob": -0.19361906889826058, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.19361906889826058, + "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 inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.14934784173965454 + ], + "opposite_logprob_tokens": [ + -2.1493477821350098 + ], + "logprob_reward": -0.14934784173965454, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14934784173965454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08094124495983124, + 0.2516718804836273, + 0.07755052298307419, + 0.0783727616071701, + 0.08524350821971893, + 0.08995652943849564, + 0.05574765056371689, + 0.12113824486732483, + 0.056984126567840576, + 0.10239351540803909 + ], + "credit_map": [ + 0.008067999966442585, + 0.01098615676164627, + 0.008514396846294403, + 0.011952491477131844, + 0.007488025818020105, + 0.007489196956157684, + 0.014198311604559422, + 0.012244664132595062, + 0.015662504360079765, + 0.012829202227294445, + 0.03225398063659668, + 0.0680852085351944, + 0.028994807973504066, + 0.03323956951498985, + 0.028795810416340828, + 0.03181079402565956, + 0.006802673917263746, + 0.009215310215950012, + 0.01152803748846054, + 0.008844628930091858, + 0.0065684267319738865, + 0.006854694802314043, + 0.014742008410394192, + 0.012994743883609772, + 0.00557019142434001, + 0.008850332349538803, + 0.017740249633789062, + 0.012971635907888412, + 0.007437132764607668, + 0.004665574990212917, + 0.01137625239789486, + 0.009761395864188671, + 0.011271221563220024, + 0.010024727322161198, + 0.012593559920787811, + 0.007283411920070648, + 0.011596941389143467, + 0.0109456367790699, + 0.01117736380547285, + 0.01035064272582531, + 0.007662186399102211, + 0.01023551169782877, + 0.013337879441678524, + 0.008761453442275524, + 0.009573983028531075, + 0.011082409881055355, + 0.016078857704997063, + 0.013224253430962563, + 0.005771166644990444, + 0.006325636990368366, + 0.014851806685328484, + 0.013504000380635262, + 0.0036637240555137396, + 0.0034622771199792624, + 0.003617400536313653, + 0.004551637917757034, + 0.015515395440161228, + 0.0177751574665308, + 0.01828077621757984, + 0.015956368297338486, + 0.015705695375800133, + 0.011402818374335766, + 0.012283534742891788, + 0.01421850174665451, + 0.013531249947845936, + 0.008152236230671406, + 0.008549029007554054, + 0.007817474193871021, + 0.004662899300456047, + 0.004910395946353674, + 0.004647271241992712, + 0.004713567439466715, + 0.008300798013806343, + 0.010915481485426426, + 0.017704417929053307, + 0.013212802819907665, + 0.016877543181180954, + 0.01177606824785471, + 0.014295699074864388, + 0.009310705587267876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.14934784173965454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.00678366981446743 + ], + "opposite_logprob_tokens": [ + -5.006783485412598 + ], + "logprob_reward": -0.00678366981446743, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00678366981446743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06870546191930771, + 0.21963264048099518, + 0.049004893749952316, + 0.03703517094254494, + 0.03724098578095436, + 0.021074382588267326, + 0.2975558638572693, + 0.1238345056772232, + 0.0595744363963604, + 0.08634167909622192 + ], + "credit_map": [ + 0.0085020512342453, + 0.004262227565050125, + 0.006007680203765631, + 0.008281640708446503, + 0.003724241629242897, + 0.005976395215839148, + 0.014958705753087997, + 0.016992516815662384, + 0.024751851335167885, + 0.02113865502178669, + 0.02582857757806778, + 0.04735371470451355, + 0.021539244800806046, + 0.015384141355752945, + 0.028760135173797607, + 0.034876320511102676, + 0.004015257116407156, + 0.0077976747415959835, + 0.005375184118747711, + 0.00651491479948163, + 0.005388177465647459, + 0.00609269505366683, + 0.00640637893229723, + 0.00741460919380188, + 0.0024981622118502855, + 0.003685843199491501, + 0.008379489183425903, + 0.006877884268760681, + 0.0022257643286138773, + 0.0022852481342852116, + 0.006471643224358559, + 0.004611136391758919, + 0.0036201803013682365, + 0.0029921100940555334, + 0.005437210202217102, + 0.004836671520024538, + 0.00355003634467721, + 0.004084442742168903, + 0.0069598848931491375, + 0.005760451313108206, + 0.0014568888582289219, + 0.0015687169507145882, + 0.003575942711904645, + 0.002254874911159277, + 0.0020197059493511915, + 0.003954481333494186, + 0.00268552009947598, + 0.0035582527052611113, + 0.0290523748844862, + 0.03995316103100777, + 0.06901249289512634, + 0.06628762930631638, + 0.022428717464208603, + 0.021020671352744102, + 0.02513221465051174, + 0.02466859668493271, + 0.01600131206214428, + 0.019410300999879837, + 0.018961003050208092, + 0.016765518113970757, + 0.016085930168628693, + 0.010979793965816498, + 0.012849506922066212, + 0.012781144119799137, + 0.007790193427354097, + 0.007134715560823679, + 0.008136438205838203, + 0.009390573017299175, + 0.005514212418347597, + 0.006601549685001373, + 0.006951882969588041, + 0.008054870180785656, + 0.007352394051849842, + 0.01116885431110859, + 0.01387264858931303, + 0.010892510414123535, + 0.012190192937850952, + 0.011449392884969711, + 0.011316432617604733, + 0.008099254220724106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.00678366981446743 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.01108358334749937 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.01108358334749937, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108358334749937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062244247645139694, + 0.1518794149160385, + 0.05548691004514694, + 0.04330337047576904, + 0.05088453367352486, + 0.03965974226593971, + 0.17849911749362946, + 0.18566803634166718, + 0.0602397583425045, + 0.1721349060535431 + ], + "credit_map": [ + 0.007047424558550119, + 0.00462902057915926, + 0.0075533282943069935, + 0.008072403259575367, + 0.004219199065119028, + 0.0076905484311282635, + 0.009826375171542168, + 0.013205943629145622, + 0.01438975054770708, + 0.008360033854842186, + 0.015964943915605545, + 0.02258680947124958, + 0.02063102275133133, + 0.019728809595108032, + 0.0219236109405756, + 0.028294425457715988, + 0.0033671162091195583, + 0.006983109749853611, + 0.007453620433807373, + 0.0053998627699911594, + 0.005802992265671492, + 0.0060545653104782104, + 0.010518977418541908, + 0.00990666076540947, + 0.0035505485720932484, + 0.003711241064593196, + 0.006998366676270962, + 0.00500769866630435, + 0.003283088793978095, + 0.004694060888141394, + 0.008157768286764622, + 0.007900593802332878, + 0.004306142684072256, + 0.003960557747632265, + 0.00681524071842432, + 0.005408376455307007, + 0.0046149916015565395, + 0.005700915120542049, + 0.011273160576820374, + 0.00880514457821846, + 0.0024093864485621452, + 0.002049473114311695, + 0.007913191802799702, + 0.00415751151740551, + 0.0036344206891953945, + 0.006546501535922289, + 0.006076788995414972, + 0.006872464902698994, + 0.00966099463403225, + 0.0108876908197999, + 0.06378065049648285, + 0.053988710045814514, + 0.009767074137926102, + 0.008817755617201328, + 0.01042082253843546, + 0.011175418272614479, + 0.027536919340491295, + 0.02375834248960018, + 0.02964276447892189, + 0.026061201468110085, + 0.018353648483753204, + 0.014533433131873608, + 0.023197930306196213, + 0.02258378081023693, + 0.008145738393068314, + 0.007700026500970125, + 0.008173367008566856, + 0.01179971918463707, + 0.004995143506675959, + 0.005183867644518614, + 0.006702134385704994, + 0.007539753802120686, + 0.015720821917057037, + 0.015162662602961063, + 0.026128074154257774, + 0.01501724123954773, + 0.04369058459997177, + 0.01849636435508728, + 0.022505128756165504, + 0.01541401632130146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.01108358334749937 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.043078359216451645 + ], + "opposite_logprob_tokens": [ + -3.1680784225463867 + ], + "logprob_reward": -0.043078359216451645, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043078359216451645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08055904507637024, + 0.1410374939441681, + 0.07753309607505798, + 0.0613272525370121, + 0.03710184246301651, + 0.03202233090996742, + 0.2956524193286896, + 0.0886070653796196, + 0.06591730564832687, + 0.12024208903312683 + ], + "credit_map": [ + 0.007586002815514803, + 0.005771417636424303, + 0.010135169140994549, + 0.008268051780760288, + 0.004605165217071772, + 0.010907081887125969, + 0.014689153991639614, + 0.018597016111016273, + 0.010870009660720825, + 0.010478068143129349, + 0.014526105485856533, + 0.025805775076150894, + 0.01608629710972309, + 0.0256311297416687, + 0.016958169639110565, + 0.02068195678293705, + 0.007319136522710323, + 0.009417261928319931, + 0.011650925502181053, + 0.010814577341079712, + 0.006644259672611952, + 0.006580553017556667, + 0.011962970718741417, + 0.013143418356776237, + 0.00419442867860198, + 0.00438581220805645, + 0.016819844022393227, + 0.011169376783072948, + 0.004050206858664751, + 0.003530655987560749, + 0.008463956415653229, + 0.008712979964911938, + 0.0028793616220355034, + 0.003169223666191101, + 0.0045401486568152905, + 0.006023497320711613, + 0.00346773536875844, + 0.0033164264168590307, + 0.006952567491680384, + 0.006752887275069952, + 0.0023473810870200396, + 0.0034624875988811255, + 0.004068928770720959, + 0.003711876692250371, + 0.0028895854484289885, + 0.005237895995378494, + 0.004843384958803654, + 0.005460797343403101, + 0.02983012981712818, + 0.04153800755739212, + 0.06658957898616791, + 0.06311012804508209, + 0.023955143988132477, + 0.02113794907927513, + 0.025833580642938614, + 0.02365793287754059, + 0.011437020264565945, + 0.012849608436226845, + 0.015660908073186874, + 0.010544206015765667, + 0.013268101029098034, + 0.008739123120903969, + 0.0077243708074092865, + 0.008383732289075851, + 0.01835358887910843, + 0.009373420849442482, + 0.009719287045300007, + 0.005446851719170809, + 0.0064473808743059635, + 0.004574863705784082, + 0.006337075959891081, + 0.005664849188178778, + 0.01290158275514841, + 0.00886708963662386, + 0.018941421061754227, + 0.012579905800521374, + 0.027217058464884758, + 0.014035302214324474, + 0.012865014374256134, + 0.012834727764129639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.043078359216451645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.75780189037323 + ], + "opposite_logprob_tokens": [ + -0.63280189037323 + ], + "logprob_reward": -0.75780189037323, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.75780189037323, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.058251913636922836, + 0.12291470915079117, + 0.04936572536826134, + 0.05733082816004753, + 0.03746698796749115, + 0.03404917195439339, + 0.5237148404121399, + 0.04289473593235016, + 0.05103345960378647, + 0.022977685555815697 + ], + "credit_map": [ + 0.00814408715814352, + 0.0065119764767587185, + 0.006717204581946135, + 0.0077244979329407215, + 0.003715906059369445, + 0.00482555478811264, + 0.011794420890510082, + 0.008818265050649643, + 0.010306361131370068, + 0.009640287607908249, + 0.012577651999890804, + 0.021530881524086, + 0.013059605844318867, + 0.021265191957354546, + 0.015290902927517891, + 0.01924382522702217, + 0.005821657832711935, + 0.005668109282851219, + 0.006400308571755886, + 0.007845932617783546, + 0.0038865492679178715, + 0.003905143588781357, + 0.00792083702981472, + 0.007917189039289951, + 0.003934761509299278, + 0.0043571810238063335, + 0.01796366460621357, + 0.013676656410098076, + 0.0038755733985453844, + 0.004481203388422728, + 0.004189690575003624, + 0.004852096550166607, + 0.0032619377598166466, + 0.0037050300743430853, + 0.003783839289098978, + 0.0047151921316981316, + 0.0034024862106889486, + 0.004779245238751173, + 0.006731593515723944, + 0.0070876674726605415, + 0.0027272789739072323, + 0.003070847364142537, + 0.005816072225570679, + 0.003571673994883895, + 0.003448129165917635, + 0.003013209905475378, + 0.006057261023670435, + 0.006344700697809458, + 0.042730145156383514, + 0.059223588556051254, + 0.13997715711593628, + 0.12900486588478088, + 0.03544454649090767, + 0.03443554788827896, + 0.042851902544498444, + 0.04004707559943199, + 0.004856382496654987, + 0.0046958341263234615, + 0.006010726559907198, + 0.004412305541336536, + 0.006625565700232983, + 0.005018371623009443, + 0.005555087700486183, + 0.00572045985609293, + 0.010530710220336914, + 0.00721136387437582, + 0.006282016169279814, + 0.007251047994941473, + 0.005444606300443411, + 0.0041849007830023766, + 0.004845737479627132, + 0.005283075850456953, + 0.002353517571464181, + 0.0024027733597904444, + 0.0033809226006269455, + 0.002274692989885807, + 0.003794498974457383, + 0.0031892796978354454, + 0.0026572297792881727, + 0.002924771048128605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.75780189037323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium close-up shot, slightly overlooking, slowly zooms out, bringing us into a bedroom in a fresh, softly styled Huizhou-style building. The first rays of morning's soft, golden sunlight filter through an intricately carved wooden window lattice, casting a dappled, warm light across the room. Motes of dust seem to dance quietly in the light. The camera focuses on a rustic, carved wooden bed. A young woman, dressed in a simple white cotton nightgown, lies sprawled on her side, her long, dark hair casually scattered like silk on the soft, white cotton pillow. A ray of light falls upon her delicate face, casting a warm glow over her skin. Her long, curled eyelashes begin to flutter slightly. Then, ever so slowly, her eyelids open, revealing a pair of brown eyes, initially hazy with sleep, but quickly becoming clear and bright, silently reflecting the light from the window pane. She narrowed her eyes slightly, as if silently feeling the tranquility. The whole awakening process was slow and leisurely, and the picture was filled with an atmosphere of tranquility and peace.", + "seed": 23495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_15/index_1733_seed_23495_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": -0.2070770390331745, + "vlm_logprob": -0.2070770390331745, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.2070770390331745, + "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 inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.23852190375328064 + ], + "opposite_logprob_tokens": [ + -1.613521933555603 + ], + "logprob_reward": -0.23852190375328064, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23852190375328064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47724515199661255, + 0.08102767169475555, + 0.1178063303232193, + 0.04894183576107025, + 0.04742614924907684, + 0.048222508281469345, + 0.029385264962911606, + 0.07062263786792755, + 0.041539501398801804, + 0.037783022969961166 + ], + "credit_map": [ + 0.02914820797741413, + 0.11219721287488937, + 0.0370890274643898, + 0.03639434278011322, + 0.03601262718439102, + 0.015751129016280174, + 0.17464765906333923, + 0.036004871129989624, + 0.00710637541487813, + 0.010466044768691063, + 0.010930877178907394, + 0.010029474273324013, + 0.009467688389122486, + 0.008386369794607162, + 0.013831772841513157, + 0.010809059254825115, + 0.00988790299743414, + 0.008206991478800774, + 0.011574587784707546, + 0.030610421672463417, + 0.007177023217082024, + 0.0058537316508591175, + 0.012919011525809765, + 0.03157665207982063, + 0.0037391064688563347, + 0.005525992251932621, + 0.005711264442652464, + 0.00590956537052989, + 0.003983612637966871, + 0.004588480107486248, + 0.010680129751563072, + 0.008803680539131165, + 0.0034393006935715675, + 0.01221082266420126, + 0.00508160050958395, + 0.005280735436826944, + 0.003465650137513876, + 0.005204296205192804, + 0.006002242676913738, + 0.006741494871675968, + 0.0055283429101109505, + 0.0033870225306600332, + 0.00842008925974369, + 0.004441145807504654, + 0.00693119689822197, + 0.005017777439206839, + 0.00805473979562521, + 0.0064421906135976315, + 0.0017928786110132933, + 0.003882149001583457, + 0.0038865183014422655, + 0.003984216600656509, + 0.0034713358618319035, + 0.0042523532174527645, + 0.004412234760820866, + 0.003703574649989605, + 0.003906457684934139, + 0.004834682680666447, + 0.014654201455414295, + 0.005588305648416281, + 0.01089316513389349, + 0.008264817297458649, + 0.005892347544431686, + 0.01658865623176098, + 0.0034704869613051414, + 0.008171862922608852, + 0.005022979341447353, + 0.007711813319474459, + 0.0030278502963483334, + 0.0024717538617551327, + 0.006026313174515963, + 0.005636438727378845, + 0.0031105573289096355, + 0.00553680956363678, + 0.006476628128439188, + 0.004538334906101227, + 0.0026146082673221827, + 0.002465213881805539, + 0.006925949826836586, + 0.0061149196699261665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.23852190375328064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.006768277380615473 + ], + "opposite_logprob_tokens": [ + -5.006768226623535 + ], + "logprob_reward": -0.006768277380615473, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006768277380615473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3256019949913025, + 0.05438341572880745, + 0.13498538732528687, + 0.04798285290598869, + 0.09140429645776749, + 0.036933910101652145, + 0.03447265177965164, + 0.07135763019323349, + 0.12879984080791473, + 0.07407806813716888 + ], + "credit_map": [ + 0.02372918650507927, + 0.060950662940740585, + 0.03248238563537598, + 0.031430523842573166, + 0.031699344515800476, + 0.018838128075003624, + 0.09687244892120361, + 0.029599273577332497, + 0.004951779264956713, + 0.007062072865664959, + 0.006295666564255953, + 0.006307725328952074, + 0.004658462479710579, + 0.004208168014883995, + 0.011874924413859844, + 0.009024612605571747, + 0.01474341843277216, + 0.020513545721769333, + 0.02555239573121071, + 0.01993410475552082, + 0.010031421668827534, + 0.009133454412221909, + 0.01854562945663929, + 0.01653141900897026, + 0.004769324325025082, + 0.0046286871656775475, + 0.00738521758466959, + 0.006191100925207138, + 0.003542470745742321, + 0.004668346140533686, + 0.007121582515537739, + 0.009676116518676281, + 0.004485508427023888, + 0.024932030588388443, + 0.006131700240075588, + 0.010303869843482971, + 0.0089943902567029, + 0.01280289702117443, + 0.011453804560005665, + 0.012300084345042706, + 0.0027257869951426983, + 0.007737280800938606, + 0.003487244714051485, + 0.004040814004838467, + 0.005078279413282871, + 0.004239578265696764, + 0.005431611090898514, + 0.0041933124884963036, + 0.002337291371077299, + 0.003585988888517022, + 0.005602850113064051, + 0.0048154243268072605, + 0.0026662847958505154, + 0.004138730000704527, + 0.0055904509499669075, + 0.005735628306865692, + 0.006556153763085604, + 0.0073029822669923306, + 0.01957477070391178, + 0.0053755189292132854, + 0.003784983651712537, + 0.0032481111120432615, + 0.0071590496227145195, + 0.018356051295995712, + 0.007062551565468311, + 0.007338077761232853, + 0.04485609009861946, + 0.009412907063961029, + 0.003905544290319085, + 0.006056190002709627, + 0.00866144709289074, + 0.041507016867399216, + 0.006335257552564144, + 0.008117049001157284, + 0.009000258520245552, + 0.010221969336271286, + 0.005774033721536398, + 0.003784371307119727, + 0.01961921527981758, + 0.011225901544094086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.006768277380615473 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.005954384338110685 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954384338110685, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954384338110685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20499664545059204, + 0.0957995057106018, + 0.11780726164579391, + 0.048995014280080795, + 0.04331623390316963, + 0.07815799117088318, + 0.04118882119655609, + 0.056060630828142166, + 0.18330669403076172, + 0.13037115335464478 + ], + "credit_map": [ + 0.015226124785840511, + 0.04034677520394325, + 0.0204426571726799, + 0.016850987449288368, + 0.01961422525346279, + 0.010896626859903336, + 0.06560016423463821, + 0.016019092872738838, + 0.01658380962908268, + 0.008624930866062641, + 0.010044959373772144, + 0.010424237698316574, + 0.006743732839822769, + 0.005321063566952944, + 0.01481848768889904, + 0.023238282650709152, + 0.00487927533686161, + 0.009190688841044903, + 0.010931666940450668, + 0.029636433348059654, + 0.005840259604156017, + 0.006409592926502228, + 0.014582466334104538, + 0.03633687645196915, + 0.0072424751706421375, + 0.004871897399425507, + 0.008961420506238937, + 0.004942651372402906, + 0.004607927519828081, + 0.0028795949183404446, + 0.008120015263557434, + 0.007369033992290497, + 0.00498114712536335, + 0.003930919338017702, + 0.008860026486217976, + 0.005681303329765797, + 0.0035144640132784843, + 0.0033736987970769405, + 0.00824311375617981, + 0.004731559194624424, + 0.010180993936955929, + 0.01721731387078762, + 0.008196160197257996, + 0.007373586297035217, + 0.009284665808081627, + 0.005768654868006706, + 0.008838169276714325, + 0.01129844505339861, + 0.0034188840072602034, + 0.004267612937837839, + 0.007028970867395401, + 0.004527834709733725, + 0.004420315846800804, + 0.005136060528457165, + 0.005350283347070217, + 0.007038856856524944, + 0.00377318705432117, + 0.007740753702819347, + 0.011136817745864391, + 0.009001083672046661, + 0.004639369901269674, + 0.005138796754181385, + 0.007657073903828859, + 0.006973547860980034, + 0.010283066891133785, + 0.017247645184397697, + 0.05852888897061348, + 0.012913438491523266, + 0.00727456109598279, + 0.009394684806466103, + 0.011489557102322578, + 0.0561748631298542, + 0.009323237463831902, + 0.010781854391098022, + 0.01809941604733467, + 0.011280456557869911, + 0.010838552378118038, + 0.015239984728395939, + 0.030167633667588234, + 0.02464001625776291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.005954384338110685 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.02634863369166851 + ], + "opposite_logprob_tokens": [ + -3.65134859085083 + ], + "logprob_reward": -0.02634863369166851, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634863369166851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3395797312259674, + 0.06577201932668686, + 0.15218350291252136, + 0.03764089569449425, + 0.06079975888133049, + 0.12237457185983658, + 0.05243486166000366, + 0.048903368413448334, + 0.08444640785455704, + 0.03586477413773537 + ], + "credit_map": [ + 0.018430903553962708, + 0.09020216017961502, + 0.02410203218460083, + 0.02183312550187111, + 0.023466290906071663, + 0.009386876598000526, + 0.12639619410037994, + 0.025762196630239487, + 0.008846237324178219, + 0.00850585661828518, + 0.008165267296135426, + 0.009547362104058266, + 0.005723270587623119, + 0.003357995767146349, + 0.012570011429488659, + 0.009056031703948975, + 0.007926881313323975, + 0.011616618372499943, + 0.012522739358246326, + 0.04263835772871971, + 0.005991215817630291, + 0.007600297685712576, + 0.015576262027025223, + 0.04831116646528244, + 0.0025244064163416624, + 0.0030652210116386414, + 0.005350756458938122, + 0.004435613751411438, + 0.0032234659884124994, + 0.002948480425402522, + 0.0068648140877485275, + 0.009228143841028214, + 0.005113333463668823, + 0.006482281722128391, + 0.0044391523115336895, + 0.0057985009625554085, + 0.0052793268114328384, + 0.009401073679327965, + 0.013360699638724327, + 0.010925399139523506, + 0.014432662166655064, + 0.020367713645100594, + 0.013915834948420525, + 0.012066725641489029, + 0.013210557401180267, + 0.008167924359440804, + 0.018720772117376328, + 0.021492401137948036, + 0.0039350856095552444, + 0.006093135103583336, + 0.009766537696123123, + 0.005556571763008833, + 0.006920520681887865, + 0.006883108988404274, + 0.00501887546852231, + 0.008261030539870262, + 0.003212605370208621, + 0.004883029963821173, + 0.007901431992650032, + 0.005545720458030701, + 0.00823687482625246, + 0.005448704585433006, + 0.0052422694861888885, + 0.008432743139564991, + 0.005056323949247599, + 0.007892918772995472, + 0.027162425220012665, + 0.006422161590307951, + 0.003100368194282055, + 0.005354838911443949, + 0.0035581071861088276, + 0.025899283587932587, + 0.003158580046147108, + 0.006162196397781372, + 0.0032877158373594284, + 0.00490203220397234, + 0.0028900548350065947, + 0.0016616520006209612, + 0.008316791616380215, + 0.005485753528773785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.02634863369166851 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.7577919960021973 + ], + "opposite_logprob_tokens": [ + -0.6327919960021973 + ], + "logprob_reward": -0.7577919960021973, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577919960021973, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2759687006473541, + 0.08272693306207657, + 0.31516799330711365, + 0.07921331375837326, + 0.0697503313422203, + 0.02532983385026455, + 0.02256028540432453, + 0.031219439581036568, + 0.07890293747186661, + 0.019160225987434387 + ], + "credit_map": [ + 0.018420612439513206, + 0.06432382762432098, + 0.020061179995536804, + 0.018759412690997124, + 0.021709850057959557, + 0.009222157299518585, + 0.08980125188827515, + 0.03367040678858757, + 0.00815048348158598, + 0.012714283540844917, + 0.008953756652772427, + 0.011569870635867119, + 0.010945769026875496, + 0.004680352751165628, + 0.015303193591535091, + 0.010409224778413773, + 0.013236191123723984, + 0.0190422423183918, + 0.022909464314579964, + 0.1051052138209343, + 0.00613771378993988, + 0.013094473630189896, + 0.023322582244873047, + 0.11232014000415802, + 0.00567094748839736, + 0.010762667283415794, + 0.010096929967403412, + 0.010921521112322807, + 0.005328675732016563, + 0.006266025826334953, + 0.016564948484301567, + 0.013601601123809814, + 0.004305395297706127, + 0.017372874543070793, + 0.007991502992808819, + 0.010227050632238388, + 0.005566331557929516, + 0.004263065289705992, + 0.011690855957567692, + 0.008333254605531693, + 0.002851237542927265, + 0.003408736316487193, + 0.004426421131938696, + 0.0030900235287845135, + 0.001629348611459136, + 0.0026773286517709494, + 0.003148699412122369, + 0.00409803818911314, + 0.002094835974276066, + 0.0017400316428393126, + 0.005740698426961899, + 0.0026200541760772467, + 0.0016544641694054008, + 0.0017850128933787346, + 0.0030352321919053793, + 0.0038899555802345276, + 0.0031272140331566334, + 0.002391727641224861, + 0.0068987333215773106, + 0.0021837397944182158, + 0.0038814309518784285, + 0.004436718299984932, + 0.003165675327181816, + 0.005134199280291796, + 0.004644561093300581, + 0.0033921170979738235, + 0.02875227853655815, + 0.0040393429808318615, + 0.002010930795222521, + 0.0035853716544806957, + 0.005273600574582815, + 0.027204737067222595, + 0.001243716455064714, + 0.0030458637047559023, + 0.0014040148817002773, + 0.002282195258885622, + 0.0015676519833505154, + 0.0014159523416310549, + 0.005343383178114891, + 0.002857448533177376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.7577919960021973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium close-up shot, slightly overlooking, slowly zooms out, bringing us into a bedroom in a fresh, softly styled Huizhou-style building. The first rays of morning's soft, golden sunlight filter through an intricately carved wooden window lattice, casting a dappled, warm light across the room. Motes of dust seem to dance quietly in the light. The camera focuses on a rustic, carved wooden bed. A young woman, dressed in a simple white cotton nightgown, lies sprawled on her side, her long, dark hair casually scattered like silk on the soft, white cotton pillow. A ray of light falls upon her delicate face, casting a warm glow over her skin. Her long, curled eyelashes begin to flutter slightly. Then, ever so slowly, her eyelids open, revealing a pair of brown eyes, initially hazy with sleep, but quickly becoming clear and bright, silently reflecting the light from the window pane. She narrowed her eyes slightly, as if silently feeling the tranquility. The whole awakening process was slow and leisurely, and the picture was filled with an atmosphere of tranquility and peace.", + "seed": 23496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_15/index_1733_seed_23496_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": -0.15018162997439505, + "vlm_logprob": -0.15018162997439505, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.15018162997439505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.1957172304391861 + ], + "opposite_logprob_tokens": [ + -1.945717215538025 + ], + "logprob_reward": -0.1957172304391861, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1957172304391861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07099667936563492, + 0.3113066256046295, + 0.18084533512592316, + 0.07064593583345413, + 0.03573956713080406, + 0.02348262444138527, + 0.053788866847753525, + 0.03373777121305466, + 0.09702499955892563, + 0.12243161350488663 + ], + "credit_map": [ + 0.007882815785706043, + 0.010074061341583729, + 0.012071005068719387, + 0.010045839473605156, + 0.015109837055206299, + 0.005430628079921007, + 0.006053891032934189, + 0.004328599199652672, + 0.06522823870182037, + 0.028523407876491547, + 0.08502314239740372, + 0.0456014946103096, + 0.020321855321526527, + 0.015568433329463005, + 0.030602071434259415, + 0.020437980070710182, + 0.028352227061986923, + 0.01039084792137146, + 0.03318403661251068, + 0.03011489100754261, + 0.011182951740920544, + 0.011259544640779495, + 0.03402552381157875, + 0.022335313260555267, + 0.005942927673459053, + 0.00800357200205326, + 0.011664590798318386, + 0.005241990555077791, + 0.01055906806141138, + 0.008483914658427238, + 0.006420844234526157, + 0.014329031109809875, + 0.0026567752938717604, + 0.002861443907022476, + 0.005685950629413128, + 0.004149147309362888, + 0.004567496012896299, + 0.0059301299042999744, + 0.004517565947026014, + 0.005371054634451866, + 0.001979610649868846, + 0.002274226862937212, + 0.0037279571406543255, + 0.002694998634979129, + 0.001905149663798511, + 0.002783737611025572, + 0.003824169747531414, + 0.004292772151529789, + 0.0040521807968616486, + 0.00739214988425374, + 0.005846781190484762, + 0.007544682361185551, + 0.008793259039521217, + 0.009396480396389961, + 0.004055619239807129, + 0.006707714404910803, + 0.0028645822312682867, + 0.004535906482487917, + 0.0028723878785967827, + 0.004295228514820337, + 0.004116413649171591, + 0.0037485116627067327, + 0.003154940204694867, + 0.008149798959493637, + 0.008386427536606789, + 0.012045682407915592, + 0.01071907952427864, + 0.011116270907223225, + 0.01308745052665472, + 0.013367054983973503, + 0.009871916845440865, + 0.01843111403286457, + 0.009366299025714397, + 0.022181300446391106, + 0.012202776968479156, + 0.008359873667359352, + 0.00991366058588028, + 0.012441995553672314, + 0.008394880220293999, + 0.039570827037096024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.1957172304391861 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.0046735103242099285 + ], + "opposite_logprob_tokens": [ + -5.379673480987549 + ], + "logprob_reward": -0.0046735103242099285, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046735103242099285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06150301918387413, + 0.2218918800354004, + 0.10128581523895264, + 0.07352606952190399, + 0.06637591868638992, + 0.03156851977109909, + 0.2682134211063385, + 0.07733223587274551, + 0.03320089355111122, + 0.06510225683450699 + ], + "credit_map": [ + 0.004821342416107655, + 0.008259671740233898, + 0.009737427346408367, + 0.007784281857311726, + 0.01233989093452692, + 0.00592083903029561, + 0.0070711299777030945, + 0.005568438209593296, + 0.023614779114723206, + 0.0255128126591444, + 0.03606167808175087, + 0.04981909319758415, + 0.016008827835321426, + 0.013291790150105953, + 0.03577578440308571, + 0.02180710807442665, + 0.01360455434769392, + 0.006442656274884939, + 0.020186597481369972, + 0.020055219531059265, + 0.006023000925779343, + 0.00561138242483139, + 0.017403947189450264, + 0.011958454735577106, + 0.006324904505163431, + 0.0055353958159685135, + 0.01400549616664648, + 0.00985852163285017, + 0.005521554499864578, + 0.007693669758737087, + 0.013329158537089825, + 0.011257370933890343, + 0.005142224486917257, + 0.011313995346426964, + 0.006888379342854023, + 0.007070065476000309, + 0.004121103323996067, + 0.0049956319853663445, + 0.007936110720038414, + 0.01890840381383896, + 0.0018859636038541794, + 0.001457886421121657, + 0.006760152522474527, + 0.003891139989718795, + 0.0018944534240290523, + 0.0031078860629349947, + 0.005693758372217417, + 0.00687728077173233, + 0.012751471251249313, + 0.030943958088755608, + 0.03783693164587021, + 0.04673412814736366, + 0.019047755748033524, + 0.019756168127059937, + 0.025118382647633553, + 0.076024629175663, + 0.006052676122635603, + 0.011335745453834534, + 0.008198422379791737, + 0.009979142807424068, + 0.009414148516952991, + 0.009112623520195484, + 0.005027432460337877, + 0.01821204274892807, + 0.002622934989631176, + 0.004315472673624754, + 0.0060924263671040535, + 0.004364635795354843, + 0.004013086203485727, + 0.0036899379920214415, + 0.0028399229049682617, + 0.005262475460767746, + 0.005043139215558767, + 0.01234612986445427, + 0.007985164411365986, + 0.004654274322092533, + 0.005136718973517418, + 0.0049408501945436, + 0.004945345688611269, + 0.02005063369870186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.0046735103242099285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.014207287691533566 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207287691533566, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207287691533566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11295303702354431, + 0.1528993844985962, + 0.10841963440179825, + 0.08311723917722702, + 0.10171347111463547, + 0.0814938172698021, + 0.16481178998947144, + 0.06670079380273819, + 0.058570586144924164, + 0.0693201944231987 + ], + "credit_map": [ + 0.0106012262403965, + 0.015775559470057487, + 0.02025349996984005, + 0.014372318051755428, + 0.019951244816184044, + 0.012483392842113972, + 0.011267499066889286, + 0.008248308673501015, + 0.03361457958817482, + 0.01732822321355343, + 0.04420005530118942, + 0.019397204741835594, + 0.010491740889847279, + 0.006799379363656044, + 0.01210740115493536, + 0.008960828185081482, + 0.017857950180768967, + 0.005526257678866386, + 0.026941755786538124, + 0.011236987076699734, + 0.009958063252270222, + 0.010580770671367645, + 0.01420931238681078, + 0.012108544819056988, + 0.007969554513692856, + 0.010180116631090641, + 0.01203500758856535, + 0.008219810202717781, + 0.011050564236938953, + 0.012035056948661804, + 0.009734849445521832, + 0.011892286129295826, + 0.008134862408041954, + 0.008617500774562359, + 0.025440238416194916, + 0.013582763262093067, + 0.006084215361624956, + 0.010052569210529327, + 0.01721445471048355, + 0.012586877681314945, + 0.005462677218019962, + 0.007564852479845285, + 0.01226774975657463, + 0.009148717857897282, + 0.01182330772280693, + 0.010313824750483036, + 0.012288522906601429, + 0.012624173425137997, + 0.007848838344216347, + 0.024070097133517265, + 0.01841549202799797, + 0.024012744426727295, + 0.010232295840978622, + 0.012333502992987633, + 0.012104076333343983, + 0.05579476058483124, + 0.004974510986357927, + 0.009968369267880917, + 0.008809288032352924, + 0.007144473027437925, + 0.006217645946890116, + 0.006175125017762184, + 0.009894917719066143, + 0.013516468927264214, + 0.005595402792096138, + 0.007070617284625769, + 0.006366979796439409, + 0.005859421566128731, + 0.008382727392017841, + 0.007367123384028673, + 0.006406610365957022, + 0.011521709151566029, + 0.004510428290814161, + 0.006284669972956181, + 0.010629583150148392, + 0.011873927898705006, + 0.007215061224997044, + 0.008189446292817593, + 0.010716868564486504, + 0.009900214150547981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.014207287691533566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.06205455958843231 + ], + "opposite_logprob_tokens": [ + -2.8120546340942383 + ], + "logprob_reward": -0.06205455958843231, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205455958843231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11575452983379364, + 0.15732720494270325, + 0.1048973873257637, + 0.031153004616498947, + 0.23805353045463562, + 0.06365345418453217, + 0.0770353302359581, + 0.16716189682483673, + 0.024792656302452087, + 0.02017100527882576 + ], + "credit_map": [ + 0.010658911429345608, + 0.015302066691219807, + 0.017176026478409767, + 0.013414482586085796, + 0.025875616818666458, + 0.011974402703344822, + 0.01173565536737442, + 0.009617356583476067, + 0.028380664065480232, + 0.01566332019865513, + 0.035479363054037094, + 0.029383007436990738, + 0.007629588712006807, + 0.008124158717691898, + 0.020275229588150978, + 0.012391860596835613, + 0.015467612072825432, + 0.005900486838072538, + 0.02073069103062153, + 0.018135875463485718, + 0.006246436387300491, + 0.0064303274266421795, + 0.019646167755126953, + 0.012339785695075989, + 0.0020082988776266575, + 0.0025423055049031973, + 0.007599142845720053, + 0.0031935605220496655, + 0.0026661495212465525, + 0.003062339499592781, + 0.005082499701529741, + 0.0049987053498625755, + 0.018621427938342094, + 0.04843524470925331, + 0.03225108981132507, + 0.02241334319114685, + 0.01563241519033909, + 0.029970042407512665, + 0.028173305094242096, + 0.042556654661893845, + 0.004016803577542305, + 0.004442049656063318, + 0.0086948461830616, + 0.010418015532195568, + 0.006153040565550327, + 0.004785693250596523, + 0.017293885350227356, + 0.007849113084375858, + 0.004195429850369692, + 0.006727063097059727, + 0.010163714177906513, + 0.013662080280482769, + 0.007728515192866325, + 0.007825490087270737, + 0.007671410217881203, + 0.019061626866459846, + 0.013598996214568615, + 0.028883779421448708, + 0.01612960733473301, + 0.016793856397271156, + 0.013322374783456326, + 0.014743613079190254, + 0.017010200768709183, + 0.046679459512233734, + 0.0023418613709509373, + 0.003329809056594968, + 0.0028986067045480013, + 0.003687134012579918, + 0.0032029703725129366, + 0.0032555873040109873, + 0.001949915662407875, + 0.004126768559217453, + 0.0014375039609149098, + 0.0029221412260085344, + 0.002810228615999222, + 0.002075880067422986, + 0.0019893809221684933, + 0.001732716802507639, + 0.0019927441608160734, + 0.005210408940911293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.06205455958843231 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a bedroom within a Huizhou-style building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rustic, carved wooden bed visible in the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman wearing a simple white cotton nightgown lying on her side on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's action include slowly opening her eyes to reveal brown eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium close-up and slowly zoom 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": [ + -0.4742555618286133 + ], + "opposite_logprob_tokens": [ + -0.9742555618286133 + ], + "logprob_reward": -0.4742555618286133, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742555618286133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21491582691669464, + 0.12784595787525177, + 0.3709242641925812, + 0.14177566766738892, + 0.03611227869987488, + 0.015750549733638763, + 0.01866958476603031, + 0.013342433609068394, + 0.03197038546204567, + 0.02869298867881298 + ], + "credit_map": [ + 0.018189042806625366, + 0.025013308972120285, + 0.03786298632621765, + 0.025187168270349503, + 0.0464673712849617, + 0.019359851256012917, + 0.020727042108774185, + 0.02210906334221363, + 0.01852872036397457, + 0.012607702054083347, + 0.024494851008057594, + 0.028903627768158913, + 0.007646551355719566, + 0.006721363868564367, + 0.01848791167140007, + 0.010455231182277203, + 0.06735781580209732, + 0.02033214271068573, + 0.07621666043996811, + 0.050172120332717896, + 0.02165779285132885, + 0.020100805908441544, + 0.06350395083427429, + 0.05158296972513199, + 0.010440768674015999, + 0.013548215851187706, + 0.02271885797381401, + 0.016119159758090973, + 0.01677802950143814, + 0.014113477431237698, + 0.026111211627721786, + 0.02194594405591488, + 0.003143533831462264, + 0.0024549514055252075, + 0.0076688118278980255, + 0.005089772865176201, + 0.002684518927708268, + 0.0042429338209331036, + 0.0062408908270299435, + 0.004586865659803152, + 0.0012801470002159476, + 0.00177846010774374, + 0.002145320177078247, + 0.001822080696001649, + 0.0019083389779552817, + 0.0020480910316109657, + 0.002269051969051361, + 0.0024990581441670656, + 0.0015057837590575218, + 0.0023646815679967403, + 0.002798667410388589, + 0.0028450952377170324, + 0.0019614507909864187, + 0.0021367473527789116, + 0.0018173244316130877, + 0.0032398339826613665, + 0.0009335516369901597, + 0.001249609631486237, + 0.0024629735853523016, + 0.00170590088237077, + 0.001430381671525538, + 0.0012008482590317726, + 0.0021047929767519236, + 0.0022543747909367085, + 0.002564292401075363, + 0.003665270283818245, + 0.004239493049681187, + 0.004263597074896097, + 0.004496149718761444, + 0.004594047088176012, + 0.004089014604687691, + 0.004058518912643194, + 0.0024590531829744577, + 0.004311985801905394, + 0.0041175708174705505, + 0.002803696086630225, + 0.0023360245395451784, + 0.004047435242682695, + 0.0029036260675638914, + 0.005713597871363163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.4742555618286133 + } + ], + "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/0000125/group_2/index_tensor([6853, 6853])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_2/index_tensor([6853, 6853])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ba62e7c5c16cf18ce41b128c2b889089c210eb32 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_2/index_tensor([6853, 6853])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle macro shot captures a sleek, modern hair dryer lying on its side on a pristine white marble countertop with subtle gray veining. The hair dryer's body is crafted from a sophisticated matte black, soft-touch material, accented by a polished rose gold ring surrounding the nozzle. Tied around this nozzle are several thin, luminous ribbons in vibrant magenta, electric blue, and sunny yellow, initially soft and still. Without warning, the hair dryer silently powers on, and the ribbons are instantly jolted to life by invisible jets of air. They begin to flutter violently, whipping back and forth in a chaotic yet graceful dance, their light fabric twisting and snapping, creating a mesmerizing, high-speed blur of colors that contrasts starkly with the appliance's inert, dark form. Soft, directional studio lighting captures the delicate sheen of the flowing silk, shimmering with rose gold accents, casting rapidly shifting, subtle shadows on the marble surface below, creating an atmosphere of vibrant energy and visual allure, unleashing power and aesthetic elegance.", + "seed": 10085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_2/index_6853_seed_10085_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": -0.054693245142698285, + "vlm_logprob": -0.054693245142698285, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.054693245142698285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03972100093960762 + ], + "opposite_logprob_tokens": [ + -3.4147210121154785 + ], + "logprob_reward": -0.03972100093960762, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03972100093960762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04486794024705887, + 0.07589167356491089, + 0.43514496088027954, + 0.12689942121505737, + 0.060054413974285126, + 0.03963405638933182, + 0.07685021311044693, + 0.061469510197639465, + 0.03776613250374794, + 0.04142167791724205 + ], + "credit_map": [ + 0.007288339547812939, + 0.006218729540705681, + 0.0059621911495924, + 0.007971328683197498, + 0.0048931366764009, + 0.0031825790647417307, + 0.00500913942232728, + 0.004342501983046532, + 0.011375196278095245, + 0.009892342612147331, + 0.009484327398240566, + 0.012675389647483826, + 0.007163878995925188, + 0.0064308419823646545, + 0.008063539862632751, + 0.010806163772940636, + 0.04447963833808899, + 0.1442413181066513, + 0.12691503763198853, + 0.033213112503290176, + 0.024301480501890182, + 0.010394391603767872, + 0.016762681305408478, + 0.03483731672167778, + 0.013969923369586468, + 0.024124713614583015, + 0.0139007568359375, + 0.014512920752167702, + 0.015042667277157307, + 0.010729199275374413, + 0.024750083684921265, + 0.009869156405329704, + 0.008144467137753963, + 0.008800152689218521, + 0.006265594623982906, + 0.011806553229689598, + 0.005356801673769951, + 0.007588149048388004, + 0.006097746547311544, + 0.0059949555434286594, + 0.004386451095342636, + 0.0036949897184967995, + 0.0069679077714681625, + 0.005947448778897524, + 0.0036038279067724943, + 0.006502355448901653, + 0.004819643218070269, + 0.0037114382721483707, + 0.022802583873271942, + 0.022737305611371994, + 0.009380795992910862, + 0.0063149831257760525, + 0.005753559060394764, + 0.004343947861343622, + 0.003101987997069955, + 0.002415055874735117, + 0.009214731864631176, + 0.008581329137086868, + 0.006105741020292044, + 0.010028299875557423, + 0.0038995391223579645, + 0.0064734481275081635, + 0.00818247813731432, + 0.008983946405351162, + 0.007715608458966017, + 0.0032813833095133305, + 0.012004876509308815, + 0.002982086269184947, + 0.0023161787539720535, + 0.004781658295542002, + 0.0026520946994423866, + 0.002032248768955469, + 0.013705841265618801, + 0.006071636453270912, + 0.004268735647201538, + 0.004153342451900244, + 0.003384672338142991, + 0.0025272981729358435, + 0.0021803348790854216, + 0.0051298197358846664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.03972100093960762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023331981152296066 + ], + "opposite_logprob_tokens": [ + -3.773331880569458 + ], + "logprob_reward": -0.023331981152296066, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023331981152296066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0943613350391388, + 0.13236528635025024, + 0.1524350643157959, + 0.38903459906578064, + 0.04783353954553604, + 0.02570720948278904, + 0.04336002841591835, + 0.04082484543323517, + 0.04716752842068672, + 0.026910535991191864 + ], + "credit_map": [ + 0.01211145706474781, + 0.013754112645983696, + 0.012459611520171165, + 0.013746798038482666, + 0.011028007604181767, + 0.00692822365090251, + 0.010185680352151394, + 0.01414745394140482, + 0.020953653380274773, + 0.013072404079139233, + 0.02482655458152294, + 0.022014593705534935, + 0.009956732392311096, + 0.012131965719163418, + 0.013279673643410206, + 0.016129720956087112, + 0.018304264172911644, + 0.0474306158721447, + 0.046640727669000626, + 0.010518031194806099, + 0.008073357865214348, + 0.003037509508430958, + 0.006566931027919054, + 0.01186364982277155, + 0.056841399520635605, + 0.08553691953420639, + 0.04109840840101242, + 0.04984075948596001, + 0.053667694330215454, + 0.0326436311006546, + 0.04392583668231964, + 0.02547995187342167, + 0.007432999555021524, + 0.008257543668150902, + 0.0077921561896800995, + 0.009965019300580025, + 0.0032921810634434223, + 0.0036738887429237366, + 0.003110242076218128, + 0.004309517331421375, + 0.0036654802970588207, + 0.002539609558880329, + 0.004330446012318134, + 0.0040405262261629105, + 0.0024728693533688784, + 0.003706117160618305, + 0.002715152222663164, + 0.0022370098158717155, + 0.015540092252194881, + 0.013929761946201324, + 0.0035816263407468796, + 0.002518946537747979, + 0.002292763441801071, + 0.0015242487424984574, + 0.0020383920054882765, + 0.001934204250574112, + 0.0057892161421477795, + 0.0062263417057693005, + 0.003934606444090605, + 0.005431306082755327, + 0.0034223198890686035, + 0.003959029912948608, + 0.0066813030280172825, + 0.00538072781637311, + 0.01534517202526331, + 0.004504316486418247, + 0.016007285565137863, + 0.00340177770704031, + 0.0030922777950763702, + 0.0014924131100997329, + 0.0018913743551820517, + 0.0014329178957268596, + 0.007011055015027523, + 0.0037900509778410196, + 0.002653856296092272, + 0.0025569519493728876, + 0.0033538159914314747, + 0.0032084297854453325, + 0.0016853464767336845, + 0.002651032991707325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.023331981152296066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108322784304619 + ], + "opposite_logprob_tokens": [ + -5.504108428955078 + ], + "logprob_reward": -0.004108322784304619, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108322784304619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05049210786819458, + 0.1371476799249649, + 0.5309121012687683, + 0.04982847720384598, + 0.04231572523713112, + 0.02365637943148613, + 0.0683518722653389, + 0.03108774870634079, + 0.04112539067864418, + 0.025082489475607872 + ], + "credit_map": [ + 0.005933510605245829, + 0.005818330682814121, + 0.006131038069725037, + 0.008554660715162754, + 0.003645966062322259, + 0.0052125160582363605, + 0.006473828107118607, + 0.008722261525690556, + 0.02075268141925335, + 0.015351259149610996, + 0.014898876659572124, + 0.021843258291482925, + 0.01222878135740757, + 0.013845602050423622, + 0.014435087330639362, + 0.023792125284671783, + 0.05040060356259346, + 0.18283632397651672, + 0.15929637849330902, + 0.031463079154491425, + 0.036718256771564484, + 0.008141942322254181, + 0.01951468735933304, + 0.042540837079286575, + 0.007323296740651131, + 0.008754365146160126, + 0.0052229915745556355, + 0.0062395078130066395, + 0.007203314453363419, + 0.0043850732035934925, + 0.007064592558890581, + 0.003635332453995943, + 0.006109483074396849, + 0.00922814104706049, + 0.005208099260926247, + 0.005888761021196842, + 0.0035051044542342424, + 0.004464512690901756, + 0.003876728704199195, + 0.004034894984215498, + 0.0025545356329530478, + 0.003533733543008566, + 0.0043577272444963455, + 0.003378961468115449, + 0.002701716497540474, + 0.0025353466626256704, + 0.0021223167423158884, + 0.0024720411747694016, + 0.02000913955271244, + 0.01810537464916706, + 0.007179415784776211, + 0.005521030630916357, + 0.0043508256785571575, + 0.004010965581983328, + 0.005160477012395859, + 0.004014648497104645, + 0.005427052732557058, + 0.003716276492923498, + 0.0031239434611052275, + 0.004431677982211113, + 0.002515220083296299, + 0.0032119820825755596, + 0.00418480858206749, + 0.004476788453757763, + 0.012144847773015499, + 0.0038916536141186953, + 0.012830231338739395, + 0.0035308359656482935, + 0.002541128545999527, + 0.0020338722970336676, + 0.001992889679968357, + 0.002159931929782033, + 0.0054372260347008705, + 0.003852775087580085, + 0.003088955534622073, + 0.00282503804191947, + 0.0024373859632760286, + 0.002139228628948331, + 0.0019470840925350785, + 0.003354796441271901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.004108322784304619 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361985132098198 + ], + "opposite_logprob_tokens": [ + -3.6513619422912598 + ], + "logprob_reward": -0.026361985132098198, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361985132098198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.063359335064888, + 0.08285371214151382, + 0.3744897246360779, + 0.11839233338832855, + 0.042728323489427567, + 0.03781028836965561, + 0.17187167704105377, + 0.018824666738510132, + 0.022953089326620102, + 0.06671684235334396 + ], + "credit_map": [ + 0.009862071834504604, + 0.007075710222125053, + 0.007683649659156799, + 0.008894111961126328, + 0.005730244796723127, + 0.005705816205590963, + 0.008250556886196136, + 0.010157179087400436, + 0.012188544496893883, + 0.010706745088100433, + 0.010904423892498016, + 0.013886167667806149, + 0.00822275597602129, + 0.009046902880072594, + 0.008069882169365883, + 0.009828293696045876, + 0.0314268060028553, + 0.12384230643510818, + 0.09430457651615143, + 0.04180604964494705, + 0.03080565296113491, + 0.009090281091630459, + 0.01595008186995983, + 0.027263976633548737, + 0.011137600988149643, + 0.021895237267017365, + 0.01535432506352663, + 0.015163757838308811, + 0.010226011276245117, + 0.00773298554122448, + 0.025478912517428398, + 0.011403506621718407, + 0.0060405791737139225, + 0.009048514999449253, + 0.006703211925923824, + 0.006307484116405249, + 0.0027390082832425833, + 0.004054941236972809, + 0.003663264214992523, + 0.004171322099864483, + 0.0040360684506595135, + 0.006582253612577915, + 0.007418227382004261, + 0.004710872191935778, + 0.004739142954349518, + 0.0031414730474352837, + 0.003717610612511635, + 0.003464639186859131, + 0.04502321407198906, + 0.04385446384549141, + 0.02609236352145672, + 0.015172802843153477, + 0.014148416928946972, + 0.01119137927889824, + 0.008344889618456364, + 0.008044137619435787, + 0.003361455164849758, + 0.0029854404274374247, + 0.002270859433338046, + 0.0017260394524782896, + 0.001906028832308948, + 0.0020305279176682234, + 0.0020055449567735195, + 0.002538772765547037, + 0.007382028736174107, + 0.001960199326276779, + 0.007524002809077501, + 0.0017234768019989133, + 0.0009711678139865398, + 0.0011497917585074902, + 0.0011717220768332481, + 0.0010706993052735925, + 0.019163863733410835, + 0.010909774340689182, + 0.007008134853094816, + 0.005248857196420431, + 0.008394007571041584, + 0.006265026517212391, + 0.0028712735511362553, + 0.006855907384306192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.026361985132098198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994293570518494 + ], + "opposite_logprob_tokens": [ + -1.8049429655075073 + ], + "logprob_reward": -0.17994293570518494, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994293570518494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11778988689184189, + 0.08211077004671097, + 0.2030913382768631, + 0.1951700896024704, + 0.10279712826013565, + 0.02015811577439308, + 0.030377482995390892, + 0.05065837875008583, + 0.11086991429328918, + 0.08697692304849625 + ], + "credit_map": [ + 0.01560020912438631, + 0.01572534441947937, + 0.016854427754878998, + 0.02024099789559841, + 0.010846365243196487, + 0.012422408908605576, + 0.009942688047885895, + 0.016157440841197968, + 0.012802706100046635, + 0.01140542235225439, + 0.00959719531238079, + 0.013485817238688469, + 0.008663441985845566, + 0.006969162728637457, + 0.008683093823492527, + 0.010503929108381271, + 0.02006295882165432, + 0.06249944493174553, + 0.05395280197262764, + 0.014890125952661037, + 0.016774466261267662, + 0.009911737404763699, + 0.008777112700045109, + 0.01622268185019493, + 0.020793236792087555, + 0.035963334143161774, + 0.024391593411564827, + 0.023115698248147964, + 0.01819019764661789, + 0.01562933437526226, + 0.04011378437280655, + 0.01697290688753128, + 0.015033148229122162, + 0.013431183993816376, + 0.020193664357066154, + 0.017565742135047913, + 0.007884659804403782, + 0.008175925351679325, + 0.012442078441381454, + 0.008070722222328186, + 0.0025873465929180384, + 0.002315615536645055, + 0.003312461543828249, + 0.003766761627048254, + 0.0016471275594085455, + 0.0024444616865366697, + 0.0022293124347925186, + 0.001855028560385108, + 0.004429669585078955, + 0.004282796755433083, + 0.005651704501360655, + 0.004204219672828913, + 0.0027093361131846905, + 0.0029364859219640493, + 0.0030293373856693506, + 0.0031339311972260475, + 0.008409866131842136, + 0.008807722479104996, + 0.007066773250699043, + 0.007396130822598934, + 0.004005665425211191, + 0.004906030837446451, + 0.005069975275546312, + 0.004996215924620628, + 0.03825126960873604, + 0.009766011498868465, + 0.038294292986392975, + 0.007264218758791685, + 0.0036527167540043592, + 0.0047559188678860664, + 0.004036081489175558, + 0.004849390126764774, + 0.026459762826561928, + 0.013708028942346573, + 0.009044715203344822, + 0.006825840566307306, + 0.009876042604446411, + 0.007944938726723194, + 0.0037027504295110703, + 0.009414846077561378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.17994293570518494 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle macro shot captures a sleek, modern hair dryer lying on its side on a pristine white marble countertop with subtle gray veining. The hair dryer's body is crafted from a sophisticated matte black, soft-touch material, accented by a polished rose gold ring surrounding the nozzle. Tied around this nozzle are several thin, luminous ribbons in vibrant magenta, electric blue, and sunny yellow, initially soft and still. Without warning, the hair dryer silently powers on, and the ribbons are instantly jolted to life by invisible jets of air. They begin to flutter violently, whipping back and forth in a chaotic yet graceful dance, their light fabric twisting and snapping, creating a mesmerizing, high-speed blur of colors that contrasts starkly with the appliance's inert, dark form. Soft, directional studio lighting captures the delicate sheen of the flowing silk, shimmering with rose gold accents, casting rapidly shifting, subtle shadows on the marble surface below, creating an atmosphere of vibrant energy and visual allure, unleashing power and aesthetic elegance.", + "seed": 10086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_2/index_6853_seed_10086_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": -0.2856898743659258, + "vlm_logprob": -0.2856898743659258, + "vlm_yes_no_margin": 1.85, + "vlm_token_credit": -0.2856898743659258, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5816084742546082 + ], + "opposite_logprob_tokens": [ + -0.8316084742546082 + ], + "logprob_reward": -0.5816084742546082, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5816084742546082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34949401021003723, + 0.1466737985610962, + 0.16383104026317596, + 0.02351081743836403, + 0.04084864258766174, + 0.044367674738168716, + 0.10962214320898056, + 0.06993649154901505, + 0.03239821270108223, + 0.01931721903383732 + ], + "credit_map": [ + 0.04705984890460968, + 0.044817958027124405, + 0.05159240588545799, + 0.0446193590760231, + 0.031112471595406532, + 0.03947008028626442, + 0.05688199773430824, + 0.033939871937036514, + 0.01780572161078453, + 0.01448886375874281, + 0.021736793220043182, + 0.024804741144180298, + 0.009050528518855572, + 0.00796504132449627, + 0.03304779902100563, + 0.017774317413568497, + 0.03968668356537819, + 0.021264811977744102, + 0.027986489236354828, + 0.019076157361268997, + 0.009201609529554844, + 0.007047226652503014, + 0.02404071018099785, + 0.015527351759374142, + 0.00400384608656168, + 0.0020937910303473473, + 0.0037930388934910297, + 0.0026833948213607073, + 0.0023489557206630707, + 0.003526933491230011, + 0.0024360944516956806, + 0.002624763874337077, + 0.006259430665522814, + 0.003055212553590536, + 0.00665450282394886, + 0.004521663766354322, + 0.003997880965471268, + 0.003271929919719696, + 0.008352097123861313, + 0.004735923372209072, + 0.007318192161619663, + 0.0029573324136435986, + 0.005141590256243944, + 0.004881137516349554, + 0.006336226128041744, + 0.003765771398320794, + 0.008978552184998989, + 0.004988871980458498, + 0.008788865990936756, + 0.005994491279125214, + 0.012380837462842464, + 0.014759019017219543, + 0.009421493858098984, + 0.00848778709769249, + 0.03174307942390442, + 0.01804656907916069, + 0.01159896980971098, + 0.006404762156307697, + 0.00951782800257206, + 0.00860190112143755, + 0.006479422561824322, + 0.005321052856743336, + 0.013882172293961048, + 0.008130384609103203, + 0.003882819088175893, + 0.002120533026754856, + 0.004426630213856697, + 0.0035980893298983574, + 0.002453151624649763, + 0.002425889251753688, + 0.008336343802511692, + 0.005154757294803858, + 0.001976137747988105, + 0.0011352458968758583, + 0.0025025398936122656, + 0.0025782270822674036, + 0.001135354395955801, + 0.001387308700941503, + 0.005394898355007172, + 0.0032075049821287394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.5816084742546082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05504162609577179 + ], + "opposite_logprob_tokens": [ + -2.930041551589966 + ], + "logprob_reward": -0.05504162609577179, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05504162609577179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37179288268089294, + 0.12308664619922638, + 0.1687646359205246, + 0.03619290515780449, + 0.035649433732032776, + 0.046356331557035446, + 0.09751339256763458, + 0.06803204864263535, + 0.02843976393342018, + 0.024171965196728706 + ], + "credit_map": [ + 0.04967723786830902, + 0.044190533459186554, + 0.056138914078474045, + 0.04259495064616203, + 0.034272678196430206, + 0.044729992747306824, + 0.06266824156045914, + 0.03752032667398453, + 0.014771746471524239, + 0.012052476406097412, + 0.018009645864367485, + 0.02032039314508438, + 0.007696499582380056, + 0.006877884268760681, + 0.02791529893875122, + 0.015442702919244766, + 0.03674053028225899, + 0.021295960992574692, + 0.027366871014237404, + 0.019832031801342964, + 0.009376594796776772, + 0.007646955084055662, + 0.030059562996029854, + 0.016446132212877274, + 0.005294597242027521, + 0.0029905594419687986, + 0.004524628631770611, + 0.004082323983311653, + 0.0035896077752113342, + 0.004891103133559227, + 0.005927283316850662, + 0.004892804194241762, + 0.005558321252465248, + 0.002590841380879283, + 0.005661558825522661, + 0.004154385067522526, + 0.002952985232695937, + 0.003044570330530405, + 0.0074625397101044655, + 0.004224229138344526, + 0.008396039716899395, + 0.003588927211239934, + 0.0056497869081795216, + 0.005312536843121052, + 0.0065412623807787895, + 0.003920586779713631, + 0.007821031846106052, + 0.005126158706843853, + 0.008303748443722725, + 0.005755338817834854, + 0.011056339368224144, + 0.013003055937588215, + 0.00904108863323927, + 0.00776698300614953, + 0.02706744335591793, + 0.015519395470619202, + 0.012950779870152473, + 0.005105637479573488, + 0.010200301185250282, + 0.008697749115526676, + 0.005867348052561283, + 0.004846277646720409, + 0.011943342164158821, + 0.008420612663030624, + 0.0037085285875946283, + 0.002080951351672411, + 0.003963892813771963, + 0.0030383483972400427, + 0.0022269971668720245, + 0.001925273914821446, + 0.006975671276450157, + 0.004520100075751543, + 0.0021916611585766077, + 0.0012483386090025306, + 0.002989523345604539, + 0.0031344236340373755, + 0.0013300423743203282, + 0.001819233875721693, + 0.00728128757327795, + 0.004177452996373177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.05504162609577179 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087958589196205 + ], + "opposite_logprob_tokens": [ + -4.141088008880615 + ], + "logprob_reward": -0.016087958589196205, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087958589196205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.406447172164917, + 0.10028468817472458, + 0.14593926072120667, + 0.14059849083423615, + 0.027953989803791046, + 0.033570270985364914, + 0.031212732195854187, + 0.05231650173664093, + 0.02799954079091549, + 0.03367733582854271 + ], + "credit_map": [ + 0.05234324187040329, + 0.06597477942705154, + 0.060782067477703094, + 0.06815650314092636, + 0.039042338728904724, + 0.03562536463141441, + 0.04720069468021393, + 0.03732218220829964, + 0.013943268917500973, + 0.010682298801839352, + 0.014305179938673973, + 0.015627728775143623, + 0.007960900664329529, + 0.005556588061153889, + 0.022139307111501694, + 0.01006941869854927, + 0.02296566031873226, + 0.016763249412178993, + 0.022506017237901688, + 0.017464283853769302, + 0.007943038828670979, + 0.007514684461057186, + 0.03707459941506386, + 0.013707726262509823, + 0.01876985654234886, + 0.011411759071052074, + 0.016565006226301193, + 0.01327358279377222, + 0.017467286437749863, + 0.017070168629288673, + 0.024053119122982025, + 0.021987710148096085, + 0.004248993471264839, + 0.0021323179826140404, + 0.003739384701475501, + 0.0036123234312981367, + 0.002627616049721837, + 0.001902267918922007, + 0.006755958776921034, + 0.002935128752142191, + 0.0065213763155043125, + 0.0028371524531394243, + 0.004553103353828192, + 0.004561729729175568, + 0.00438656285405159, + 0.002994989277794957, + 0.004298876971006393, + 0.0034164797980338335, + 0.005243366118520498, + 0.002156083006411791, + 0.004281522706151009, + 0.003440714906901121, + 0.0034994767047464848, + 0.003190192859619856, + 0.005876778159290552, + 0.0035245963372290134, + 0.011278954334557056, + 0.010405641980469227, + 0.00798859540373087, + 0.0060517690144479275, + 0.003666471689939499, + 0.0029711094684898853, + 0.006404628045856953, + 0.0035493308678269386, + 0.0038935591001063585, + 0.002725962782278657, + 0.0035534086637198925, + 0.00312215369194746, + 0.0022067190147936344, + 0.002653786214068532, + 0.00648022536188364, + 0.00336372759193182, + 0.005681953858584166, + 0.0034754187799990177, + 0.004451426211744547, + 0.003948289901018143, + 0.0038706958293914795, + 0.0028021633625030518, + 0.005240145605057478, + 0.004207241348922253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.016087958589196205 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284159243106842 + ], + "opposite_logprob_tokens": [ + -2.0178415775299072 + ], + "logprob_reward": -0.14284159243106842, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284159243106842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31354251503944397, + 0.10426367074251175, + 0.06704039126634598, + 0.13107262551784515, + 0.04792036861181259, + 0.09829119592905045, + 0.09183256328105927, + 0.10835370421409607, + 0.0281399916857481, + 0.009542958810925484 + ], + "credit_map": [ + 0.041897937655448914, + 0.04978059232234955, + 0.05034370720386505, + 0.06223487854003906, + 0.024310147389769554, + 0.020617282018065453, + 0.03444506973028183, + 0.029912913218140602, + 0.02031349390745163, + 0.011097919195890427, + 0.016025139018893242, + 0.015084479935467243, + 0.006905821617692709, + 0.007788622751832008, + 0.01558839064091444, + 0.011459807865321636, + 0.01001725159585476, + 0.006309053394943476, + 0.008670376613736153, + 0.007756069768220186, + 0.004612669814378023, + 0.00413008127361536, + 0.01842680387198925, + 0.007118090055882931, + 0.01865391619503498, + 0.010224251076579094, + 0.0193327683955431, + 0.01874634064733982, + 0.013716800138354301, + 0.009093645960092545, + 0.025267189368605614, + 0.01603773422539234, + 0.008155552670359612, + 0.003920113667845726, + 0.008564075455069542, + 0.005986635573208332, + 0.003984810784459114, + 0.0035456314217299223, + 0.008549388498067856, + 0.0052141654305160046, + 0.012553656473755836, + 0.007310128305107355, + 0.013665078207850456, + 0.010762977413833141, + 0.010088334791362286, + 0.008127265609800816, + 0.022292669862508774, + 0.013491088524460793, + 0.009795892052352428, + 0.005994880571961403, + 0.010546629317104816, + 0.012590516358613968, + 0.006007108837366104, + 0.008443444967269897, + 0.024921337142586708, + 0.01353276427835226, + 0.028349988162517548, + 0.021827705204486847, + 0.01745934598147869, + 0.011402984149754047, + 0.0054898569360375404, + 0.004227874800562859, + 0.009919069707393646, + 0.009676886722445488, + 0.005324606318026781, + 0.0027155964635312557, + 0.004257942549884319, + 0.002997123636305332, + 0.0026322246994823217, + 0.0023492206819355488, + 0.004807198885828257, + 0.0030560807790607214, + 0.0013982129748910666, + 0.0007886811508797109, + 0.0017622592858970165, + 0.0011391285806894302, + 0.0009122794144786894, + 0.0008857105276547372, + 0.0014568861806765199, + 0.0011998015688732266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.14284159243106842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328697204589844 + ], + "opposite_logprob_tokens": [ + -0.7578697204589844 + ], + "logprob_reward": -0.6328697204589844, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328697204589844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31772491335868835, + 0.07494354993104935, + 0.03517584502696991, + 0.15017980337142944, + 0.07615114003419876, + 0.08768127113580704, + 0.06312737613916397, + 0.10543099045753479, + 0.052308402955532074, + 0.03727668896317482 + ], + "credit_map": [ + 0.04746253788471222, + 0.03983808681368828, + 0.054265931248664856, + 0.049419354647397995, + 0.023611290380358696, + 0.016847215592861176, + 0.05170385167002678, + 0.03457663208246231, + 0.00964896846562624, + 0.008402771316468716, + 0.011172051541507244, + 0.01230101939290762, + 0.004969978239387274, + 0.004125256557017565, + 0.015686744824051857, + 0.00863675493746996, + 0.005822724662721157, + 0.005071571562439203, + 0.005261986516416073, + 0.004745873156934977, + 0.0026996068190783262, + 0.002604034263640642, + 0.00510292686522007, + 0.003867123043164611, + 0.02181246690452099, + 0.012679382227361202, + 0.018150946125388145, + 0.01618862897157669, + 0.01904684491455555, + 0.015097694471478462, + 0.026514828205108643, + 0.020689019933342934, + 0.011419886723160744, + 0.004824166186153889, + 0.01235936488956213, + 0.008223770186305046, + 0.007239949889481068, + 0.0052748932503163815, + 0.01729058288037777, + 0.009518529288470745, + 0.014482680708169937, + 0.008391477167606354, + 0.011701186187565327, + 0.01056013535708189, + 0.015106282196938992, + 0.00836589839309454, + 0.009792988188564777, + 0.009280615486204624, + 0.008936205878853798, + 0.004070463590323925, + 0.007612516637891531, + 0.0061192805878818035, + 0.008328489027917385, + 0.005291911773383617, + 0.014819071628153324, + 0.007949437946081161, + 0.012503993697464466, + 0.006905386224389076, + 0.012594718486070633, + 0.012557290494441986, + 0.009616557508707047, + 0.009899618104100227, + 0.02716408111155033, + 0.0141893420368433, + 0.007360202260315418, + 0.004638382699340582, + 0.007004222832620144, + 0.006110949441790581, + 0.005426059942692518, + 0.004667192697525024, + 0.011065076105296612, + 0.006036313716322184, + 0.003956970758736134, + 0.0025501553900539875, + 0.004779531620442867, + 0.005211215931922197, + 0.0024280943907797337, + 0.002946276916190982, + 0.00986695010215044, + 0.0055374945513904095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.6328697204589844 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle macro shot captures a sleek, modern hair dryer lying on its side on a pristine white marble countertop with subtle gray veining. The hair dryer's body is crafted from a sophisticated matte black, soft-touch material, accented by a polished rose gold ring surrounding the nozzle. Tied around this nozzle are several thin, luminous ribbons in vibrant magenta, electric blue, and sunny yellow, initially soft and still. Without warning, the hair dryer silently powers on, and the ribbons are instantly jolted to life by invisible jets of air. They begin to flutter violently, whipping back and forth in a chaotic yet graceful dance, their light fabric twisting and snapping, creating a mesmerizing, high-speed blur of colors that contrasts starkly with the appliance's inert, dark form. Soft, directional studio lighting captures the delicate sheen of the flowing silk, shimmering with rose gold accents, casting rapidly shifting, subtle shadows on the marble surface below, creating an atmosphere of vibrant energy and visual allure, unleashing power and aesthetic elegance.", + "seed": 10087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_2/index_6853_seed_10087_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": -0.10094123361632228, + "vlm_logprob": -0.10094123361632228, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.10094123361632228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021850910037755966 + ], + "opposite_logprob_tokens": [ + -4.021851062774658 + ], + "logprob_reward": -0.021850910037755966, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021850910037755966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14402340352535248, + 0.18662768602371216, + 0.07467865943908691, + 0.05114287883043289, + 0.19792304933071136, + 0.04310445487499237, + 0.07806906849145889, + 0.042013853788375854, + 0.1690267026424408, + 0.013390213251113892 + ], + "credit_map": [ + 0.022260431200265884, + 0.019249750301241875, + 0.01738942414522171, + 0.020075198262929916, + 0.015719670802354813, + 0.01637946255505085, + 0.017764901742339134, + 0.015184566378593445, + 0.031415920704603195, + 0.029263295233249664, + 0.023208193480968475, + 0.018066419288516045, + 0.03119419328868389, + 0.019451621919870377, + 0.016441520303487778, + 0.01758652552962303, + 0.010313989594578743, + 0.009900791570544243, + 0.009646727703511715, + 0.005933209788054228, + 0.009725544601678848, + 0.007310323417186737, + 0.01617375575006008, + 0.00567432027310133, + 0.00875370018184185, + 0.007640012539923191, + 0.006737766321748495, + 0.007827646099030972, + 0.0043663387186825275, + 0.006730697117745876, + 0.004743437748402357, + 0.004343281965702772, + 0.039097029715776443, + 0.01478533074259758, + 0.028312362730503082, + 0.02437017671763897, + 0.018611544743180275, + 0.02404656447470188, + 0.013646343722939491, + 0.035053692758083344, + 0.01102773193269968, + 0.003422902664169669, + 0.007797561585903168, + 0.0065404633060097694, + 0.002961265156045556, + 0.003228571731597185, + 0.0036641908809542656, + 0.004461771342903376, + 0.015718895941972733, + 0.008246298879384995, + 0.011646823026239872, + 0.013815462589263916, + 0.007344840560108423, + 0.007777579128742218, + 0.00686001917347312, + 0.0066591426730155945, + 0.0077406964264810085, + 0.004164195153862238, + 0.005423597060143948, + 0.005909029860049486, + 0.003172823693603277, + 0.005410465877503157, + 0.005576622672379017, + 0.004616419784724712, + 0.028918348252773285, + 0.02201901189982891, + 0.017936265096068382, + 0.014217499643564224, + 0.026535646989941597, + 0.01460176520049572, + 0.01855764538049698, + 0.026240510866045952, + 0.0023923644330352545, + 0.0016802788013592362, + 0.001695143524557352, + 0.002052526455372572, + 0.001630681101232767, + 0.0011685896897688508, + 0.0012964787892997265, + 0.0014741498744115233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.021850910037755966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026379982009530067 + ], + "opposite_logprob_tokens": [ + -3.6513800621032715 + ], + "logprob_reward": -0.026379982009530067, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026379982009530067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10891492664813995, + 0.07459407299757004, + 0.028053956106305122, + 0.029749710112810135, + 0.13511094450950623, + 0.03624822199344635, + 0.06103195250034332, + 0.01323426142334938, + 0.49258953332901, + 0.020472349599003792 + ], + "credit_map": [ + 0.015824802219867706, + 0.012700995430350304, + 0.013204536400735378, + 0.014554056338965893, + 0.010334120132029057, + 0.015478802844882011, + 0.016136687248945236, + 0.010680939070880413, + 0.010992064140737057, + 0.011677639558911324, + 0.009342711418867111, + 0.008608785457909107, + 0.0135254617780447, + 0.0060134464874863625, + 0.00714210094884038, + 0.007291877176612616, + 0.004076005425304174, + 0.002788300858810544, + 0.003820856800302863, + 0.0037718811072409153, + 0.0037852413952350616, + 0.0027472658548504114, + 0.004804690834134817, + 0.002259717555716634, + 0.004475959111005068, + 0.0053768944926559925, + 0.0037617727648466825, + 0.0030835289508104324, + 0.00270052719861269, + 0.004702040459960699, + 0.002982411300763488, + 0.002666580257937312, + 0.02589290589094162, + 0.010776673443615437, + 0.019453207030892372, + 0.013284917920827866, + 0.012932712212204933, + 0.01659485325217247, + 0.010491261258721352, + 0.02568444237112999, + 0.005293393973261118, + 0.003687799209728837, + 0.004056029953062534, + 0.005494975950568914, + 0.004203314892947674, + 0.0038340373430401087, + 0.005547587759792805, + 0.004131087101995945, + 0.011238181963562965, + 0.007388313300907612, + 0.008288739249110222, + 0.006857675965875387, + 0.008674153126776218, + 0.00696905329823494, + 0.005214279517531395, + 0.006401564925909042, + 0.0022851123940199614, + 0.0014065656578168273, + 0.0016122037777677178, + 0.0013454293366521597, + 0.0014151796931400895, + 0.0015989464009180665, + 0.0020036250352859497, + 0.0015672004083171487, + 0.08499933779239655, + 0.06546337902545929, + 0.05375128239393234, + 0.041897859424352646, + 0.07324071228504181, + 0.04261792451143265, + 0.05254626274108887, + 0.07807278633117676, + 0.002880363492295146, + 0.002189015503972769, + 0.0022749530617147684, + 0.002759245689958334, + 0.002663188148289919, + 0.0029494597110897303, + 0.0025305699091404676, + 0.002225557342171669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.026379982009530067 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262214224785566 + ], + "opposite_logprob_tokens": [ + -5.25526237487793 + ], + "logprob_reward": -0.005262214224785566, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262214224785566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016652079299092293, + 0.024073876440525055, + 0.017538199201226234, + 0.008551194332540035, + 0.020979532971978188, + 0.008566566742956638, + 0.01809495873749256, + 0.02066115103662014, + 0.8607826828956604, + 0.004099724348634481 + ], + "credit_map": [ + 0.003130085999146104, + 0.0017031123861670494, + 0.0024392963387072086, + 0.0024057968985289335, + 0.0016361026791855693, + 0.0017947449814528227, + 0.001987446565181017, + 0.0015554935671389103, + 0.0035998087842017412, + 0.0032755485735833645, + 0.002947577042505145, + 0.0038382727652788162, + 0.002888472517952323, + 0.0019246197771281004, + 0.0019633707124739885, + 0.003636204870417714, + 0.0023881131783127785, + 0.0022067222744226456, + 0.0020228931680321693, + 0.0020272431429475546, + 0.002439659321680665, + 0.001493834308348596, + 0.0031126108951866627, + 0.001847122679464519, + 0.001544941682368517, + 0.000997512019239366, + 0.0011778839398175478, + 0.0013572914758697152, + 0.0006916568381711841, + 0.0011268300004303455, + 0.0006780277471989393, + 0.0009770503966137767, + 0.0037295909132808447, + 0.002012170385569334, + 0.0029123062267899513, + 0.00294448877684772, + 0.0016954663442447782, + 0.002187039703130722, + 0.0017943429993465543, + 0.0037041299510747194, + 0.001246199244633317, + 0.000698741408996284, + 0.0010223094141110778, + 0.0016525359824299812, + 0.0008878230582922697, + 0.0008159196586348116, + 0.001311364583671093, + 0.0009316738578490913, + 0.0034811594523489475, + 0.0023823322262614965, + 0.0025336509570479393, + 0.0016739886486902833, + 0.0022670826874673367, + 0.001980759669095278, + 0.0018390475306659937, + 0.0019369374494999647, + 0.003470509545877576, + 0.0018918663263320923, + 0.0024726081173866987, + 0.0028435164131224155, + 0.0019346019253134727, + 0.0024590392131358385, + 0.0026782755739986897, + 0.002910735085606575, + 0.14747127890586853, + 0.10809431225061417, + 0.09155070781707764, + 0.07905654609203339, + 0.1305440217256546, + 0.07431157678365707, + 0.0939243957400322, + 0.135829895734787, + 0.0006645839312113822, + 0.00044059468200430274, + 0.00048497572424821556, + 0.0005414645420387387, + 0.0004690331406891346, + 0.0005728047108277678, + 0.00041106267599388957, + 0.0005152048543095589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.005262214224785566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256159484386444 + ], + "opposite_logprob_tokens": [ + -1.6006159782409668 + ], + "logprob_reward": -0.2256159484386444, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256159484386444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16487795114517212, + 0.0675712302327156, + 0.16714343428611755, + 0.0648835226893425, + 0.137678325176239, + 0.0586688332259655, + 0.05386904999613762, + 0.06302733719348907, + 0.19518901407718658, + 0.02709135040640831 + ], + "credit_map": [ + 0.03622942790389061, + 0.018825000151991844, + 0.024722659960389137, + 0.02158055640757084, + 0.015743451192975044, + 0.02004753239452839, + 0.015708312392234802, + 0.01202100794762373, + 0.011021728627383709, + 0.009935795329511166, + 0.008684931322932243, + 0.007413039915263653, + 0.007264596410095692, + 0.007538137026131153, + 0.005519148427993059, + 0.010193862952291965, + 0.021179260686039925, + 0.02363251894712448, + 0.020769814029335976, + 0.01569783128798008, + 0.019769635051488876, + 0.01799684390425682, + 0.03380919247865677, + 0.014288319274783134, + 0.01588110625743866, + 0.0055936174467206, + 0.01093883067369461, + 0.009032048285007477, + 0.0050994581542909145, + 0.005076736211776733, + 0.005864907056093216, + 0.007396820466965437, + 0.027268823236227036, + 0.013602307066321373, + 0.019485022872686386, + 0.026183217763900757, + 0.008199709467589855, + 0.012554961256682873, + 0.01072345394641161, + 0.01966082490980625, + 0.010392374359071255, + 0.006741589866578579, + 0.008284049108624458, + 0.008567478507757187, + 0.006168517749756575, + 0.005070067476481199, + 0.007585981395095587, + 0.005858772899955511, + 0.01215747743844986, + 0.006843038834631443, + 0.007247184403240681, + 0.005078017245978117, + 0.004525001160800457, + 0.005403747782111168, + 0.007387591525912285, + 0.005226995330303907, + 0.011430403217673302, + 0.007726073730736971, + 0.008137557655572891, + 0.01024213619530201, + 0.006550117861479521, + 0.006113875657320023, + 0.005408516153693199, + 0.007418654393404722, + 0.028454329818487167, + 0.01834658347070217, + 0.01863296516239643, + 0.024152427911758423, + 0.02579982578754425, + 0.022084534168243408, + 0.0281339380890131, + 0.029584405943751335, + 0.005912615917623043, + 0.003341029630973935, + 0.0033901494462043047, + 0.004264621529728174, + 0.0021611538249999285, + 0.0037897960282862186, + 0.002380334073677659, + 0.0018516505369916558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.2256159484386444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559711337089539 + ], + "opposite_logprob_tokens": [ + -1.6005971431732178 + ], + "logprob_reward": -0.22559711337089539, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559711337089539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10074087232351303, + 0.04871635138988495, + 0.14078035950660706, + 0.05846000835299492, + 0.082648366689682, + 0.05122144892811775, + 0.05485311895608902, + 0.07451900839805603, + 0.3484213352203369, + 0.0396391823887825 + ], + "credit_map": [ + 0.01733279600739479, + 0.008565806783735752, + 0.015835238620638847, + 0.02145206183195114, + 0.008438403718173504, + 0.012257348746061325, + 0.007322760298848152, + 0.009536448866128922, + 0.008883656933903694, + 0.005516093224287033, + 0.007755164057016373, + 0.009509533643722534, + 0.003281997749581933, + 0.0045257071033120155, + 0.00419639702886343, + 0.005047799553722143, + 0.029462771490216255, + 0.013098394498229027, + 0.022552238777279854, + 0.031882479786872864, + 0.008467067033052444, + 0.00890952255576849, + 0.011770335957407951, + 0.014637544751167297, + 0.010566791519522667, + 0.0063891359604895115, + 0.007535998243838549, + 0.010482694953680038, + 0.00415064999833703, + 0.005460991524159908, + 0.005566481966525316, + 0.008307254873216152, + 0.01914796605706215, + 0.007061951328068972, + 0.01247606985270977, + 0.01743713766336441, + 0.004838965833187103, + 0.006607276853173971, + 0.005884095560759306, + 0.009194895625114441, + 0.00877531711012125, + 0.006299625616520643, + 0.0070309932343661785, + 0.0071939569897949696, + 0.005720452405512333, + 0.004597908351570368, + 0.007106826175004244, + 0.004496363457292318, + 0.01002566423267126, + 0.006811931729316711, + 0.006443957332521677, + 0.005567742511630058, + 0.005298976320773363, + 0.005356170237064362, + 0.00902014784514904, + 0.006328522693365812, + 0.012701145373284817, + 0.007304669823497534, + 0.00888331513851881, + 0.00897296518087387, + 0.006597640924155712, + 0.010579768568277359, + 0.010425821878015995, + 0.009053673595190048, + 0.06097957119345665, + 0.04540746286511421, + 0.04038118198513985, + 0.029863959178328514, + 0.05398040637373924, + 0.028070520609617233, + 0.03666246309876442, + 0.05307573452591896, + 0.011587921530008316, + 0.0037135756574571133, + 0.007211959920823574, + 0.004933301825076342, + 0.0033760482911020517, + 0.002649490488693118, + 0.0032224906608462334, + 0.00294439191929996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.22559711337089539 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle macro shot captures a sleek, modern hair dryer lying on its side on a pristine white marble countertop with subtle gray veining. The hair dryer's body is crafted from a sophisticated matte black, soft-touch material, accented by a polished rose gold ring surrounding the nozzle. Tied around this nozzle are several thin, luminous ribbons in vibrant magenta, electric blue, and sunny yellow, initially soft and still. Without warning, the hair dryer silently powers on, and the ribbons are instantly jolted to life by invisible jets of air. They begin to flutter violently, whipping back and forth in a chaotic yet graceful dance, their light fabric twisting and snapping, creating a mesmerizing, high-speed blur of colors that contrasts starkly with the appliance's inert, dark form. Soft, directional studio lighting captures the delicate sheen of the flowing silk, shimmering with rose gold accents, casting rapidly shifting, subtle shadows on the marble surface below, creating an atmosphere of vibrant energy and visual allure, unleashing power and aesthetic elegance.", + "seed": 10088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_2/index_6853_seed_10088_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": -0.34859764352440836, + "vlm_logprob": -0.34859764352440836, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.34859764352440836, + "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 a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0572028160095215 + ], + "opposite_logprob_tokens": [ + -0.4322027862071991 + ], + "logprob_reward": -1.0572028160095215, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0572028160095215, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22275632619857788, + 0.07157434523105621, + 0.09251635521650314, + 0.14885134994983673, + 0.14468738436698914, + 0.09174983203411102, + 0.040036339312791824, + 0.06802315264940262, + 0.07457289844751358, + 0.045232079923152924 + ], + "credit_map": [ + 0.03191229701042175, + 0.02382335066795349, + 0.03219003975391388, + 0.02074502408504486, + 0.025744423270225525, + 0.03475432097911835, + 0.0207294300198555, + 0.03285742178559303, + 0.011247079819440842, + 0.00953246932476759, + 0.006049302406609058, + 0.005128175485879183, + 0.006322862580418587, + 0.014214844442903996, + 0.008657693862915039, + 0.010421914048492908, + 0.010069549083709717, + 0.009957974776625633, + 0.011125506833195686, + 0.00877248402684927, + 0.009547044523060322, + 0.014775226823985577, + 0.013837768696248531, + 0.014430781826376915, + 0.017556535080075264, + 0.017442595213651657, + 0.009865153580904007, + 0.009561609476804733, + 0.027993252500891685, + 0.02897379733622074, + 0.01803511381149292, + 0.01942327246069908, + 0.016241002827882767, + 0.015214280225336552, + 0.01635752059519291, + 0.014995868317782879, + 0.01595168374478817, + 0.031018510460853577, + 0.01649506948888302, + 0.018413439393043518, + 0.009988024830818176, + 0.010083337314426899, + 0.013059994205832481, + 0.017410030588507652, + 0.009679012931883335, + 0.010200838558375835, + 0.010885056108236313, + 0.010443530976772308, + 0.004855882376432419, + 0.006887755356729031, + 0.0025093453004956245, + 0.0039008583407849073, + 0.003172964556142688, + 0.006082228384912014, + 0.008684174157679081, + 0.003943126183003187, + 0.0075576333329081535, + 0.006090313661843538, + 0.00603299168869853, + 0.007509532384574413, + 0.0068656811490654945, + 0.012049907818436623, + 0.014485359191894531, + 0.0074317255057394505, + 0.007292204536497593, + 0.007010068744421005, + 0.011497235856950283, + 0.008847433142364025, + 0.006694310810416937, + 0.011317592114210129, + 0.015135985799133778, + 0.006778060458600521, + 0.004481550771743059, + 0.003126315074041486, + 0.003162264358252287, + 0.0044600339606404305, + 0.0044249799102544785, + 0.009548320434987545, + 0.009074756875634193, + 0.006953854113817215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -1.0572028160095215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055074457079172134 + ], + "opposite_logprob_tokens": [ + -2.930074453353882 + ], + "logprob_reward": -0.055074457079172134, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055074457079172134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13524769246578217, + 0.06673412024974823, + 0.07941748946905136, + 0.18804150819778442, + 0.17104050517082214, + 0.07723961770534515, + 0.07653212547302246, + 0.06235434487462044, + 0.08043450117111206, + 0.06295816600322723 + ], + "credit_map": [ + 0.018750622868537903, + 0.014455805532634258, + 0.019937915727496147, + 0.01604798249900341, + 0.013206755742430687, + 0.02041982114315033, + 0.013474377803504467, + 0.018954405561089516, + 0.009764962829649448, + 0.007918265648186207, + 0.005295589566230774, + 0.005084366071969271, + 0.00655377795919776, + 0.011617082171142101, + 0.010464755818247795, + 0.010035314597189426, + 0.008430383168160915, + 0.0076101976446807384, + 0.0074693881906569, + 0.005835747811943293, + 0.0068259709514677525, + 0.014539996162056923, + 0.013550853356719017, + 0.015154951252043247, + 0.021652838215231895, + 0.021083461120724678, + 0.013751527294516563, + 0.012880867347121239, + 0.024851206690073013, + 0.023733634501695633, + 0.0485612228512764, + 0.02152673713862896, + 0.02100413851439953, + 0.016721300780773163, + 0.018931029364466667, + 0.016144854947924614, + 0.01958007551729679, + 0.0365758053958416, + 0.018279578536748886, + 0.023803718388080597, + 0.008463985286653042, + 0.006799660623073578, + 0.011055528186261654, + 0.014096022583544254, + 0.008132757619023323, + 0.01229457464069128, + 0.008101779967546463, + 0.008295306004583836, + 0.007938755676150322, + 0.014241797849535942, + 0.004714489448815584, + 0.008318387903273106, + 0.007672234904021025, + 0.014107558876276016, + 0.010185970924794674, + 0.009352938272058964, + 0.007042984943836927, + 0.006078910082578659, + 0.00517796678468585, + 0.0068430653773248196, + 0.00596666568890214, + 0.008987448178231716, + 0.01616147719323635, + 0.0060958256945014, + 0.009157069958746433, + 0.006464812904596329, + 0.011511115357279778, + 0.008955619297921658, + 0.006830065976828337, + 0.013728823512792587, + 0.015356853604316711, + 0.008430141024291515, + 0.006724145263433456, + 0.004462905693799257, + 0.004061192274093628, + 0.005498984362930059, + 0.006122165825217962, + 0.015007155947387218, + 0.012025532312691212, + 0.009056080132722855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.055074457079172134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820462942123413 + ], + "opposite_logprob_tokens": [ + -4.018204689025879 + ], + "logprob_reward": -0.01820462942123413, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820462942123413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24016931653022766, + 0.07956276834011078, + 0.12115732580423355, + 0.09902101010084152, + 0.13054217398166656, + 0.12213987112045288, + 0.039882875978946686, + 0.05399720370769501, + 0.08442535996437073, + 0.0291020218282938 + ], + "credit_map": [ + 0.034093696624040604, + 0.02531246282160282, + 0.03690630942583084, + 0.025091156363487244, + 0.02380947582423687, + 0.04203611984848976, + 0.020687285810709, + 0.03223280981183052, + 0.010271244682371616, + 0.011853352189064026, + 0.007711753249168396, + 0.007201624568551779, + 0.006623974535614252, + 0.011998073197901249, + 0.013230221346020699, + 0.010672533884644508, + 0.015627745538949966, + 0.015799786895513535, + 0.01486685499548912, + 0.014819443225860596, + 0.010905221104621887, + 0.02012246660888195, + 0.013912737369537354, + 0.01510307751595974, + 0.009544456377625465, + 0.01759614609181881, + 0.01284483727067709, + 0.009979655034840107, + 0.014495763927698135, + 0.013217843137681484, + 0.011651168577373028, + 0.009691134095191956, + 0.015816885977983475, + 0.011574381031095982, + 0.014596276916563511, + 0.012906149961054325, + 0.01580181159079075, + 0.029062412679195404, + 0.01337420754134655, + 0.017410049214959145, + 0.013767124153673649, + 0.014136148616671562, + 0.01647941954433918, + 0.02290915884077549, + 0.011206131428480148, + 0.016448283568024635, + 0.013598868623375893, + 0.013594732619822025, + 0.004581217188388109, + 0.005055768880993128, + 0.0037793524097651243, + 0.004283270798623562, + 0.004851154517382383, + 0.005300499033182859, + 0.008185110054910183, + 0.003846504492685199, + 0.005104797426611185, + 0.003459946019575, + 0.006531439255923033, + 0.004997123498469591, + 0.006440997123718262, + 0.012048497796058655, + 0.009747139178216457, + 0.005667262710630894, + 0.008944015949964523, + 0.0075402772054076195, + 0.013664801605045795, + 0.008037430234253407, + 0.00747958617284894, + 0.016256222501397133, + 0.013319650664925575, + 0.009183378890156746, + 0.002775635104626417, + 0.0028924508951604366, + 0.002452652668580413, + 0.0036281824577599764, + 0.003601220203563571, + 0.00477469339966774, + 0.005111107602715492, + 0.00386608112603426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.01820462942123413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909420669078827 + ], + "opposite_logprob_tokens": [ + -2.4640941619873047 + ], + "logprob_reward": -0.08909420669078827, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909420669078827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23769719898700714, + 0.053460877388715744, + 0.13038834929466248, + 0.06207120791077614, + 0.048194169998168945, + 0.25950783491134644, + 0.08358924090862274, + 0.058514002710580826, + 0.05763067305088043, + 0.008946413174271584 + ], + "credit_map": [ + 0.03417505696415901, + 0.028383634984493256, + 0.0412866435945034, + 0.02925570122897625, + 0.023434607312083244, + 0.02833358384668827, + 0.0221965741366148, + 0.030631422996520996, + 0.006269180215895176, + 0.007668611127883196, + 0.0062711345963180065, + 0.004504856187850237, + 0.005228073336184025, + 0.008682367391884327, + 0.009901311248540878, + 0.004935343284159899, + 0.016424700617790222, + 0.018633035942912102, + 0.014033677987754345, + 0.013555491343140602, + 0.019952718168497086, + 0.016755012795329094, + 0.013098434545099735, + 0.017935283482074738, + 0.006065176334232092, + 0.007860333658754826, + 0.005260221194475889, + 0.006281791254878044, + 0.010241745039820671, + 0.010707609355449677, + 0.008512189611792564, + 0.0071421461179852486, + 0.00527669582515955, + 0.005115630105137825, + 0.005854208022356033, + 0.005851384252309799, + 0.006742368917912245, + 0.006461398210376501, + 0.00656387023627758, + 0.00632861303165555, + 0.027849489822983742, + 0.029327664524316788, + 0.034179605543613434, + 0.04926910251379013, + 0.02494235523045063, + 0.03501105681061745, + 0.025498077273368835, + 0.033430472016334534, + 0.009152677841484547, + 0.011842436157166958, + 0.009099356830120087, + 0.012850344181060791, + 0.007176399230957031, + 0.012406942434608936, + 0.012265478260815144, + 0.008795605972409248, + 0.00680387020111084, + 0.004638523329049349, + 0.007553904317319393, + 0.00569354509934783, + 0.004462162498384714, + 0.011690056882798672, + 0.010705226100981236, + 0.006966718006879091, + 0.00572857353836298, + 0.004475452937185764, + 0.00602363096550107, + 0.006419506389647722, + 0.005215446464717388, + 0.009381690993905067, + 0.01406609732657671, + 0.006320273503661156, + 0.0009843440493568778, + 0.0011410763254389167, + 0.001024385099299252, + 0.001412153709679842, + 0.001152750221081078, + 0.0009455439867451787, + 0.001371317426674068, + 0.0009148421813733876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.08909420669078827 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle macro shot of a hair dryer on a white marble countertop with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hair dryer matte black with a polished rose gold ring around the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thin ribbons in magenta, electric blue, and sunny yellow tied around the hair dryer nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ribbons flutter violently and create a high-speed blur of colors due to invisible jets of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting cast rapidly shifting, subtle shadows on the marble surface below the hair dryer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234121084213257 + ], + "opposite_logprob_tokens": [ + -0.8984121084213257 + ], + "logprob_reward": -0.5234121084213257, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234121084213257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25886523723602295, + 0.07715501636266708, + 0.19456417858600616, + 0.0869322121143341, + 0.06169411540031433, + 0.08348146826028824, + 0.052351295948028564, + 0.08110311627388, + 0.08126498758792877, + 0.022588400170207024 + ], + "credit_map": [ + 0.03884362801909447, + 0.02846764586865902, + 0.04583367705345154, + 0.026024335995316505, + 0.02418680675327778, + 0.0364106111228466, + 0.02236166223883629, + 0.0367368645966053, + 0.011163996532559395, + 0.01253574714064598, + 0.00640509370714426, + 0.00551598658785224, + 0.006606344133615494, + 0.013783730566501617, + 0.009066709317266941, + 0.012077405117452145, + 0.025697654113173485, + 0.025024302303791046, + 0.019880931824445724, + 0.019716182723641396, + 0.031509943306446075, + 0.025878774002194405, + 0.021874966099858284, + 0.0249814223498106, + 0.0116691580042243, + 0.020307229831814766, + 0.008709799498319626, + 0.007168370299041271, + 0.00946212187409401, + 0.01014697179198265, + 0.009608848951756954, + 0.009859715588390827, + 0.007889222353696823, + 0.00612118374556303, + 0.008717029355466366, + 0.006248410325497389, + 0.0055922712199389935, + 0.010439882054924965, + 0.007755196187645197, + 0.00893092155456543, + 0.008657978847622871, + 0.008768666535615921, + 0.005881847348064184, + 0.0076742335222661495, + 0.00964270904660225, + 0.023524731397628784, + 0.007635017391294241, + 0.011696276254951954, + 0.0054301670752465725, + 0.006842297036200762, + 0.004498526453971863, + 0.005577020347118378, + 0.005731239914894104, + 0.00865922961384058, + 0.011028887704014778, + 0.004583926871418953, + 0.010493669658899307, + 0.0070335171185433865, + 0.008122765459120274, + 0.008836374618113041, + 0.007626770995557308, + 0.01565292477607727, + 0.014096860773861408, + 0.009240233339369297, + 0.008105197921395302, + 0.007378125563263893, + 0.009619690477848053, + 0.006521891802549362, + 0.005849137436598539, + 0.017195742577314377, + 0.016774721443653107, + 0.009820476174354553, + 0.0021765597630292177, + 0.0014150463975965977, + 0.0029324679635465145, + 0.0029861389193683863, + 0.0024550568778067827, + 0.003330553648993373, + 0.004032706376165152, + 0.0032598685938864946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.5234121084213257 + } + ], + "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/0000125/group_3/index_tensor([2031, 2031])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_3/index_tensor([2031, 2031])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2030d19cbb20ffb4ade2cbc31524c6303443fbb3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_3/index_tensor([2031, 2031])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking, low-angle fisheye lens follows a tall, muscular athlete, his eyes wide and teeth clenched beneath his signature thick, black unibrow, his face strained with intense concentration. Clad in a bright red, pixelated jersey and white shorts, he clutches a brown, blocky football in his arms, sprinting forward with unstoppable force. His world is composed entirely of three-dimensional pixels: beneath him lies a grassy field of varying shades of green, and every powerful kick sends tiny fragments flying into the air. The extreme distortion of the fisheye lens distorts this already geometric world into a spherical shape. The distant finish line—a striking strip of black and white—is stretched and magnified in the center of the field of view, while the spectator stands, stacked with multicolored blocks, curve sharply outward, forming exaggerated arcs. The entire picture is illuminated by bright, even top light, with extremely high color saturation, creating a retro yet surreal video game climax atmosphere, full of tense sense of speed and unstoppable kinetic energy.", + "seed": 13663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_3/index_2031_seed_13663_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": -0.021699166437610985, + "vlm_logprob": -0.021699166437610985, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.021699166437610985, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011638598516583443 + ], + "opposite_logprob_tokens": [ + -4.511638641357422 + ], + "logprob_reward": -0.011638598516583443, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011638598516583443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6355403661727905, + 0.09451515972614288, + 0.0658598393201828, + 0.046779144555330276, + 0.017643960192799568, + 0.023803649470210075, + 0.019681265577673912, + 0.05251618102192879, + 0.03196028247475624, + 0.011700139380991459 + ], + "credit_map": [ + 0.04440191388130188, + 0.0435357466340065, + 0.20161937177181244, + 0.03599578142166138, + 0.05036451295018196, + 0.03453851863741875, + 0.0366891473531723, + 0.18839535117149353, + 0.009889249689877033, + 0.015960579738020897, + 0.013347926549613476, + 0.012919868342578411, + 0.010103988461196423, + 0.006882976274937391, + 0.013614475727081299, + 0.01179608702659607, + 0.008132846094667912, + 0.007220747880637646, + 0.007233237847685814, + 0.010114307515323162, + 0.010527233593165874, + 0.008202309720218182, + 0.006596545688807964, + 0.007832605391740799, + 0.006807256489992142, + 0.006676108110696077, + 0.0037142219953238964, + 0.0042641558684408665, + 0.009646517224609852, + 0.0037063388153910637, + 0.004304147325456142, + 0.007660395465791225, + 0.0025271549820899963, + 0.00177634391002357, + 0.0017386655090376735, + 0.0032444458920508623, + 0.002232662169262767, + 0.0016418397426605225, + 0.0021390521433204412, + 0.0023437945637851954, + 0.0036992088425904512, + 0.003934936597943306, + 0.0025118112098425627, + 0.0033211165573447943, + 0.002344684675335884, + 0.0026053832843899727, + 0.0031502998899668455, + 0.0022362053859978914, + 0.0032651096116751432, + 0.0032521625980734825, + 0.0014520069817081094, + 0.002744832308962941, + 0.0024313160683959723, + 0.0021935398690402508, + 0.002485664328560233, + 0.0018566310172900558, + 0.005915087182074785, + 0.00904797948896885, + 0.003184502013027668, + 0.007391605991870165, + 0.006669299211353064, + 0.005781366024166346, + 0.008120458573102951, + 0.006405878812074661, + 0.004044434987008572, + 0.0037414736580103636, + 0.0027398488018661737, + 0.0038920887745916843, + 0.005653465632349253, + 0.0050536696799099445, + 0.0029740026220679283, + 0.0038612938951700926, + 0.002614371245726943, + 0.0016877918969839811, + 0.000826830742880702, + 0.0013853382552042603, + 0.0013283774023875594, + 0.001254257163964212, + 0.0010662900749593973, + 0.0015368815511465073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011638598516583443 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636727273464203 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636727273464203, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636727273464203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.508982241153717, + 0.1501385122537613, + 0.029550792649388313, + 0.04899635165929794, + 0.02430044673383236, + 0.031957317143678665, + 0.050871457904577255, + 0.09757505357265472, + 0.030763890594244003, + 0.026863936334848404 + ], + "credit_map": [ + 0.039861466735601425, + 0.042347367852926254, + 0.1459980607032776, + 0.031130345538258553, + 0.04905693605542183, + 0.029008081182837486, + 0.038046229630708694, + 0.13353371620178223, + 0.01792173460125923, + 0.02072540856897831, + 0.016907623037695885, + 0.017716165632009506, + 0.0257148128002882, + 0.013586585409939289, + 0.01442402321845293, + 0.02314216084778309, + 0.003949400503188372, + 0.004376598633825779, + 0.002201158320531249, + 0.0038120606914162636, + 0.0064978571608662605, + 0.0030635870061814785, + 0.0031632089521735907, + 0.0024869218468666077, + 0.0063033862970769405, + 0.005515994504094124, + 0.004630726762115955, + 0.004837160464376211, + 0.009250043891370296, + 0.004137012176215649, + 0.005914278328418732, + 0.00840775016695261, + 0.003927511163055897, + 0.002728631254285574, + 0.0020696143619716167, + 0.003354292828589678, + 0.004665229003876448, + 0.002090025693178177, + 0.0031824130564928055, + 0.0022827291395515203, + 0.006241489201784134, + 0.004849810618907213, + 0.0029467556159943342, + 0.0035181567072868347, + 0.00430384511128068, + 0.0036733904853463173, + 0.002822642447426915, + 0.003601224161684513, + 0.008621841669082642, + 0.008341746404767036, + 0.0051117511466145515, + 0.007236042059957981, + 0.005508571397513151, + 0.0061772847548127174, + 0.0044087497517466545, + 0.005465466063469648, + 0.011570344679057598, + 0.01963183656334877, + 0.00706692086532712, + 0.014855366200208664, + 0.01324219536036253, + 0.007644856348633766, + 0.014508143067359924, + 0.009055392816662788, + 0.007860511541366577, + 0.003481782041490078, + 0.0024133767001330853, + 0.0035366469528526068, + 0.00303117698058486, + 0.0036791888996958733, + 0.002741422737017274, + 0.004019786603748798, + 0.00555190397426486, + 0.0037199361249804497, + 0.0020352406427264214, + 0.0032637363765388727, + 0.0032217525877058506, + 0.0037302246782928705, + 0.0020947162993252277, + 0.0032464272808283567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004636727273464203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106613501906395 + ], + "opposite_logprob_tokens": [ + -4.51106595993042 + ], + "logprob_reward": -0.01106613501906395, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106613501906395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.305300235748291, + 0.21014130115509033, + 0.06341477483510971, + 0.08480485528707504, + 0.0750197172164917, + 0.0463329516351223, + 0.05523737892508507, + 0.07672808319330215, + 0.05742844566702843, + 0.025592252612113953 + ], + "credit_map": [ + 0.02182677760720253, + 0.027263719588518143, + 0.07554245740175247, + 0.023084569722414017, + 0.03335903212428093, + 0.01780080422759056, + 0.03718201816082001, + 0.06924083828926086, + 0.022677989676594734, + 0.03546338528394699, + 0.02571556344628334, + 0.032706499099731445, + 0.017472684383392334, + 0.021303575485944748, + 0.02832673490047455, + 0.02647487260401249, + 0.007944264449179173, + 0.006093480158597231, + 0.010306493379175663, + 0.011266756802797318, + 0.006948068272322416, + 0.0062762838788330555, + 0.007001593243330717, + 0.007577830925583839, + 0.011038707569241524, + 0.009826170280575752, + 0.007031966466456652, + 0.010086570866405964, + 0.016915783286094666, + 0.005949738901108503, + 0.011395754292607307, + 0.012560158967971802, + 0.006909760180860758, + 0.006324063055217266, + 0.009688004851341248, + 0.015612057410180569, + 0.0060004559345543385, + 0.012080864049494267, + 0.010705120861530304, + 0.007699390407651663, + 0.004398747812956572, + 0.005537841934710741, + 0.00373068917542696, + 0.009206360206007957, + 0.005664102733135223, + 0.006256904918700457, + 0.007026804611086845, + 0.0045114983804523945, + 0.005898481700569391, + 0.008509395644068718, + 0.005279256496578455, + 0.009716086089611053, + 0.008409438654780388, + 0.006188647821545601, + 0.00629056803882122, + 0.0049454993568360806, + 0.01459579635411501, + 0.008849342353641987, + 0.005913608707487583, + 0.011585439555346966, + 0.00847198162227869, + 0.009947956539690495, + 0.009624912403523922, + 0.007739043794572353, + 0.007815194316208363, + 0.007094198372215033, + 0.0044793132692575455, + 0.0070171174593269825, + 0.008766830898821354, + 0.0069310348480939865, + 0.005474559497088194, + 0.009850199334323406, + 0.0050809565000236034, + 0.0034686499275267124, + 0.0017791044665500522, + 0.0034186369739472866, + 0.002524008508771658, + 0.0037340179551392794, + 0.0024525669869035482, + 0.0031343100126832724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01106613501906395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00219533103518188 + ], + "opposite_logprob_tokens": [ + -6.127195358276367 + ], + "logprob_reward": -0.00219533103518188, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00219533103518188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3818749487400055, + 0.14429070055484772, + 0.1592588573694229, + 0.05257828161120415, + 0.025795621797442436, + 0.031207073479890823, + 0.024617208167910576, + 0.09290652722120285, + 0.05608043074607849, + 0.03139035031199455 + ], + "credit_map": [ + 0.04278535023331642, + 0.05528317019343376, + 0.08456182479858398, + 0.032706551253795624, + 0.034069664776325226, + 0.023859074339270592, + 0.02384888380765915, + 0.08476044237613678, + 0.018058227375149727, + 0.018492596223950386, + 0.022105319425463676, + 0.015800826251506805, + 0.016093725338578224, + 0.012525486759841442, + 0.01549672894179821, + 0.02571779489517212, + 0.018781336024403572, + 0.01859848015010357, + 0.01666523702442646, + 0.015280366875231266, + 0.02365848422050476, + 0.030700843781232834, + 0.015338714234530926, + 0.02023538388311863, + 0.007364523597061634, + 0.005574934650212526, + 0.006372600328177214, + 0.006109519861638546, + 0.008098859339952469, + 0.00482855224981904, + 0.005601188633590937, + 0.00862810667604208, + 0.003368681063875556, + 0.003229565219953656, + 0.002320396713912487, + 0.00424264557659626, + 0.003444892819970846, + 0.0034458721056580544, + 0.0028333566151559353, + 0.002910211682319641, + 0.00508582778275013, + 0.006263835821300745, + 0.002284675370901823, + 0.004019622225314379, + 0.0050440202467143536, + 0.002703901147469878, + 0.0029125644359737635, + 0.002892628312110901, + 0.004150880500674248, + 0.003909215331077576, + 0.0025147167034447193, + 0.002943813567981124, + 0.0030318114440888166, + 0.002646137960255146, + 0.0026653327513486147, + 0.002755299210548401, + 0.014843602664768696, + 0.013894762843847275, + 0.006019785534590483, + 0.01289356779307127, + 0.012156156823039055, + 0.009573562070727348, + 0.013614262454211712, + 0.009910824708640575, + 0.01376263052225113, + 0.006764007266610861, + 0.0047545419074594975, + 0.007225136272609234, + 0.0058092521503567696, + 0.0065664006397128105, + 0.005167920608073473, + 0.006030542775988579, + 0.006608029827475548, + 0.0034022836480289698, + 0.0028404397889971733, + 0.004671860486268997, + 0.0033565799240022898, + 0.0039928206242620945, + 0.0025841081514954567, + 0.003934225998818874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00219533103518188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895904034376144 + ], + "opposite_logprob_tokens": [ + -2.5789589881896973 + ], + "logprob_reward": -0.07895904034376144, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895904034376144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3425980806350708, + 0.24350842833518982, + 0.04006796330213547, + 0.02259257808327675, + 0.023320723325014114, + 0.03339104726910591, + 0.06755459308624268, + 0.14929670095443726, + 0.04826653376221657, + 0.029403334483504295 + ], + "credit_map": [ + 0.0318794809281826, + 0.04124870523810387, + 0.05580669641494751, + 0.07736149430274963, + 0.03079705312848091, + 0.018200701102614403, + 0.031144626438617706, + 0.0561593659222126, + 0.03513706848025322, + 0.02693384326994419, + 0.034019164741039276, + 0.028699496760964394, + 0.029120689257979393, + 0.023867670446634293, + 0.022634850814938545, + 0.043095652014017105, + 0.005039495415985584, + 0.004456529393792152, + 0.0035213096998631954, + 0.00496431952342391, + 0.004701334983110428, + 0.00808379054069519, + 0.0025341433938592672, + 0.0067670405842363834, + 0.0031625228002667427, + 0.003148220246657729, + 0.002041763626039028, + 0.002765441546216607, + 0.003432618221268058, + 0.0019517007749527693, + 0.0028554098680615425, + 0.0032349000684916973, + 0.003172077937051654, + 0.003968257922679186, + 0.00231382017955184, + 0.0026293897535651922, + 0.0027587274089455605, + 0.0028476722072809935, + 0.0030802199617028236, + 0.002550556557253003, + 0.005259796977043152, + 0.005480132065713406, + 0.0030755505431443453, + 0.004252721555531025, + 0.004800721071660519, + 0.003699843306094408, + 0.00358826806768775, + 0.003234011586755514, + 0.011084121651947498, + 0.01108981017023325, + 0.006900745443999767, + 0.008912337012588978, + 0.007271601352840662, + 0.009452399797737598, + 0.006402541883289814, + 0.006441033910959959, + 0.016864201053977013, + 0.026005107909440994, + 0.009116078726947308, + 0.0254655834287405, + 0.019218439236283302, + 0.014013990759849548, + 0.02422480285167694, + 0.014388494193553925, + 0.008826060220599174, + 0.006667173933237791, + 0.0034645190462470055, + 0.006580192595720291, + 0.005793051328510046, + 0.006153539288789034, + 0.004321616142988205, + 0.006460382137447596, + 0.005114984232932329, + 0.0036144652403891087, + 0.002729834523051977, + 0.00393256638199091, + 0.003853874746710062, + 0.0033897715620696545, + 0.0028951363638043404, + 0.0038727009668946266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.07895904034376144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, low-angle fisheye lens follows a tall, muscular athlete, his eyes wide and teeth clenched beneath his signature thick, black unibrow, his face strained with intense concentration. Clad in a bright red, pixelated jersey and white shorts, he clutches a brown, blocky football in his arms, sprinting forward with unstoppable force. His world is composed entirely of three-dimensional pixels: beneath him lies a grassy field of varying shades of green, and every powerful kick sends tiny fragments flying into the air. The extreme distortion of the fisheye lens distorts this already geometric world into a spherical shape. The distant finish line—a striking strip of black and white—is stretched and magnified in the center of the field of view, while the spectator stands, stacked with multicolored blocks, curve sharply outward, forming exaggerated arcs. The entire picture is illuminated by bright, even top light, with extremely high color saturation, creating a retro yet surreal video game climax atmosphere, full of tense sense of speed and unstoppable kinetic energy.", + "seed": 13664, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_3/index_2031_seed_13664_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": -0.026647272007539868, + "vlm_logprob": -0.026647272007539868, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.026647272007539868, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03874647617340088 + ], + "opposite_logprob_tokens": [ + -3.2887463569641113 + ], + "logprob_reward": -0.03874647617340088, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03874647617340088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22502534091472626, + 0.15035995841026306, + 0.038033805787563324, + 0.028451988473534584, + 0.07210902869701385, + 0.06653692573308945, + 0.12574905157089233, + 0.17615266144275665, + 0.07205279171466827, + 0.04552846774458885 + ], + "credit_map": [ + 0.06625418365001678, + 0.01492699608206749, + 0.033794134855270386, + 0.02085292711853981, + 0.01414791215211153, + 0.027755025774240494, + 0.025875041261315346, + 0.021419117227196693, + 0.02075132727622986, + 0.01727491430938244, + 0.008123508654534817, + 0.015616354532539845, + 0.028065837919712067, + 0.00943505298346281, + 0.04180383309721947, + 0.009289120323956013, + 0.004498675465583801, + 0.003481789492070675, + 0.003691748483106494, + 0.00512545183300972, + 0.006793577689677477, + 0.006931515876203775, + 0.004254892002791166, + 0.0032561523839831352, + 0.003775116056203842, + 0.002843311754986644, + 0.003814662341028452, + 0.003121219575405121, + 0.0023718634620308876, + 0.005289961118251085, + 0.0029248835053294897, + 0.004310972057282925, + 0.013817786239087582, + 0.008218200877308846, + 0.011052994057536125, + 0.009373749606311321, + 0.007344461511820555, + 0.008368179202079773, + 0.006511991377919912, + 0.007421664893627167, + 0.005911785177886486, + 0.006296096369624138, + 0.010414826683700085, + 0.008303536102175713, + 0.015612320974469185, + 0.005180560518056154, + 0.009847771376371384, + 0.004970034118741751, + 0.024209991097450256, + 0.007316846866160631, + 0.009948354214429855, + 0.014123335480690002, + 0.009116573259234428, + 0.038497138768434525, + 0.009044034406542778, + 0.01349277701228857, + 0.023902229964733124, + 0.01907590962946415, + 0.015671081840991974, + 0.02258557640016079, + 0.014773888513445854, + 0.018175840377807617, + 0.044026732444763184, + 0.017941400408744812, + 0.005979531444609165, + 0.005157568957656622, + 0.011025425978004932, + 0.0069986144080758095, + 0.013774368911981583, + 0.014120102860033512, + 0.006623289547860622, + 0.00837389100342989, + 0.006844285875558853, + 0.0035239052958786488, + 0.003563179634511471, + 0.00466381898149848, + 0.004802444484084845, + 0.006698692217469215, + 0.003919378854334354, + 0.011512762866914272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03874647617340088 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.18147684633731842, + 0.13851068913936615, + 0.05954517051577568, + 0.1255633682012558, + 0.07661733776330948, + 0.037974074482917786, + 0.05394954979419708, + 0.1525161862373352, + 0.06555648148059845, + 0.10829029977321625 + ], + "credit_map": [ + 0.014479792676866055, + 0.012053487822413445, + 0.016021769493818283, + 0.019859638065099716, + 0.014057110995054245, + 0.04571240395307541, + 0.04546526074409485, + 0.013827381655573845, + 0.02222728170454502, + 0.010102705098688602, + 0.011905551888048649, + 0.014058149419724941, + 0.014513298869132996, + 0.010881914757192135, + 0.0460696741938591, + 0.008752093650400639, + 0.006502863951027393, + 0.006683585699647665, + 0.007079662289470434, + 0.008308809250593185, + 0.011499791406095028, + 0.007198089733719826, + 0.005932142958045006, + 0.006340226158499718, + 0.01523847971111536, + 0.013836936093866825, + 0.015634922310709953, + 0.01572202518582344, + 0.01598830707371235, + 0.02118059992790222, + 0.013600017875432968, + 0.014362084679305553, + 0.015446039848029613, + 0.0067231967113912106, + 0.021840298548340797, + 0.009080618619918823, + 0.005653596017509699, + 0.00570671958848834, + 0.0049750348553061485, + 0.007191839162260294, + 0.004978001117706299, + 0.0035758349113166332, + 0.0044091735035181046, + 0.007038559298962355, + 0.006589960772544146, + 0.0031426982022821903, + 0.004806477576494217, + 0.0034333711955696344, + 0.006211331579834223, + 0.0053460304625332355, + 0.0029479239601641893, + 0.0076487623155117035, + 0.009022414684295654, + 0.011091550812125206, + 0.005422726273536682, + 0.006258809939026833, + 0.01604652777314186, + 0.01738603413105011, + 0.024742314592003822, + 0.01956678181886673, + 0.024783821776509285, + 0.013866588473320007, + 0.020560117438435555, + 0.015564005821943283, + 0.005068058613687754, + 0.00462590204551816, + 0.005274413153529167, + 0.008582577109336853, + 0.008400081656873226, + 0.015242291614413261, + 0.006752111483365297, + 0.011611048132181168, + 0.015699753537774086, + 0.009533013217151165, + 0.008419876918196678, + 0.010744399391114712, + 0.009913005866110325, + 0.015114892274141312, + 0.010146775282919407, + 0.028718581423163414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0036202622577548027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341667398810387 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.026341667398810387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341667398810387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14216654002666473, + 0.23717035353183746, + 0.05042966082692146, + 0.04720814526081085, + 0.10157513618469238, + 0.061744194477796555, + 0.044815532863140106, + 0.08418770879507065, + 0.20685625076293945, + 0.0238465815782547 + ], + "credit_map": [ + 0.01672685332596302, + 0.009265883825719357, + 0.035606466233730316, + 0.016869334504008293, + 0.010100189596414566, + 0.02958470769226551, + 0.00981990434229374, + 0.014193182811141014, + 0.04996396601200104, + 0.01796678453683853, + 0.0210918877273798, + 0.024877939373254776, + 0.016585595905780792, + 0.015020633116364479, + 0.07602902501821518, + 0.015634508803486824, + 0.005380683578550816, + 0.004882889799773693, + 0.009593984112143517, + 0.008081686682999134, + 0.006878362502902746, + 0.006479498464614153, + 0.004734196700155735, + 0.004398354794830084, + 0.006510418839752674, + 0.004674823954701424, + 0.007946922443807125, + 0.007498876191675663, + 0.005715698469430208, + 0.0035382357891649008, + 0.005039543379098177, + 0.006283618044108152, + 0.015782147645950317, + 0.009274768643081188, + 0.02295178733766079, + 0.016567103564739227, + 0.009134606458246708, + 0.009354106150567532, + 0.010392747819423676, + 0.0081178592517972, + 0.005325575824826956, + 0.006904880050569773, + 0.0057820966467261314, + 0.011763193644583225, + 0.010950548574328423, + 0.005822473671287298, + 0.008888368494808674, + 0.006307054311037064, + 0.00437241792678833, + 0.004136934410780668, + 0.004182789474725723, + 0.007574096787720919, + 0.005588209722191095, + 0.004186953883618116, + 0.0059938677586615086, + 0.008780259639024734, + 0.008725784718990326, + 0.00855997297912836, + 0.01154548954218626, + 0.010281462222337723, + 0.012630223296582699, + 0.008471446111798286, + 0.014226926490664482, + 0.009746397845447063, + 0.016800381243228912, + 0.014210546389222145, + 0.014583607204258442, + 0.022487947717308998, + 0.03659280389547348, + 0.05069621652364731, + 0.019746417179703712, + 0.03173832222819328, + 0.003355472581461072, + 0.002219690475612879, + 0.0019467200618237257, + 0.003033524611964822, + 0.003546151565387845, + 0.0021995895076543093, + 0.0035219741985201836, + 0.004023456946015358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.026341667398810387 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024889700580388308 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.0024889700580388308, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024889700580388308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12125347554683685, + 0.083603635430336, + 0.09378408640623093, + 0.04348883777856827, + 0.08365985751152039, + 0.12592536211013794, + 0.09952066093683243, + 0.13671553134918213, + 0.174112930893898, + 0.037935562431812286 + ], + "credit_map": [ + 0.011234179139137268, + 0.008222722448408604, + 0.01947673037648201, + 0.012385208159685135, + 0.006498207338154316, + 0.023190757259726524, + 0.02748822793364525, + 0.012757445685565472, + 0.014852630905807018, + 0.006748845335096121, + 0.009026474319398403, + 0.007749953772872686, + 0.010644789785146713, + 0.005292017478495836, + 0.0223655104637146, + 0.006923414766788483, + 0.010973663069307804, + 0.011261016130447388, + 0.013002412393689156, + 0.015983207151293755, + 0.008619804866611958, + 0.014736950397491455, + 0.009562362916767597, + 0.009644675999879837, + 0.006601979024708271, + 0.0038949446752667427, + 0.005506016314029694, + 0.006080176681280136, + 0.006705736741423607, + 0.005168314557522535, + 0.004260415676981211, + 0.005271256435662508, + 0.010566533543169498, + 0.008973291143774986, + 0.015844233334064484, + 0.013061837293207645, + 0.010181892663240433, + 0.009212746284902096, + 0.008252487517893314, + 0.007566840387880802, + 0.012363304384052753, + 0.016652455553412437, + 0.013491113670170307, + 0.030094124376773834, + 0.012025789357721806, + 0.009515724144876003, + 0.01848192885518074, + 0.013300909660756588, + 0.0187571682035923, + 0.006394439842551947, + 0.006856485735625029, + 0.00918932817876339, + 0.0078020100481808186, + 0.030480848625302315, + 0.00730358250439167, + 0.012736796401441097, + 0.02294081822037697, + 0.008012693375349045, + 0.011387175880372524, + 0.013709274120628834, + 0.008747299201786518, + 0.007857772521674633, + 0.05152786895632744, + 0.01253262534737587, + 0.012211396358907223, + 0.013566176407039165, + 0.017721563577651978, + 0.01866462454199791, + 0.03509948030114174, + 0.03752852976322174, + 0.0170974712818861, + 0.022223690524697304, + 0.005416231229901314, + 0.0031394909601658583, + 0.004505613818764687, + 0.0043587046675384045, + 0.003909230697900057, + 0.004849336575716734, + 0.0035543255507946014, + 0.008202625438570976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.0024889700580388308 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203898414969444 + ], + "opposite_logprob_tokens": [ + -2.8120388984680176 + ], + "logprob_reward": -0.06203898414969444, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203898414969444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2734975516796112, + 0.08675610274076462, + 0.07822972536087036, + 0.033579472452402115, + 0.0837073102593422, + 0.03458378463983536, + 0.09938612580299377, + 0.10513051599264145, + 0.18518707156181335, + 0.019942423328757286 + ], + "credit_map": [ + 0.027605174109339714, + 0.016065362840890884, + 0.05221055820584297, + 0.034126296639442444, + 0.022067386656999588, + 0.041143447160720825, + 0.049439989030361176, + 0.030839312821626663, + 0.014209273271262646, + 0.009465519338846207, + 0.008585128001868725, + 0.008783703669905663, + 0.012340241111814976, + 0.005430870223790407, + 0.02243215963244438, + 0.005509198643267155, + 0.011949393898248672, + 0.006592178717255592, + 0.009535960853099823, + 0.010059808380901814, + 0.004970909561961889, + 0.021224940195679665, + 0.006039231549948454, + 0.007857299409806728, + 0.004085188265889883, + 0.0027363740373402834, + 0.00465362099930644, + 0.004288790747523308, + 0.0031388027127832174, + 0.005759591236710548, + 0.0027273136656731367, + 0.006189784966409206, + 0.00884444173425436, + 0.011224821209907532, + 0.010678313672542572, + 0.012843490578234196, + 0.012605959549546242, + 0.011009463109076023, + 0.008883945643901825, + 0.007616867311298847, + 0.0038851802237331867, + 0.0032386789098381996, + 0.005349633749574423, + 0.00411178870126605, + 0.00685320608317852, + 0.003087487770244479, + 0.005050568375736475, + 0.0030072408262640238, + 0.015513278543949127, + 0.007925419136881828, + 0.008598561398684978, + 0.016183262690901756, + 0.006961900275200605, + 0.025008568540215492, + 0.008597441017627716, + 0.01059768721461296, + 0.01672546938061714, + 0.007407265715301037, + 0.011044614017009735, + 0.011208289302885532, + 0.007940566167235374, + 0.007231724914163351, + 0.03275347873568535, + 0.010819097980856895, + 0.013528045266866684, + 0.012956897728145123, + 0.02890501357614994, + 0.01747722737491131, + 0.026201365515589714, + 0.0470859669148922, + 0.016037873923778534, + 0.022994650527834892, + 0.002809581346809864, + 0.0019062486244365573, + 0.0020580689888447523, + 0.0024956127163022757, + 0.0025565701071172953, + 0.0025124847888946533, + 0.0025805134791880846, + 0.0030233401339501143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.06203898414969444 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, low-angle fisheye lens follows a tall, muscular athlete, his eyes wide and teeth clenched beneath his signature thick, black unibrow, his face strained with intense concentration. Clad in a bright red, pixelated jersey and white shorts, he clutches a brown, blocky football in his arms, sprinting forward with unstoppable force. His world is composed entirely of three-dimensional pixels: beneath him lies a grassy field of varying shades of green, and every powerful kick sends tiny fragments flying into the air. The extreme distortion of the fisheye lens distorts this already geometric world into a spherical shape. The distant finish line—a striking strip of black and white—is stretched and magnified in the center of the field of view, while the spectator stands, stacked with multicolored blocks, curve sharply outward, forming exaggerated arcs. The entire picture is illuminated by bright, even top light, with extremely high color saturation, creating a retro yet surreal video game climax atmosphere, full of tense sense of speed and unstoppable kinetic energy.", + "seed": 13665, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_3/index_2031_seed_13665_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": -0.022846662532538177, + "vlm_logprob": -0.022846662532538177, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.022846662532538177, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013019553385674953 + ], + "opposite_logprob_tokens": [ + -4.513019561767578 + ], + "logprob_reward": -0.013019553385674953, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013019553385674953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16516682505607605, + 0.21537023782730103, + 0.08706694096326828, + 0.025586197152733803, + 0.1924138367176056, + 0.0627410039305687, + 0.11725331842899323, + 0.023610644042491913, + 0.07498390227556229, + 0.03580711781978607 + ], + "credit_map": [ + 0.015070769935846329, + 0.023777637630701065, + 0.02267768420279026, + 0.01693142019212246, + 0.02189032919704914, + 0.013835019432008266, + 0.02574225328862667, + 0.02524169348180294, + 0.0368681475520134, + 0.02120720036327839, + 0.06535721570253372, + 0.026206625625491142, + 0.020902559161186218, + 0.016671614721417427, + 0.013652040623128414, + 0.0145048126578331, + 0.012475553900003433, + 0.009483013302087784, + 0.009210536256432533, + 0.017384853214025497, + 0.008914312347769737, + 0.010728508234024048, + 0.008806979283690453, + 0.01006318163126707, + 0.004765281453728676, + 0.002971319481730461, + 0.002784944837912917, + 0.002705069025978446, + 0.003058007452636957, + 0.004654728341847658, + 0.0019299625419080257, + 0.0027168807573616505, + 0.027553468942642212, + 0.03730772063136101, + 0.02799636498093605, + 0.02039482817053795, + 0.022802526131272316, + 0.022868933156132698, + 0.01603119634091854, + 0.017458787187933922, + 0.005029505118727684, + 0.005719832144677639, + 0.005666008684784174, + 0.009297029115259647, + 0.0131862573325634, + 0.005411038640886545, + 0.010998755693435669, + 0.00743256788700819, + 0.007014002185314894, + 0.005173653829842806, + 0.03195847198367119, + 0.007843279279768467, + 0.004751905798912048, + 0.010222148150205612, + 0.0117950439453125, + 0.03849480673670769, + 0.002424520207569003, + 0.0035803006030619144, + 0.0017936110962182283, + 0.0023645726032555103, + 0.004352610558271408, + 0.0020915500354021788, + 0.003504213411360979, + 0.0034992622677236795, + 0.010990669950842857, + 0.00962082203477621, + 0.006003158632665873, + 0.00836419127881527, + 0.0061505944468081, + 0.015931742265820503, + 0.009972330182790756, + 0.007950382307171822, + 0.0048746392130851746, + 0.003449505427852273, + 0.0031310904305428267, + 0.0045478735119104385, + 0.004420110490173101, + 0.004893552977591753, + 0.006203989498317242, + 0.004286355804651976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.013019553385674953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252846051007509 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252846051007509, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252846051007509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2223362773656845, + 0.0667448565363884, + 0.1540141999721527, + 0.09221870452165604, + 0.09044180810451508, + 0.0610998198390007, + 0.10268181562423706, + 0.03645222634077072, + 0.08384981006383896, + 0.09016051143407822 + ], + "credit_map": [ + 0.037240345031023026, + 0.02637477032840252, + 0.02243596874177456, + 0.02035536989569664, + 0.02418677881360054, + 0.03204952925443649, + 0.02882097102701664, + 0.030872542411088943, + 0.00586318876594305, + 0.007291365880519152, + 0.005960606969892979, + 0.02043936215341091, + 0.008721023797988892, + 0.005967026576399803, + 0.006764620076864958, + 0.005737662315368652, + 0.012353845871984959, + 0.0133878318592906, + 0.015638679265975952, + 0.022560186684131622, + 0.0370255783200264, + 0.019123824313282967, + 0.019085686653852463, + 0.014838576316833496, + 0.01296690758317709, + 0.010423236526548862, + 0.011218025349080563, + 0.009167253039777279, + 0.01683005690574646, + 0.009449977427721024, + 0.010425537824630737, + 0.01173771545290947, + 0.012297543697059155, + 0.017119016498327255, + 0.01250450313091278, + 0.010519023053348064, + 0.011795996688306332, + 0.008546452037990093, + 0.007745082955807447, + 0.00991419330239296, + 0.004828141536563635, + 0.004963801708072424, + 0.006185867823660374, + 0.009286784566938877, + 0.011709257028996944, + 0.004883561283349991, + 0.011146092787384987, + 0.008096317760646343, + 0.00833488255739212, + 0.004422102589160204, + 0.027998246252536774, + 0.007238775957375765, + 0.005011576693505049, + 0.00809413380920887, + 0.01119370199739933, + 0.030388399958610535, + 0.003712680656462908, + 0.003161897649988532, + 0.002934893826022744, + 0.00541089428588748, + 0.004047378897666931, + 0.0037345890887081623, + 0.007661151699721813, + 0.005788743030279875, + 0.007729603908956051, + 0.012114329263567924, + 0.006581097841262817, + 0.007486360613256693, + 0.011710403487086296, + 0.012142919935286045, + 0.014456558972597122, + 0.01162853091955185, + 0.012045026756823063, + 0.008211910724639893, + 0.008909119293093681, + 0.015869854018092155, + 0.00979426596313715, + 0.009247364476323128, + 0.015265600755810738, + 0.010817370377480984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.005252846051007509 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418695505708456 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.01418695505708456, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418695505708456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10611650347709656, + 0.18699532747268677, + 0.060116108506917953, + 0.023336132988333702, + 0.11354400217533112, + 0.06791115552186966, + 0.3108098804950714, + 0.0621943324804306, + 0.03607051447033882, + 0.03290611132979393 + ], + "credit_map": [ + 0.010836792178452015, + 0.013322343118488789, + 0.010624948889017105, + 0.013961591757833958, + 0.013684961944818497, + 0.014335127547383308, + 0.013461562804877758, + 0.01588916964828968, + 0.052820879966020584, + 0.012423262931406498, + 0.06577074527740479, + 0.02275608479976654, + 0.008823136799037457, + 0.006626345682889223, + 0.008770080283284187, + 0.009004777297377586, + 0.012032072991132736, + 0.0068140821531414986, + 0.006040562409907579, + 0.010235014371573925, + 0.008801417425274849, + 0.005499256309121847, + 0.0050504449754953384, + 0.0056432513520121574, + 0.0039434838108718395, + 0.0024308692663908005, + 0.0028880364261567593, + 0.0024685394018888474, + 0.003954184241592884, + 0.0031170763541013002, + 0.002298942534253001, + 0.00223500095307827, + 0.01749487966299057, + 0.023147715255618095, + 0.017643509432673454, + 0.01098062563687563, + 0.014604596421122551, + 0.011641661636531353, + 0.008223881013691425, + 0.009807123802602291, + 0.005542296916246414, + 0.0069586955942213535, + 0.0063782138749957085, + 0.009386159479618073, + 0.012852765619754791, + 0.0066908253356814384, + 0.011593835428357124, + 0.008508355356752872, + 0.019518032670021057, + 0.01156078465282917, + 0.08661350607872009, + 0.02100660465657711, + 0.012072240933775902, + 0.025751203298568726, + 0.031935591250658035, + 0.10235191881656647, + 0.005768760573118925, + 0.004937422927469015, + 0.0055632214061915874, + 0.010261868126690388, + 0.009555717930197716, + 0.005233457311987877, + 0.012152842245995998, + 0.008721032179892063, + 0.003153298981487751, + 0.004236347507685423, + 0.0027794037014245987, + 0.0042723724618554115, + 0.0051107993349432945, + 0.00548939174041152, + 0.005915434565395117, + 0.00511346198618412, + 0.0032133653294295073, + 0.004160765558481216, + 0.003963574301451445, + 0.004339737817645073, + 0.003985290415585041, + 0.002937071956694126, + 0.005628748796880245, + 0.0046775550581514835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.01418695505708456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817234955728054 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817234955728054, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817234955728054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17104634642601013, + 0.14322753250598907, + 0.1041913628578186, + 0.05715760216116905, + 0.03715328872203827, + 0.04556826129555702, + 0.3253210783004761, + 0.0285407155752182, + 0.04685071110725403, + 0.04094306007027626 + ], + "credit_map": [ + 0.026112250983715057, + 0.0307016484439373, + 0.01612972840666771, + 0.014269402250647545, + 0.02411239594221115, + 0.022563356906175613, + 0.014693164266645908, + 0.022464385256171227, + 0.03327552601695061, + 0.009449826553463936, + 0.04725179448723793, + 0.018457017838954926, + 0.012549122795462608, + 0.006918375380337238, + 0.007859920151531696, + 0.007465951144695282, + 0.01293997373431921, + 0.009968835860490799, + 0.00922432355582714, + 0.016309209167957306, + 0.019307753071188927, + 0.016546443104743958, + 0.010897831059992313, + 0.008996989578008652, + 0.008152595721185207, + 0.006169901695102453, + 0.009375370107591152, + 0.006881883833557367, + 0.005643133074045181, + 0.009484472684562206, + 0.0052879988215863705, + 0.006162247620522976, + 0.004681054502725601, + 0.006917624268680811, + 0.005242298822849989, + 0.0035724774934351444, + 0.004961567930877209, + 0.004373463802039623, + 0.0036275468301028013, + 0.0037772534415125847, + 0.00457781134173274, + 0.0056099677458405495, + 0.004958645906299353, + 0.005261546932160854, + 0.00710242660716176, + 0.006296386476606131, + 0.00594399543479085, + 0.005817481316626072, + 0.019996171817183495, + 0.012471696361899376, + 0.08951257169246674, + 0.021831093356013298, + 0.012992721050977707, + 0.027303103357553482, + 0.033558595925569534, + 0.1076551303267479, + 0.0024839905090630054, + 0.002862250665202737, + 0.0024413978680968285, + 0.005468865390866995, + 0.004700532648712397, + 0.0021662686485797167, + 0.005199366249144077, + 0.0032180454581975937, + 0.006636238656938076, + 0.006719783879816532, + 0.0036759062204509974, + 0.004560051951557398, + 0.005852098111063242, + 0.006823779083788395, + 0.0075846402905881405, + 0.004998211283236742, + 0.00616101361811161, + 0.004320025444030762, + 0.006308429408818483, + 0.003581785364076495, + 0.004410711582750082, + 0.003921250347048044, + 0.004656380042433739, + 0.007583462167531252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.002817234955728054 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789567232131958 + ], + "opposite_logprob_tokens": [ + -2.5789566040039062 + ], + "logprob_reward": -0.0789567232131958, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789567232131958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18184010684490204, + 0.14525556564331055, + 0.07218887656927109, + 0.053224507719278336, + 0.04636405035853386, + 0.049554478377103806, + 0.36887240409851074, + 0.024148203432559967, + 0.04053835570812225, + 0.01801348105072975 + ], + "credit_map": [ + 0.030584068968892097, + 0.03230246901512146, + 0.019167162477970123, + 0.01879573054611683, + 0.020930521190166473, + 0.023628758266568184, + 0.013846555724740028, + 0.0225848238915205, + 0.023472663015127182, + 0.011934613808989525, + 0.03752169385552406, + 0.028766373172402382, + 0.014664232730865479, + 0.008776356466114521, + 0.011276631616055965, + 0.00884298887103796, + 0.008357384242117405, + 0.006065085530281067, + 0.009995142929255962, + 0.010281185619533062, + 0.010999404825270176, + 0.011545608751475811, + 0.00840011890977621, + 0.006544935051351786, + 0.008655248209834099, + 0.005963923409581184, + 0.009077400900423527, + 0.007024358958005905, + 0.005540365353226662, + 0.006005721166729927, + 0.005092008970677853, + 0.005865477491170168, + 0.005779706407338381, + 0.0074869561940431595, + 0.006083971820771694, + 0.006250795442610979, + 0.006312293000519276, + 0.003967371769249439, + 0.0048215510323643684, + 0.0056614032946527, + 0.005214004777371883, + 0.005933945998549461, + 0.005379694513976574, + 0.004991760011762381, + 0.008749008178710938, + 0.0055295126512646675, + 0.007192583754658699, + 0.006563965231180191, + 0.022904694080352783, + 0.0139862559735775, + 0.10191047936677933, + 0.025003653019666672, + 0.014751910232007504, + 0.031160477548837662, + 0.03809567913413048, + 0.12105921655893326, + 0.001859093434177339, + 0.0018815912771970034, + 0.0018587345257401466, + 0.004532786086201668, + 0.004025030881166458, + 0.0021556769497692585, + 0.004656060133129358, + 0.0031792265363037586, + 0.005616704002022743, + 0.006295098457485437, + 0.0033939885906875134, + 0.003736863611266017, + 0.004895871505141258, + 0.006723201833665371, + 0.005709580611437559, + 0.004167045466601849, + 0.002402182202786207, + 0.0017461151583120227, + 0.0016500841593369842, + 0.0025721522979438305, + 0.002649031113833189, + 0.0017122689168900251, + 0.0028673321940004826, + 0.002414312679320574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0789567232131958 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, low-angle fisheye lens follows a tall, muscular athlete, his eyes wide and teeth clenched beneath his signature thick, black unibrow, his face strained with intense concentration. Clad in a bright red, pixelated jersey and white shorts, he clutches a brown, blocky football in his arms, sprinting forward with unstoppable force. His world is composed entirely of three-dimensional pixels: beneath him lies a grassy field of varying shades of green, and every powerful kick sends tiny fragments flying into the air. The extreme distortion of the fisheye lens distorts this already geometric world into a spherical shape. The distant finish line—a striking strip of black and white—is stretched and magnified in the center of the field of view, while the spectator stands, stacked with multicolored blocks, curve sharply outward, forming exaggerated arcs. The entire picture is illuminated by bright, even top light, with extremely high color saturation, creating a retro yet surreal video game climax atmosphere, full of tense sense of speed and unstoppable kinetic energy.", + "seed": 13666, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_3/index_2031_seed_13666_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": -0.02632259698584676, + "vlm_logprob": -0.02632259698584676, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02632259698584676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02068927325308323 + ], + "opposite_logprob_tokens": [ + -4.020689487457275 + ], + "logprob_reward": -0.02068927325308323, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02068927325308323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13035865128040314, + 0.23505020141601562, + 0.07318387180566788, + 0.05585223063826561, + 0.08810704946517944, + 0.0595150962471962, + 0.04064440354704857, + 0.058653853833675385, + 0.16276490688323975, + 0.09586974233388901 + ], + "credit_map": [ + 0.023502446711063385, + 0.013106776401400566, + 0.021810749545693398, + 0.02012317068874836, + 0.008899216540157795, + 0.009310893714427948, + 0.012555696070194244, + 0.021049706265330315, + 0.03603660315275192, + 0.020020630210638046, + 0.035345569252967834, + 0.027663346379995346, + 0.02893911488354206, + 0.019751517102122307, + 0.030578874051570892, + 0.036714546382427216, + 0.007894308306276798, + 0.008280680514872074, + 0.007375817745923996, + 0.01177488174289465, + 0.01409458089619875, + 0.006633280776441097, + 0.008100401610136032, + 0.009029917418956757, + 0.006658272352069616, + 0.005097674671560526, + 0.006244860123842955, + 0.012398532591760159, + 0.005705936811864376, + 0.004281910136342049, + 0.005851265508681536, + 0.009613779373466969, + 0.011812735348939896, + 0.009273689240217209, + 0.009963935241103172, + 0.019568949937820435, + 0.01014829520136118, + 0.008353929966688156, + 0.01040634699165821, + 0.008579166606068611, + 0.005633289460092783, + 0.005851461552083492, + 0.00748615013435483, + 0.010234513320028782, + 0.008068435825407505, + 0.007920399308204651, + 0.007055317517369986, + 0.007265526335686445, + 0.004555798601359129, + 0.005347824189811945, + 0.004515497013926506, + 0.006721372250467539, + 0.006132692098617554, + 0.0034155873581767082, + 0.005071339663118124, + 0.004884290974587202, + 0.00831560231745243, + 0.00810731016099453, + 0.008979453705251217, + 0.008870127610862255, + 0.007253018207848072, + 0.004774572793394327, + 0.005135910585522652, + 0.007217856589704752, + 0.01425319816917181, + 0.025945890694856644, + 0.022853650152683258, + 0.019513878971338272, + 0.013259009458124638, + 0.015875114127993584, + 0.017033524811267853, + 0.034030646085739136, + 0.013161560520529747, + 0.012362445704638958, + 0.012473535723984241, + 0.009246833622455597, + 0.014424518682062626, + 0.01054645050317049, + 0.013800345361232758, + 0.009854045696556568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.02068927325308323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067379651591181755 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.0067379651591181755, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067379651591181755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14909332990646362, + 0.2193698287010193, + 0.11636119335889816, + 0.050785329192876816, + 0.060642629861831665, + 0.08337563276290894, + 0.04678094759583473, + 0.07085905969142914, + 0.07384460419416428, + 0.12888741493225098 + ], + "credit_map": [ + 0.025436362251639366, + 0.014326700009405613, + 0.02306249551475048, + 0.022212499752640724, + 0.010578709654510021, + 0.014414351433515549, + 0.014965680427849293, + 0.024096541106700897, + 0.027915220707654953, + 0.025916310027241707, + 0.03046761453151703, + 0.02858254872262478, + 0.019435513764619827, + 0.01924288459122181, + 0.033541545271873474, + 0.03426820784807205, + 0.014070629142224789, + 0.013578911311924458, + 0.010852918960154057, + 0.017152560874819756, + 0.017220625653862953, + 0.009735575877130032, + 0.020486624911427498, + 0.013263355009257793, + 0.006542779039591551, + 0.00597057631239295, + 0.006281191948801279, + 0.006293451879173517, + 0.006880942266434431, + 0.005034083034843206, + 0.0069810389541089535, + 0.006801267620176077, + 0.008109284564852715, + 0.012124842032790184, + 0.006560313981026411, + 0.008750684559345245, + 0.006300938781350851, + 0.006202398799359798, + 0.005694575607776642, + 0.00689959479495883, + 0.007146934047341347, + 0.007146934047341347, + 0.012646016664803028, + 0.0169348306953907, + 0.008728617802262306, + 0.013455617241561413, + 0.007526970002800226, + 0.009789712727069855, + 0.007028101943433285, + 0.005818143952637911, + 0.006144246552139521, + 0.007609164342284203, + 0.005092083476483822, + 0.003953856416046619, + 0.0053292145021259785, + 0.005806136876344681, + 0.011442556977272034, + 0.008783433586359024, + 0.009233447723090649, + 0.010394352488219738, + 0.00899233017116785, + 0.008747415617108345, + 0.005352548323571682, + 0.007912983186542988, + 0.010588578879833221, + 0.006434560287743807, + 0.007298193871974945, + 0.01001405157148838, + 0.005421758163720369, + 0.008795293048024178, + 0.008197491057217121, + 0.017094681039452553, + 0.022755661979317665, + 0.010125244036316872, + 0.010908172465860844, + 0.011166360229253769, + 0.02614709362387657, + 0.01692984811961651, + 0.02051299624145031, + 0.010342034511268139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0067379651591181755 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106967218220234 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.01106967218220234, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106967218220234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09377845376729965, + 0.23337231576442719, + 0.10235974937677383, + 0.02247103862464428, + 0.040075648576021194, + 0.04412960261106491, + 0.01809929497539997, + 0.0365777425467968, + 0.3811646103858948, + 0.027971558272838593 + ], + "credit_map": [ + 0.015961021184921265, + 0.008934265933930874, + 0.01408357359468937, + 0.014993560500442982, + 0.009809516370296478, + 0.007621973752975464, + 0.0090196393430233, + 0.013354906812310219, + 0.030297882854938507, + 0.023604094982147217, + 0.03645055368542671, + 0.028023218736052513, + 0.01994810253381729, + 0.01901794783771038, + 0.030259931460022926, + 0.04577060416340828, + 0.015209604986011982, + 0.011692804284393787, + 0.011296642012894154, + 0.01884508691728115, + 0.01169074047356844, + 0.006791239138692617, + 0.01681193895637989, + 0.0100216930732131, + 0.0029449297580868006, + 0.002718720817938447, + 0.002427416853606701, + 0.004030042793601751, + 0.002719288691878319, + 0.002740664640441537, + 0.0028398039285093546, + 0.0020501723047345877, + 0.00411233538761735, + 0.005010151769965887, + 0.003475737292319536, + 0.007780127227306366, + 0.005040137097239494, + 0.004538111854344606, + 0.00427066907286644, + 0.005848378874361515, + 0.003747486975044012, + 0.0037212411407381296, + 0.005406595300883055, + 0.009165504015982151, + 0.004325200337916613, + 0.0069107916206121445, + 0.005463093519210815, + 0.005389689467847347, + 0.002267025876790285, + 0.0015623559011146426, + 0.002500650705769658, + 0.0029901238158345222, + 0.0024653479922562838, + 0.0019326764158904552, + 0.0017040036618709564, + 0.0026771104894578457, + 0.005789624527096748, + 0.004730056505650282, + 0.003923072945326567, + 0.005078684072941542, + 0.00593005632981658, + 0.0035022154916077852, + 0.0031755203381180763, + 0.004448512569069862, + 0.03406607732176781, + 0.08457939326763153, + 0.05886680260300636, + 0.03976207599043846, + 0.022628139704465866, + 0.03222544491291046, + 0.03646464645862579, + 0.07257203012704849, + 0.00449859444051981, + 0.0036313741002231836, + 0.002962235826998949, + 0.002530236029997468, + 0.004903498105704784, + 0.00303041934967041, + 0.0040151504799723625, + 0.002400050638243556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.01106967218220234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095263313502073 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095263313502073, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095263313502073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2150457501411438, + 0.12834736704826355, + 0.1656508594751358, + 0.054723434150218964, + 0.06808117032051086, + 0.057890329509973526, + 0.04126753658056259, + 0.09613782167434692, + 0.11435190588235855, + 0.05850377306342125 + ], + "credit_map": [ + 0.028783155605196953, + 0.026415793225169182, + 0.03246307745575905, + 0.030185110867023468, + 0.024120446294546127, + 0.019209034740924835, + 0.022145887836813927, + 0.03172324597835541, + 0.025349756702780724, + 0.010850832797586918, + 0.021922484040260315, + 0.011462085880339146, + 0.015664566308259964, + 0.01074681244790554, + 0.012833526358008385, + 0.019517309963703156, + 0.020834151655435562, + 0.014434944838285446, + 0.018259558826684952, + 0.027959195896983147, + 0.022989092394709587, + 0.013470224104821682, + 0.033048029989004135, + 0.014655671082437038, + 0.005831476300954819, + 0.007676813751459122, + 0.006757297553122044, + 0.010978310368955135, + 0.00581135181710124, + 0.004878591746091843, + 0.007375354878604412, + 0.005414236336946487, + 0.01058259792625904, + 0.013137507252395153, + 0.006955529563128948, + 0.01338726282119751, + 0.005131115205585957, + 0.004748104605823755, + 0.005956597160547972, + 0.00818245206028223, + 0.007044398225843906, + 0.0061694560572505, + 0.007471880875527859, + 0.009875246323645115, + 0.007107672281563282, + 0.0065470607951283455, + 0.007449067197740078, + 0.006225548684597015, + 0.008859873749315739, + 0.005383492447435856, + 0.005877391435205936, + 0.005111945327371359, + 0.003667686600238085, + 0.00522886635735631, + 0.003310508793219924, + 0.0038277709390968084, + 0.012050963006913662, + 0.012407598085701466, + 0.01037115603685379, + 0.017505256459116936, + 0.012054581195116043, + 0.007788761053234339, + 0.01045218389481306, + 0.013507331721484661, + 0.013889878988265991, + 0.02461019903421402, + 0.014535878784954548, + 0.013703973963856697, + 0.007243844214826822, + 0.010515078902244568, + 0.010769685730338097, + 0.01908336579799652, + 0.007482378743588924, + 0.007609113119542599, + 0.0066955373622477055, + 0.006657525431364775, + 0.009751737117767334, + 0.0063664671033620834, + 0.008812503889203072, + 0.005128505639731884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004095263313502073 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, muscular athlete wearing a bright red pixelated jersey and white shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete's face characterized by wide eyes, clenched teeth, and a thick black unibrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete sprinting forward while clutching a brown, blocky football?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a 3D pixel art style with a grassy field of varying green shades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle fisheye lens that distorts the scene into a spherical shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902081102132797 + ], + "opposite_logprob_tokens": [ + -2.4640207290649414 + ], + "logprob_reward": -0.08902081102132797, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902081102132797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17566123604774475, + 0.12485288828611374, + 0.1711127758026123, + 0.07948736101388931, + 0.046739909797906876, + 0.04725160822272301, + 0.030043430626392365, + 0.07274580001831055, + 0.2219814658164978, + 0.030123572796583176 + ], + "credit_map": [ + 0.020529575645923615, + 0.023312579840421677, + 0.030046634376049042, + 0.02111101895570755, + 0.014790002256631851, + 0.017017843201756477, + 0.0199994258582592, + 0.028854139149188995, + 0.023766258731484413, + 0.01471148245036602, + 0.017601927742362022, + 0.0138075677677989, + 0.013852303847670555, + 0.00920630432665348, + 0.015151727944612503, + 0.016755307093262672, + 0.037307146936655045, + 0.016578994691371918, + 0.021572859957814217, + 0.040256399661302567, + 0.014739129692316055, + 0.010863237082958221, + 0.014064001850783825, + 0.015730978921055794, + 0.013282131403684616, + 0.009061001241207123, + 0.010155325755476952, + 0.016915583983063698, + 0.006214207038283348, + 0.00653053866699338, + 0.0075064292177557945, + 0.009822135791182518, + 0.007451684679836035, + 0.012538798153400421, + 0.005121351685374975, + 0.007247099187225103, + 0.004014821723103523, + 0.0033815433271229267, + 0.0031925307121127844, + 0.003792070085182786, + 0.005262061487883329, + 0.005139087326824665, + 0.006321996916085482, + 0.0101548433303833, + 0.004798322916030884, + 0.004709168337285519, + 0.00539708137512207, + 0.005469044204801321, + 0.005807294510304928, + 0.003248505759984255, + 0.004525785334408283, + 0.00422454159706831, + 0.003033979795873165, + 0.0025069222319871187, + 0.0028931775595992804, + 0.0038032180164009333, + 0.014090289361774921, + 0.009285928681492805, + 0.01067754440009594, + 0.009904824197292328, + 0.005754257086664438, + 0.010394667275249958, + 0.005817415192723274, + 0.006820873357355595, + 0.02370525151491165, + 0.026819102466106415, + 0.023875487968325615, + 0.027774879708886147, + 0.014238935895264149, + 0.024198610335588455, + 0.025836408138275146, + 0.055532749742269516, + 0.0038667405024170876, + 0.0034763338044285774, + 0.004293488804250956, + 0.0029911024030297995, + 0.004973491188138723, + 0.003249600762501359, + 0.004447672050446272, + 0.0028251376934349537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.08902081102132797 + } + ], + "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/0000125/group_4/index_tensor([7391, 7391])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_4/index_tensor([7391, 7391])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1cf34409e772a2b460a23b516c8415ea9c8b8a5f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_4/index_tensor([7391, 7391])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, captured from an extremely low angle, follows a streamlined maglev freight train as it travels at breakneck speed. The train's bullet-like silhouette, its hull seamlessly crafted from polished silver alloy and matte carbon fiber, is adorned with glowing cyan-blue energy strips along its sides, signaling its powerful power source. Directly ahead, two piercing white LED headlights pierce the dim, industrial-looking hangar space like sharp swords. In front of the camera, a massive, immaculate mirror, towering from the ceiling, blocks the path, its surface clearly reflecting the train's roaring image, creating a surreal sense of symmetry. At the moment of impact, the scene abruptly shifts into a series of high-speed, dazzling close-up montages: first, the sharp nose of the train impacts the mirror, causing it to deform visibly; then, countless spider-web-like cracks spread rapidly from the point of impact at the speed of light; then, with a deafening, resonant explosion, the entire mirror suddenly explodes into a storm of billions of glittering glass shards. The camera rapidly switches between the metallic details of the train's nose and the sharp glass fragments tumbling through the air, reflecting the train's lights. Some fragments gleam like diamonds, while others are stretched into blurred light in their rapid motion. The entire scene is rendered with breathtakingly realistic effects, with the physical trajectory of each fragment and every refraction and reflection of light clearly visible, creating an extremely violent aesthetic and breathtaking visual impact, creating an atmosphere of tension and grandeur.", + "seed": 32035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_4/index_7391_seed_32035_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": -0.05978844796773046, + "vlm_logprob": -0.05978844796773046, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.05978844796773046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1305030733346939 + ], + "opposite_logprob_tokens": [ + -2.1305031776428223 + ], + "logprob_reward": -0.1305030733346939, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1305030733346939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1132998839020729, + 0.04447638988494873, + 0.032328661531209946, + 0.07565576583147049, + 0.13819237053394318, + 0.1882278174161911, + 0.2131664901971817, + 0.11341647803783417, + 0.027949431911110878, + 0.053286727517843246 + ], + "credit_map": [ + 0.014704260043799877, + 0.011269877664744854, + 0.015814710408449173, + 0.014581968076527119, + 0.012551608495414257, + 0.008283736184239388, + 0.016902651637792587, + 0.0191910769790411, + 0.003795165801420808, + 0.004473944194614887, + 0.007963864132761955, + 0.0061642504297196865, + 0.003936443477869034, + 0.0034037933219224215, + 0.007184167392551899, + 0.007554762996733189, + 0.0035706795752048492, + 0.002681775949895382, + 0.004726945422589779, + 0.00559616694226861, + 0.0026046298444271088, + 0.0032662309240549803, + 0.004511713981628418, + 0.0053705195896327496, + 0.006766513921320438, + 0.009876394644379616, + 0.012495629489421844, + 0.008300647139549255, + 0.0077817183919250965, + 0.008494623005390167, + 0.010346619412302971, + 0.011593623086810112, + 0.009223698638379574, + 0.009081454016268253, + 0.01308752316981554, + 0.01865527592599392, + 0.011036542244255543, + 0.015410144813358784, + 0.023345692083239555, + 0.038352031260728836, + 0.016356296837329865, + 0.010843058116734028, + 0.04377942159771919, + 0.036781422793865204, + 0.009425935335457325, + 0.01852274127304554, + 0.025697503238916397, + 0.026821432635188103, + 0.017989452928304672, + 0.026380058377981186, + 0.03368811309337616, + 0.030443795025348663, + 0.01968696340918541, + 0.01639958657324314, + 0.04164355620741844, + 0.02693496085703373, + 0.011282945051789284, + 0.01453295536339283, + 0.007163661066442728, + 0.01779361255466938, + 0.02944195829331875, + 0.01372212078422308, + 0.012089921161532402, + 0.007389305625110865, + 0.0037679062224924564, + 0.003669167635962367, + 0.0032771476544439793, + 0.003938803914934397, + 0.005220604129135609, + 0.0026708615478128195, + 0.0032077571377158165, + 0.002197185531258583, + 0.004762083757668734, + 0.008755630813539028, + 0.008318132720887661, + 0.009693264029920101, + 0.0059297168627381325, + 0.0035976795479655266, + 0.00650089280679822, + 0.005729326978325844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.1305030733346939 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003201361745595932 + ], + "opposite_logprob_tokens": [ + -5.753201484680176 + ], + "logprob_reward": -0.003201361745595932, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003201361745595932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04127095267176628, + 0.04537368565797806, + 0.48124775290489197, + 0.054550062865018845, + 0.06440185010433197, + 0.08126991242170334, + 0.059322286397218704, + 0.07697679847478867, + 0.03958144038915634, + 0.05600531026721001 + ], + "credit_map": [ + 0.005627312697470188, + 0.0043114605359733105, + 0.006011142395436764, + 0.005766143091022968, + 0.0037257999647408724, + 0.0031538463663309813, + 0.006875296123325825, + 0.0057999491691589355, + 0.0044098105281591415, + 0.006323165260255337, + 0.007551601622253656, + 0.005819775629788637, + 0.004835821688175201, + 0.003995181061327457, + 0.005171849858015776, + 0.007266482803970575, + 0.044932130724191666, + 0.03458292409777641, + 0.11403053998947144, + 0.05462293699383736, + 0.04673384130001068, + 0.039327893406152725, + 0.04800727963447571, + 0.09901019930839539, + 0.004871510434895754, + 0.005836350377649069, + 0.010725259780883789, + 0.008784902282059193, + 0.003874929388985038, + 0.006256286520510912, + 0.004974605515599251, + 0.009226217865943909, + 0.009144197218120098, + 0.006026187911629677, + 0.007222536951303482, + 0.008220452815294266, + 0.004575488623231649, + 0.007213910575956106, + 0.00819483958184719, + 0.013804229907691479, + 0.006987948901951313, + 0.00457442132756114, + 0.018747707828879356, + 0.015507973730564117, + 0.0034551890566945076, + 0.006185889709740877, + 0.012591605074703693, + 0.01321917213499546, + 0.005488727707415819, + 0.0049393996596336365, + 0.009818431921303272, + 0.008201313205063343, + 0.005814083851873875, + 0.004810915794223547, + 0.01075505930930376, + 0.009494355879724026, + 0.008810962550342083, + 0.009356693364679813, + 0.004557719454169273, + 0.010863958857953548, + 0.0195902269333601, + 0.009430178441107273, + 0.008584415540099144, + 0.005782645661383867, + 0.00365513376891613, + 0.004995614755898714, + 0.005788479000329971, + 0.0048287701793015, + 0.009682665579020977, + 0.0041631353087723255, + 0.0038587411399930716, + 0.0026089006569236517, + 0.0050228796899318695, + 0.0076747252605855465, + 0.007124273106455803, + 0.006833272986114025, + 0.009073778055608273, + 0.00777120515704155, + 0.005332733038812876, + 0.007172443438321352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.003201361745595932 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603344976902008 + ], + "opposite_logprob_tokens": [ + -1.9103344678878784 + ], + "logprob_reward": -0.1603344976902008, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603344976902008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21789582073688507, + 0.08932588249444962, + 0.15910544991493225, + 0.14043757319450378, + 0.07298456877470016, + 0.09972827136516571, + 0.07570670545101166, + 0.07940363883972168, + 0.026743093505501747, + 0.03866903483867645 + ], + "credit_map": [ + 0.029220810160040855, + 0.022183937951922417, + 0.02697354555130005, + 0.028445450589060783, + 0.02138923667371273, + 0.01642077974975109, + 0.028675543144345284, + 0.04458652064204216, + 0.01104728877544403, + 0.005922400392591953, + 0.014640788547694683, + 0.009887932799756527, + 0.007359974551945925, + 0.007386348210275173, + 0.014401007443666458, + 0.018680136650800705, + 0.01728685013949871, + 0.011189129203557968, + 0.03214612230658531, + 0.01868436112999916, + 0.015823636204004288, + 0.01495525985956192, + 0.01935841329395771, + 0.029661664739251137, + 0.012785961851477623, + 0.011604943312704563, + 0.02630086988210678, + 0.018206393346190453, + 0.010559910908341408, + 0.01627539098262787, + 0.022007962688803673, + 0.02269614487886429, + 0.008947665803134441, + 0.011211392469704151, + 0.008020288310945034, + 0.009910368360579014, + 0.005484419409185648, + 0.007403023075312376, + 0.009321840479969978, + 0.012685568071901798, + 0.011403479613363743, + 0.005963866598904133, + 0.01385115459561348, + 0.01730320230126381, + 0.0058072395622730255, + 0.012643270194530487, + 0.018566904589533806, + 0.014189156703650951, + 0.005862210411578417, + 0.004569461103528738, + 0.012035468593239784, + 0.016106870025396347, + 0.003906991798430681, + 0.007159970235079527, + 0.013434333726763725, + 0.012631403282284737, + 0.006914340425282717, + 0.007820761762559414, + 0.005238690413534641, + 0.01172673236578703, + 0.018532730638980865, + 0.009040449745953083, + 0.008483677171170712, + 0.01164625957608223, + 0.0033954705577343702, + 0.004226406570523977, + 0.0025546413380652666, + 0.0032075040508061647, + 0.004434858448803425, + 0.0031754288356751204, + 0.003386138239875436, + 0.002362647792324424, + 0.004465492442250252, + 0.008123846724629402, + 0.0035972853656858206, + 0.006261708214879036, + 0.0034877408761531115, + 0.004895448684692383, + 0.00427691824734211, + 0.003560593817383051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1603344976902008 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017107388703152537 + ], + "opposite_logprob_tokens": [ + -6.376710891723633 + ], + "logprob_reward": -0.0017107388703152537, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017107388703152537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08789757639169693, + 0.08262282609939575, + 0.060008544474840164, + 0.2519422471523285, + 0.07017973065376282, + 0.08290965855121613, + 0.06419897079467773, + 0.08633565157651901, + 0.08763569593429565, + 0.12626908719539642 + ], + "credit_map": [ + 0.01226071361452341, + 0.008087960071861744, + 0.0128175625577569, + 0.010411900468170643, + 0.007412489037960768, + 0.008208248764276505, + 0.013945101760327816, + 0.014753596857190132, + 0.008170042186975479, + 0.01126769557595253, + 0.01551771443337202, + 0.010359852574765682, + 0.008726230822503567, + 0.007958979345858097, + 0.0069424803368747234, + 0.013679837808012962, + 0.005359681323170662, + 0.004499002825468779, + 0.01163933053612709, + 0.008234377019107342, + 0.005975285079330206, + 0.0049119810573756695, + 0.009470200166106224, + 0.009918687865138054, + 0.0200581606477499, + 0.03041570819914341, + 0.03556247428059578, + 0.024957790970802307, + 0.026652423664927483, + 0.0246150940656662, + 0.037511155009269714, + 0.052169427275657654, + 0.005899811163544655, + 0.008890839293599129, + 0.006414507050067186, + 0.012074497528374195, + 0.005267488770186901, + 0.00923185795545578, + 0.009359549731016159, + 0.013041176833212376, + 0.009014101698994637, + 0.006472712382674217, + 0.009412318468093872, + 0.013300684280693531, + 0.006071866489946842, + 0.013495204038918018, + 0.013714156113564968, + 0.01142861507833004, + 0.007171336095780134, + 0.005745977163314819, + 0.010207239538431168, + 0.00874081626534462, + 0.00802636705338955, + 0.004877841100096703, + 0.010693734511733055, + 0.008735662326216698, + 0.00915909931063652, + 0.00803793128579855, + 0.006764667574316263, + 0.010308435186743736, + 0.021656237542629242, + 0.014243331737816334, + 0.007780826650559902, + 0.008385119959712029, + 0.006781851407140493, + 0.012193364091217518, + 0.0087259067222476, + 0.014782214537262917, + 0.019006401300430298, + 0.008320295251905918, + 0.01216171309351921, + 0.00566395279020071, + 0.009710108861327171, + 0.021583354100584984, + 0.018257956951856613, + 0.015069221146404743, + 0.016896363347768784, + 0.018846901133656502, + 0.010763546451926231, + 0.015141640789806843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0017107388703152537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031925681978464127 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031925681978464127, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031925681978464127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10521125048398972, + 0.06522686779499054, + 0.10367053747177124, + 0.10805831849575043, + 0.08356734365224838, + 0.16160652041435242, + 0.1212855651974678, + 0.09990322589874268, + 0.07709059119224548, + 0.07437969744205475 + ], + "credit_map": [ + 0.012878398410975933, + 0.01103580929338932, + 0.014790994115173817, + 0.014005257748067379, + 0.012103800661861897, + 0.009008626453578472, + 0.014396234415471554, + 0.01699213869869709, + 0.0053240228444337845, + 0.005814048927277327, + 0.011489927768707275, + 0.008162464946508408, + 0.005662248935550451, + 0.005214817356318235, + 0.009691542014479637, + 0.013867790810763836, + 0.009757493622601032, + 0.007643071468919516, + 0.02343122474849224, + 0.013522965833544731, + 0.009106175042688847, + 0.009146030060946941, + 0.011973011307418346, + 0.019090576097369194, + 0.008425029926002026, + 0.015215395949780941, + 0.015151566825807095, + 0.012757644057273865, + 0.01170438900589943, + 0.009314656257629395, + 0.017561782151460648, + 0.017927858978509903, + 0.0066112191416323185, + 0.004902839194983244, + 0.007378017529845238, + 0.011609341017901897, + 0.005745117552578449, + 0.009161427617073059, + 0.014251690357923508, + 0.023907693102955818, + 0.014729108661413193, + 0.013880910351872444, + 0.033406440168619156, + 0.025256218388676643, + 0.012766738422214985, + 0.015999875962734222, + 0.02362593449652195, + 0.0219412948936224, + 0.009567681699991226, + 0.012664941139519215, + 0.0239605363458395, + 0.024137940257787704, + 0.011222421191632748, + 0.006664480548352003, + 0.018579838797450066, + 0.014487729407846928, + 0.012789003551006317, + 0.009068334475159645, + 0.006023253779858351, + 0.011749687604606152, + 0.020239291712641716, + 0.013433186337351799, + 0.009946470148861408, + 0.01665399596095085, + 0.007048961706459522, + 0.010516565293073654, + 0.011078341864049435, + 0.010234509594738483, + 0.01321196649223566, + 0.008346323855221272, + 0.011154578067362309, + 0.005499341990798712, + 0.0059860507026314735, + 0.012795993126928806, + 0.010264193639159203, + 0.009477417916059494, + 0.010633480735123158, + 0.010594896972179413, + 0.006119410041719675, + 0.008508259430527687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.0031925681978464127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, captured from an extremely low angle, follows a streamlined maglev freight train as it travels at breakneck speed. The train's bullet-like silhouette, its hull seamlessly crafted from polished silver alloy and matte carbon fiber, is adorned with glowing cyan-blue energy strips along its sides, signaling its powerful power source. Directly ahead, two piercing white LED headlights pierce the dim, industrial-looking hangar space like sharp swords. In front of the camera, a massive, immaculate mirror, towering from the ceiling, blocks the path, its surface clearly reflecting the train's roaring image, creating a surreal sense of symmetry. At the moment of impact, the scene abruptly shifts into a series of high-speed, dazzling close-up montages: first, the sharp nose of the train impacts the mirror, causing it to deform visibly; then, countless spider-web-like cracks spread rapidly from the point of impact at the speed of light; then, with a deafening, resonant explosion, the entire mirror suddenly explodes into a storm of billions of glittering glass shards. The camera rapidly switches between the metallic details of the train's nose and the sharp glass fragments tumbling through the air, reflecting the train's lights. Some fragments gleam like diamonds, while others are stretched into blurred light in their rapid motion. The entire scene is rendered with breathtakingly realistic effects, with the physical trajectory of each fragment and every refraction and reflection of light clearly visible, creating an extremely violent aesthetic and breathtaking visual impact, creating an atmosphere of tension and grandeur.", + "seed": 32036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_4/index_7391_seed_32036_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": -0.036897188797593114, + "vlm_logprob": -0.036897188797593114, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.036897188797593114, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11863197386264801 + ], + "opposite_logprob_tokens": [ + -2.2436320781707764 + ], + "logprob_reward": -0.11863197386264801, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11863197386264801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1586543768644333, + 0.06856314837932587, + 0.17908066511154175, + 0.10421284288167953, + 0.10642192512750626, + 0.0759466290473938, + 0.06693795323371887, + 0.119435615837574, + 0.05918460339307785, + 0.0615621879696846 + ], + "credit_map": [ + 0.009312174282968044, + 0.013232802972197533, + 0.02504592575132847, + 0.020536376163363457, + 0.012903247028589249, + 0.019598083570599556, + 0.024071266874670982, + 0.03395450487732887, + 0.00742533290758729, + 0.0059003378264606, + 0.010933829471468925, + 0.008738142438232899, + 0.007702690549194813, + 0.00916460994631052, + 0.006399050820618868, + 0.012299162335693836, + 0.012839975766837597, + 0.02675958350300789, + 0.020961608737707138, + 0.03330012410879135, + 0.014739995822310448, + 0.025290487334132195, + 0.018940143287181854, + 0.026248764246702194, + 0.009918545372784138, + 0.01081505324691534, + 0.018609793856739998, + 0.01470142137259245, + 0.010800548829138279, + 0.013884802348911762, + 0.011165575124323368, + 0.01431710459291935, + 0.012550948187708855, + 0.01601717621088028, + 0.013798408210277557, + 0.011307114735245705, + 0.017730580642819405, + 0.010436179116368294, + 0.011775825172662735, + 0.012805698439478874, + 0.00933007150888443, + 0.008373021148145199, + 0.012291829101741314, + 0.010462270118296146, + 0.01366443745791912, + 0.006201558280736208, + 0.008127488195896149, + 0.0074959611520171165, + 0.007501944433897734, + 0.008143982850015163, + 0.008525286801159382, + 0.007453371305018663, + 0.010301203466951847, + 0.011169904842972755, + 0.006703737657517195, + 0.007138531655073166, + 0.013470140285789967, + 0.016688769683241844, + 0.014558500610291958, + 0.01894172839820385, + 0.011795999482274055, + 0.013128879480063915, + 0.02090446464717388, + 0.009947141632437706, + 0.005867027211934328, + 0.004622630309313536, + 0.008991640992462635, + 0.0053993589244782925, + 0.009845324791967869, + 0.008319029584527016, + 0.01053521316498518, + 0.005604382138699293, + 0.005270946305245161, + 0.00569425243884325, + 0.018848076462745667, + 0.006339980289340019, + 0.0036581256426870823, + 0.0051252939738333225, + 0.006159412208944559, + 0.01046610064804554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.11863197386264801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104998894035816 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.004104998894035816, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104998894035816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1782650500535965, + 0.08755292743444443, + 0.07493792474269867, + 0.10730864852666855, + 0.08284652233123779, + 0.0693202093243599, + 0.03834233805537224, + 0.03894966468214989, + 0.13922986388206482, + 0.1832469254732132 + ], + "credit_map": [ + 0.011058692820370197, + 0.017540501430630684, + 0.034777991473674774, + 0.02382185496389866, + 0.012863057665526867, + 0.019233139231801033, + 0.024219201877713203, + 0.03475058823823929, + 0.008643732406198978, + 0.012542466633021832, + 0.009989787824451923, + 0.007084837183356285, + 0.013651990331709385, + 0.01685376465320587, + 0.007799244485795498, + 0.010987096466124058, + 0.005663369316607714, + 0.013661613687872887, + 0.00794978067278862, + 0.009116062894463539, + 0.00644904188811779, + 0.011502978391945362, + 0.005719124339520931, + 0.014875943772494793, + 0.008502056822180748, + 0.011087309569120407, + 0.019264796748757362, + 0.018638359382748604, + 0.011948538944125175, + 0.011214213445782661, + 0.010193263180553913, + 0.016460096463561058, + 0.009561401791870594, + 0.008490198291838169, + 0.009572086855769157, + 0.013004065491259098, + 0.0075938086956739426, + 0.0064888303168118, + 0.012296573258936405, + 0.015839550644159317, + 0.008584821596741676, + 0.008008280768990517, + 0.011516226455569267, + 0.008536887355148792, + 0.013362988829612732, + 0.004712933674454689, + 0.008402413688600063, + 0.00619564950466156, + 0.004625015892088413, + 0.004305513575673103, + 0.004463101737201214, + 0.0044505479745566845, + 0.0070364344865083694, + 0.006272603757679462, + 0.002992133144289255, + 0.004196980968117714, + 0.00468032481148839, + 0.005570631008595228, + 0.004288190044462681, + 0.0071823387406766415, + 0.003846618114039302, + 0.004725709557533264, + 0.005028542596846819, + 0.003627307014539838, + 0.011606386862695217, + 0.013848296366631985, + 0.0220639631152153, + 0.01699265092611313, + 0.021541761234402657, + 0.014309554360806942, + 0.02357276901602745, + 0.015294458717107773, + 0.012476077303290367, + 0.014672912657260895, + 0.06107839569449425, + 0.014928922988474369, + 0.008037458173930645, + 0.012842192314565182, + 0.01463954709470272, + 0.04457138851284981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.004104998894035816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054942674934864044 + ], + "opposite_logprob_tokens": [ + -2.9299426078796387 + ], + "logprob_reward": -0.054942674934864044, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054942674934864044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46649906039237976, + 0.08587674796581268, + 0.07016429305076599, + 0.08031872659921646, + 0.10417955368757248, + 0.03637523576617241, + 0.04308663681149483, + 0.05005497857928276, + 0.04105863720178604, + 0.02238618955016136 + ], + "credit_map": [ + 0.031208079308271408, + 0.04306742548942566, + 0.0779654011130333, + 0.06504622101783752, + 0.060108572244644165, + 0.06490568816661835, + 0.06477343291044235, + 0.05942419171333313, + 0.009108774363994598, + 0.009009606204926968, + 0.011908118613064289, + 0.010453478433191776, + 0.009144140407443047, + 0.0117555633187294, + 0.0074877808801829815, + 0.017009275034070015, + 0.006200597155839205, + 0.009486924856901169, + 0.008776482194662094, + 0.00907011516392231, + 0.005944808479398489, + 0.012662481516599655, + 0.008566327393054962, + 0.009456554427742958, + 0.007648940198123455, + 0.007502731867134571, + 0.011662866920232773, + 0.011740247718989849, + 0.011841967701911926, + 0.009317122399806976, + 0.007902431301772594, + 0.012702411971986294, + 0.01061566174030304, + 0.012888828292489052, + 0.01335038524121046, + 0.01452513039112091, + 0.015174984000623226, + 0.010095388628542423, + 0.010869513265788555, + 0.016659650951623917, + 0.005199774168431759, + 0.0042486065067350864, + 0.00555701507255435, + 0.004406353924423456, + 0.005326311569660902, + 0.0029226846527308226, + 0.003483746200799942, + 0.005230741109699011, + 0.0056359609588980675, + 0.004437606316059828, + 0.006396246142685413, + 0.003992896061390638, + 0.006623629946261644, + 0.007384206634014845, + 0.003841239959001541, + 0.004774849396198988, + 0.004813866224139929, + 0.007449640892446041, + 0.0048114582896232605, + 0.005846453830599785, + 0.005327179562300444, + 0.008154898881912231, + 0.006814899854362011, + 0.006836575921624899, + 0.005115549545735121, + 0.0034913760609924793, + 0.00906557310372591, + 0.005528529640287161, + 0.0048690130934119225, + 0.0031581411603838205, + 0.00704578822478652, + 0.0027846647426486015, + 0.002662038430571556, + 0.0022124021779745817, + 0.005032883491367102, + 0.0020859481301158667, + 0.002304322551935911, + 0.0023654713295400143, + 0.002247034339234233, + 0.0034760853741317987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.054942674934864044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031890033278614283 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.0031890033278614283, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031890033278614283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06762126088142395, + 0.051751431077718735, + 0.09383034706115723, + 0.05675743892788887, + 0.03215811401605606, + 0.08027627319097519, + 0.03920615836977959, + 0.124200239777565, + 0.3893832564353943, + 0.06481541693210602 + ], + "credit_map": [ + 0.005212065763771534, + 0.0053449030965566635, + 0.011803674511611462, + 0.008554105646908283, + 0.00621627364307642, + 0.00799830723553896, + 0.008614970371127129, + 0.013876957818865776, + 0.005694248713552952, + 0.0049782926216721535, + 0.0065091392025351524, + 0.005029955878853798, + 0.004931322764605284, + 0.010799040086567402, + 0.004586025606840849, + 0.00922340340912342, + 0.005735210608690977, + 0.01657574065029621, + 0.011492341756820679, + 0.014897318556904793, + 0.006002381909638643, + 0.014269528910517693, + 0.008507078513503075, + 0.01635074056684971, + 0.007672385778278112, + 0.005677695851773024, + 0.010384817607700825, + 0.00812576524913311, + 0.005476719234138727, + 0.005959007889032364, + 0.005098613910377026, + 0.00836243201047182, + 0.00396541366353631, + 0.0036890984047204256, + 0.0046182721853256226, + 0.0045132408849895, + 0.0036874408833682537, + 0.002776010427623987, + 0.0036931552458554506, + 0.00521548418328166, + 0.011499104090034962, + 0.007983112707734108, + 0.013804049231112003, + 0.010156737640500069, + 0.011913364753127098, + 0.006208386737853289, + 0.010154769755899906, + 0.008556751534342766, + 0.005572082009166479, + 0.0052498080767691135, + 0.004549304023385048, + 0.004183659330010414, + 0.006159873213618994, + 0.005822754930704832, + 0.0031138909980654716, + 0.004554784391075373, + 0.011031250469386578, + 0.01820729859173298, + 0.012822439894080162, + 0.015233897604048252, + 0.013013487681746483, + 0.021343110129237175, + 0.017414424568414688, + 0.015134322457015514, + 0.03181033954024315, + 0.04477531835436821, + 0.04344074800610542, + 0.042399387806653976, + 0.0629347562789917, + 0.05324748158454895, + 0.06454004347324371, + 0.04623520001769066, + 0.008475939743220806, + 0.008048078045248985, + 0.0054220594465732574, + 0.00569930300116539, + 0.006036093458533287, + 0.009772274643182755, + 0.00435735983774066, + 0.017004311084747314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.0031890033278614283 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617292968556285 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.003617292968556285, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617292968556285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21635811030864716, + 0.04459980130195618, + 0.11780969053506851, + 0.08162988722324371, + 0.13717791438102722, + 0.1335069239139557, + 0.03864727541804314, + 0.06696200370788574, + 0.09714624285697937, + 0.066162109375 + ], + "credit_map": [ + 0.014508197084069252, + 0.019639208912849426, + 0.032762281596660614, + 0.020700329914689064, + 0.018782150000333786, + 0.02015519328415394, + 0.034265223890542984, + 0.05554550513625145, + 0.0032151497434824705, + 0.0062927925027906895, + 0.005944359116256237, + 0.003920829389244318, + 0.007278206292539835, + 0.007291591726243496, + 0.004194579552859068, + 0.006462291814386845, + 0.009961413219571114, + 0.019612902775406837, + 0.015972239896655083, + 0.015721358358860016, + 0.009941334836184978, + 0.02102840133011341, + 0.00995511095970869, + 0.015616921707987785, + 0.012337377294898033, + 0.011408348567783833, + 0.012950428761541843, + 0.00977677945047617, + 0.007934845052659512, + 0.008479968644678593, + 0.007808672729879618, + 0.010933462530374527, + 0.01621847227215767, + 0.02098938450217247, + 0.01888718083500862, + 0.01649957336485386, + 0.020321104675531387, + 0.013973343186080456, + 0.014775705523788929, + 0.015513155609369278, + 0.022329924628138542, + 0.013071522116661072, + 0.021102355793118477, + 0.01874840259552002, + 0.016905631870031357, + 0.012095547281205654, + 0.014079761691391468, + 0.01517377607524395, + 0.0052971686236560345, + 0.004359608516097069, + 0.0044540404342114925, + 0.003731307340785861, + 0.005514821968972683, + 0.007530457340180874, + 0.003476160578429699, + 0.004283712711185217, + 0.006901680026203394, + 0.009476594626903534, + 0.00811633188277483, + 0.011971953324973583, + 0.0059372782707214355, + 0.010672573931515217, + 0.007696707267314196, + 0.006188881117850542, + 0.009452282451093197, + 0.009166293777525425, + 0.010806244798004627, + 0.011043551377952099, + 0.016314219683408737, + 0.01214811485260725, + 0.019051874056458473, + 0.009163661859929562, + 0.004692982882261276, + 0.008355466648936272, + 0.013388033024966717, + 0.005202025640755892, + 0.003934791777282953, + 0.006790768355131149, + 0.003183424938470125, + 0.020614618435502052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.003617292968556285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, captured from an extremely low angle, follows a streamlined maglev freight train as it travels at breakneck speed. The train's bullet-like silhouette, its hull seamlessly crafted from polished silver alloy and matte carbon fiber, is adorned with glowing cyan-blue energy strips along its sides, signaling its powerful power source. Directly ahead, two piercing white LED headlights pierce the dim, industrial-looking hangar space like sharp swords. In front of the camera, a massive, immaculate mirror, towering from the ceiling, blocks the path, its surface clearly reflecting the train's roaring image, creating a surreal sense of symmetry. At the moment of impact, the scene abruptly shifts into a series of high-speed, dazzling close-up montages: first, the sharp nose of the train impacts the mirror, causing it to deform visibly; then, countless spider-web-like cracks spread rapidly from the point of impact at the speed of light; then, with a deafening, resonant explosion, the entire mirror suddenly explodes into a storm of billions of glittering glass shards. The camera rapidly switches between the metallic details of the train's nose and the sharp glass fragments tumbling through the air, reflecting the train's lights. Some fragments gleam like diamonds, while others are stretched into blurred light in their rapid motion. The entire scene is rendered with breathtakingly realistic effects, with the physical trajectory of each fragment and every refraction and reflection of light clearly visible, creating an extremely violent aesthetic and breathtaking visual impact, creating an atmosphere of tension and grandeur.", + "seed": 32037, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_4/index_7391_seed_32037_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": -0.07513467245735228, + "vlm_logprob": -0.07513467245735228, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.07513467245735228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2048565000295639 + ], + "opposite_logprob_tokens": [ + -1.704856514930725 + ], + "logprob_reward": -0.2048565000295639, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2048565000295639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07951627671718597, + 0.3417607247829437, + 0.10596106946468353, + 0.0959760919213295, + 0.09724878519773483, + 0.12670192122459412, + 0.03650699928402901, + 0.031391873955726624, + 0.060624152421951294, + 0.024312084540724754 + ], + "credit_map": [ + 0.01026289165019989, + 0.009159719571471214, + 0.009345888160169125, + 0.016887124627828598, + 0.007440368179231882, + 0.00858705397695303, + 0.009610111825168133, + 0.00822311919182539, + 0.07411422580480576, + 0.02476956695318222, + 0.12313047051429749, + 0.03004145808517933, + 0.023893270641565323, + 0.012774240225553513, + 0.038702841848134995, + 0.014334681443870068, + 0.007598093245178461, + 0.01445607841014862, + 0.019758205860853195, + 0.01397393736988306, + 0.009273719042539597, + 0.012414674274623394, + 0.011691959574818611, + 0.01679440774023533, + 0.006449752952903509, + 0.008537333458662033, + 0.00788571685552597, + 0.017412541434168816, + 0.010397831909358501, + 0.012710059992969036, + 0.02539527229964733, + 0.007187583018094301, + 0.009431625716388226, + 0.007935477420687675, + 0.023170700296759605, + 0.010146424174308777, + 0.006362698040902615, + 0.021611928939819336, + 0.009841308929026127, + 0.008748620748519897, + 0.01813066564500332, + 0.015103895217180252, + 0.01300590019673109, + 0.011146333999931812, + 0.020955881103873253, + 0.012510479427874088, + 0.018187759444117546, + 0.01766100898385048, + 0.0035946788266301155, + 0.0038369703106582165, + 0.0027065672911703587, + 0.004500388167798519, + 0.004135832190513611, + 0.004087778739631176, + 0.0049642538651824, + 0.00868052989244461, + 0.003753593424335122, + 0.0036226341035217047, + 0.0041017718613147736, + 0.004261635709553957, + 0.0032649205531924963, + 0.0047096130438148975, + 0.003970727324485779, + 0.0037069786339998245, + 0.005505032371729612, + 0.0061117494478821754, + 0.006861032452434301, + 0.004792260937392712, + 0.010808190330862999, + 0.013077754527330399, + 0.0064721680246293545, + 0.006995959207415581, + 0.0032624048180878162, + 0.003591854590922594, + 0.003582834731787443, + 0.0025375329423695803, + 0.0031092646531760693, + 0.003321267431601882, + 0.0021621009800583124, + 0.0027448241598904133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.2048565000295639 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017210922669619322 + ], + "opposite_logprob_tokens": [ + -6.376720905303955 + ], + "logprob_reward": -0.0017210922669619322, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017210922669619322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09789752215147018, + 0.2518436908721924, + 0.115074023604393, + 0.18941961228847504, + 0.053949467837810516, + 0.07187909632921219, + 0.06674449890851974, + 0.04072275385260582, + 0.05841981992125511, + 0.05404951423406601 + ], + "credit_map": [ + 0.015495878644287586, + 0.018001431599259377, + 0.01335415430366993, + 0.00996505469083786, + 0.013385227881371975, + 0.012114935554564, + 0.00854216143488884, + 0.007038669195026159, + 0.049709778279066086, + 0.01619029976427555, + 0.09242098033428192, + 0.020966073498129845, + 0.017611896619200706, + 0.009324654936790466, + 0.03465810790657997, + 0.010961885564029217, + 0.009021946229040623, + 0.010307695716619492, + 0.02473902702331543, + 0.014187021180987358, + 0.010070361196994781, + 0.01478057075291872, + 0.011556313373148441, + 0.020411085337400436, + 0.009548620320856571, + 0.013373085297644138, + 0.014817222021520138, + 0.0381600596010685, + 0.019377008080482483, + 0.03187955915927887, + 0.045957960188388824, + 0.01630609668791294, + 0.004884534981101751, + 0.004519640933722258, + 0.009251494891941547, + 0.0074823834002017975, + 0.004129958804696798, + 0.012288355268537998, + 0.004948817193508148, + 0.006444275379180908, + 0.0069543360732495785, + 0.008957432582974434, + 0.007025600876659155, + 0.00753692165017128, + 0.010496153496205807, + 0.010574753396213055, + 0.009668765589594841, + 0.01066512893885374, + 0.0075977942906320095, + 0.00826882105320692, + 0.005629335064440966, + 0.011683589778840542, + 0.011190010234713554, + 0.006563324015587568, + 0.007233244366943836, + 0.008578372187912464, + 0.004713952541351318, + 0.0040152897126972675, + 0.004902575630694628, + 0.003960845526307821, + 0.0026299788150936365, + 0.006563165690749884, + 0.0031686744187027216, + 0.010768270120024681, + 0.0079542500898242, + 0.005450396798551083, + 0.00813312642276287, + 0.004056193865835667, + 0.009442679584026337, + 0.009105033241212368, + 0.005616131238639355, + 0.008662008680403233, + 0.005993738304823637, + 0.006374940741807222, + 0.0066412221640348434, + 0.0057854605838656425, + 0.009453457780182362, + 0.009273592382669449, + 0.005564766004681587, + 0.004962335340678692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.0017210922669619322 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603318452835083 + ], + "opposite_logprob_tokens": [ + -1.9103318452835083 + ], + "logprob_reward": -0.1603318452835083, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603318452835083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16288574039936066, + 0.07148277014493942, + 0.15348120033740997, + 0.2388429492712021, + 0.08207020908594131, + 0.07763945311307907, + 0.06268372386693954, + 0.054193153977394104, + 0.03844216465950012, + 0.05827862769365311 + ], + "credit_map": [ + 0.02849567122757435, + 0.023218046873807907, + 0.02666371688246727, + 0.018629249185323715, + 0.017178939655423164, + 0.018499089404940605, + 0.014470665715634823, + 0.01573037914931774, + 0.007958680391311646, + 0.010982475243508816, + 0.010021192021667957, + 0.008200305514037609, + 0.008414478972554207, + 0.008371864445507526, + 0.010682121850550175, + 0.006851655431091785, + 0.012185437604784966, + 0.010057642124593258, + 0.02175792306661606, + 0.015570314601063728, + 0.015630386769771576, + 0.03002648800611496, + 0.01718353107571602, + 0.031069478020071983, + 0.02368117868900299, + 0.018241915851831436, + 0.028782935813069344, + 0.041048452258110046, + 0.024262867867946625, + 0.03278782218694687, + 0.04589671269059181, + 0.02414107881486416, + 0.005380580201745033, + 0.004908641800284386, + 0.022213784977793694, + 0.007887611165642738, + 0.00396059500053525, + 0.021573955193161964, + 0.008999744430184364, + 0.0071453009732067585, + 0.017413588240742683, + 0.00780866015702486, + 0.00682649714872241, + 0.006284546572715044, + 0.010094507597386837, + 0.00748023996129632, + 0.009964855387806892, + 0.011766565032303333, + 0.006836510729044676, + 0.007274975068867207, + 0.005099866539239883, + 0.008017734624445438, + 0.0085642384365201, + 0.006743297446519136, + 0.007839974015951157, + 0.012307127937674522, + 0.006407434586435556, + 0.006318732630461454, + 0.006418912205845118, + 0.005888345185667276, + 0.0039531816728413105, + 0.008476289920508862, + 0.004768900573253632, + 0.011961361393332481, + 0.004620921332389116, + 0.004311080556362867, + 0.004474600311368704, + 0.0028898294549435377, + 0.0059264288283884525, + 0.007479325868189335, + 0.0040660095401108265, + 0.004673973191529512, + 0.0053349705412983894, + 0.00785005372017622, + 0.00981774553656578, + 0.004979551769793034, + 0.012376334518194199, + 0.007333249319344759, + 0.004505375865846872, + 0.006081352476030588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.1603318452835083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059441933408379555 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.0059441933408379555, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059441933408379555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06611365079879761, + 0.2349054217338562, + 0.09676811844110489, + 0.1237170547246933, + 0.0383145771920681, + 0.08769530802965164, + 0.03362544625997543, + 0.1830882877111435, + 0.09462802112102509, + 0.04114418476819992 + ], + "credit_map": [ + 0.013735385611653328, + 0.009407112374901772, + 0.010834122076630592, + 0.007980507798492908, + 0.00691546406596899, + 0.007255504839122295, + 0.0052893809042871, + 0.00469617173075676, + 0.04996239021420479, + 0.01614321954548359, + 0.0878056138753891, + 0.019972382113337517, + 0.013456791639328003, + 0.00799410231411457, + 0.029826302081346512, + 0.009744598530232906, + 0.008543665520846844, + 0.008760805241763592, + 0.015757083892822266, + 0.01155834924429655, + 0.009102658368647099, + 0.015926048159599304, + 0.009926456958055496, + 0.017193038016557693, + 0.006203148048371077, + 0.00833543948829174, + 0.010199190117418766, + 0.03354896977543831, + 0.010077782906591892, + 0.01206163875758648, + 0.0343509279191494, + 0.008939957246184349, + 0.004484702367335558, + 0.003944989759474993, + 0.004308568313717842, + 0.004878298379480839, + 0.0031739443074911833, + 0.008664565160870552, + 0.004424194805324078, + 0.00443531246855855, + 0.014315607026219368, + 0.01086833793669939, + 0.008265722543001175, + 0.007796525023877621, + 0.011772574856877327, + 0.011064170859754086, + 0.01185647677630186, + 0.011755882762372494, + 0.004004289861768484, + 0.003849077271297574, + 0.003090097801759839, + 0.0039380439557135105, + 0.005465504713356495, + 0.003851892426609993, + 0.003940332215279341, + 0.005486203357577324, + 0.020739208906888962, + 0.012252140790224075, + 0.024437781423330307, + 0.0159015916287899, + 0.008965722285211086, + 0.028716374188661575, + 0.012223450466990471, + 0.0598519928753376, + 0.014120491221547127, + 0.009277461096644402, + 0.013599480502307415, + 0.007194855250418186, + 0.013392471708357334, + 0.012721146456897259, + 0.010539287701249123, + 0.01378281693905592, + 0.004619892220944166, + 0.006027474999427795, + 0.004857739899307489, + 0.00405757874250412, + 0.005715172737836838, + 0.006268633063882589, + 0.003954456653445959, + 0.005643235519528389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.0059441933408379555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04068692773580551, + 0.38198044896125793, + 0.02614176832139492, + 0.1169946938753128, + 0.05474444478750229, + 0.05585150793194771, + 0.05801878124475479, + 0.12016227841377258, + 0.030948838219046593, + 0.11447033286094666 + ], + "credit_map": [ + 0.005572068504989147, + 0.006288417614996433, + 0.008077633567154408, + 0.004191246349364519, + 0.004399984609335661, + 0.005215643905103207, + 0.0037468248046934605, + 0.0031951037235558033, + 0.08386240899562836, + 0.020255999639630318, + 0.14731961488723755, + 0.029820265248417854, + 0.020630236715078354, + 0.012150399386882782, + 0.053470611572265625, + 0.01447086688131094, + 0.002225881442427635, + 0.003341029165312648, + 0.0035662080626934767, + 0.002873387886211276, + 0.00255383993498981, + 0.0031902894843369722, + 0.0030306822154670954, + 0.005360445939004421, + 0.006093322299420834, + 0.006855588871985674, + 0.008214958012104034, + 0.035071611404418945, + 0.007515180855989456, + 0.01041897851973772, + 0.03533322736620903, + 0.007491813041269779, + 0.004072944633662701, + 0.004476677626371384, + 0.013981988653540611, + 0.005747233983129263, + 0.0033479875419288874, + 0.012017843313515186, + 0.005631337873637676, + 0.005468428134918213, + 0.008225062862038612, + 0.007398118264973164, + 0.004992166068404913, + 0.005579652264714241, + 0.0071357060223817825, + 0.007756656501442194, + 0.0074224816635251045, + 0.007341659627854824, + 0.007356763817369938, + 0.005894154775887728, + 0.0053583481349051, + 0.007868476212024689, + 0.006919422186911106, + 0.007301053963601589, + 0.007158593274652958, + 0.010161958634853363, + 0.012987213209271431, + 0.009237225167453289, + 0.011353010311722755, + 0.011010156013071537, + 0.007120257243514061, + 0.025521107017993927, + 0.00788942165672779, + 0.0350438728928566, + 0.0037786313332617283, + 0.0028084833174943924, + 0.004934420809149742, + 0.00238796416670084, + 0.005402859300374985, + 0.004404430743306875, + 0.003219218924641609, + 0.00401282636448741, + 0.010520944371819496, + 0.011698716320097446, + 0.024918416514992714, + 0.009008505381643772, + 0.0322263203561306, + 0.011911599896848202, + 0.0058752805925905704, + 0.008310549892485142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.002819731365889311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, captured from an extremely low angle, follows a streamlined maglev freight train as it travels at breakneck speed. The train's bullet-like silhouette, its hull seamlessly crafted from polished silver alloy and matte carbon fiber, is adorned with glowing cyan-blue energy strips along its sides, signaling its powerful power source. Directly ahead, two piercing white LED headlights pierce the dim, industrial-looking hangar space like sharp swords. In front of the camera, a massive, immaculate mirror, towering from the ceiling, blocks the path, its surface clearly reflecting the train's roaring image, creating a surreal sense of symmetry. At the moment of impact, the scene abruptly shifts into a series of high-speed, dazzling close-up montages: first, the sharp nose of the train impacts the mirror, causing it to deform visibly; then, countless spider-web-like cracks spread rapidly from the point of impact at the speed of light; then, with a deafening, resonant explosion, the entire mirror suddenly explodes into a storm of billions of glittering glass shards. The camera rapidly switches between the metallic details of the train's nose and the sharp glass fragments tumbling through the air, reflecting the train's lights. Some fragments gleam like diamonds, while others are stretched into blurred light in their rapid motion. The entire scene is rendered with breathtakingly realistic effects, with the physical trajectory of each fragment and every refraction and reflection of light clearly visible, creating an extremely violent aesthetic and breathtaking visual impact, creating an atmosphere of tension and grandeur.", + "seed": 32038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_4/index_7391_seed_32038_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": -0.13022041488438846, + "vlm_logprob": -0.13022041488438846, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.13022041488438846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20785294473171234 + ], + "opposite_logprob_tokens": [ + -1.7078529596328735 + ], + "logprob_reward": -0.20785294473171234, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20785294473171234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048665910959243774, + 0.5562908053398132, + 0.03773794323205948, + 0.03687065467238426, + 0.017002025619149208, + 0.030722837895154953, + 0.05843787640333176, + 0.057246703654527664, + 0.06681464612483978, + 0.09021065384149551 + ], + "credit_map": [ + 0.006048313807696104, + 0.008132219314575195, + 0.005261773243546486, + 0.0045929173938930035, + 0.0024699715431779623, + 0.0033762981183826923, + 0.010224254801869392, + 0.008560160174965858, + 0.21774335205554962, + 0.07118280977010727, + 0.027962861582636833, + 0.010352782905101776, + 0.007054351270198822, + 0.021943876519799232, + 0.019603842869400978, + 0.1804468333721161, + 0.005647511221468449, + 0.00609489856287837, + 0.0025986582040786743, + 0.005701633170247078, + 0.004340124316513538, + 0.002730489708483219, + 0.006412687245756388, + 0.004211939871311188, + 0.004530956503003836, + 0.005440342705696821, + 0.0020599605049937963, + 0.0042235576547682285, + 0.008707917295396328, + 0.0023412290029227734, + 0.0042495401576161385, + 0.00531714828684926, + 0.0018698059720918536, + 0.0019486707169562578, + 0.0026281692553311586, + 0.0017208410426974297, + 0.0024259076453745365, + 0.0018615021836012602, + 0.00174889771733433, + 0.0027982296887785196, + 0.002055902499705553, + 0.0031855383422225714, + 0.00482074124738574, + 0.0035097100771963596, + 0.003132097888737917, + 0.007131833117455244, + 0.003718941006809473, + 0.003168070688843727, + 0.005168760661035776, + 0.006355566903948784, + 0.010536301881074905, + 0.008199827745556831, + 0.008386447094380856, + 0.0067590647377073765, + 0.005302027333527803, + 0.0077298711985349655, + 0.006043297704309225, + 0.006881737615913153, + 0.010663656517863274, + 0.007654156535863876, + 0.007158095017075539, + 0.005722429603338242, + 0.0068389130756258965, + 0.006284415256232023, + 0.005261015146970749, + 0.005592952482402325, + 0.007231603842228651, + 0.00979052484035492, + 0.007056801114231348, + 0.007667765021324158, + 0.012560902163386345, + 0.011653078719973564, + 0.00971008650958538, + 0.011004197411239147, + 0.009718556888401508, + 0.012549828737974167, + 0.013665569946169853, + 0.009522118605673313, + 0.013812646269798279, + 0.010227647610008717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.20785294473171234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002499315422028303 + ], + "opposite_logprob_tokens": [ + -6.002499103546143 + ], + "logprob_reward": -0.002499315422028303, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002499315422028303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19692370295524597, + 0.25636613368988037, + 0.030695674940943718, + 0.032981276512145996, + 0.06353029608726501, + 0.031113727018237114, + 0.11502300947904587, + 0.06293956935405731, + 0.09275893867015839, + 0.11766771972179413 + ], + "credit_map": [ + 0.02122778818011284, + 0.03051094338297844, + 0.021372370421886444, + 0.021033108234405518, + 0.011210513301193714, + 0.012309523299336433, + 0.04752486199140549, + 0.031734567135572433, + 0.09477869421243668, + 0.030959399417042732, + 0.014285647310316563, + 0.007932955399155617, + 0.005161582957953215, + 0.010272336192429066, + 0.01153989601880312, + 0.08143558353185654, + 0.004529194440692663, + 0.003693928709253669, + 0.002127349376678467, + 0.0038955449126660824, + 0.004540957044810057, + 0.0021467693150043488, + 0.005928077735006809, + 0.0038338513113558292, + 0.004485603887587786, + 0.004655628930777311, + 0.0023497978691011667, + 0.004012711811810732, + 0.005736067891120911, + 0.003194526070728898, + 0.004268039483577013, + 0.004278893116861582, + 0.0068822214379906654, + 0.005887921899557114, + 0.015224743634462357, + 0.005094623193144798, + 0.009532998315989971, + 0.004800694994628429, + 0.004958722274750471, + 0.011148366145789623, + 0.0024867902975529432, + 0.0038099412340670824, + 0.004234026186168194, + 0.004212907515466213, + 0.0032087911386042833, + 0.006878061685711145, + 0.003833548165857792, + 0.002449657302349806, + 0.011172950267791748, + 0.008488224819302559, + 0.027722226455807686, + 0.018348174169659615, + 0.013513983227312565, + 0.013502806425094604, + 0.008973589166998863, + 0.01330103725194931, + 0.006965240463614464, + 0.006404405925422907, + 0.013934954069554806, + 0.008846188895404339, + 0.007453961297869682, + 0.006574607454240322, + 0.00693086301907897, + 0.005829342175275087, + 0.009452680125832558, + 0.008357490412890911, + 0.009616692550480366, + 0.011289650574326515, + 0.008021439425647259, + 0.011298362165689468, + 0.016507498919963837, + 0.018215108662843704, + 0.015389751642942429, + 0.022983478382229805, + 0.011938422918319702, + 0.016506655141711235, + 0.013486742042005062, + 0.012713524512946606, + 0.01325968187302351, + 0.011389446444809437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.002499315422028303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289189875125885 + ], + "opposite_logprob_tokens": [ + -1.0539189577102661 + ], + "logprob_reward": -0.4289189875125885, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289189875125885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06777628511190414, + 0.5277091264724731, + 0.031373828649520874, + 0.019769076257944107, + 0.04996993765234947, + 0.02935565449297428, + 0.04458263888955116, + 0.12332514673471451, + 0.03517555072903633, + 0.0709628090262413 + ], + "credit_map": [ + 0.009865222498774529, + 0.011271275579929352, + 0.011741014197468758, + 0.006951419170945883, + 0.004321068990975618, + 0.003326373640447855, + 0.010928752832114697, + 0.009371154010295868, + 0.20975734293460846, + 0.07138080894947052, + 0.02609514817595482, + 0.00917721726000309, + 0.0053063128143548965, + 0.022140199318528175, + 0.01603306643664837, + 0.16781899333000183, + 0.004374900832772255, + 0.005535319447517395, + 0.0019035391742363572, + 0.0040080430917441845, + 0.00368799758143723, + 0.0026541834231466055, + 0.005369191989302635, + 0.003840653458610177, + 0.0031795846298336983, + 0.0028924085199832916, + 0.001441977801732719, + 0.002724548103287816, + 0.003245547879487276, + 0.0014803441008552909, + 0.001779399928636849, + 0.0030252656433731318, + 0.00460871122777462, + 0.00473222928121686, + 0.005219379439949989, + 0.0036680970806628466, + 0.014218250289559364, + 0.0035047344863414764, + 0.0034451549872756004, + 0.010573381558060646, + 0.0023183971643447876, + 0.002734570764005184, + 0.005193088203668594, + 0.002595791593194008, + 0.003646646160632372, + 0.007715531159192324, + 0.002759017748758197, + 0.0023926126305013895, + 0.0048285434022545815, + 0.0043855588883161545, + 0.009824731387197971, + 0.007220542524009943, + 0.00439745606854558, + 0.005466898437589407, + 0.003253690432757139, + 0.0052052210085093975, + 0.01189206913113594, + 0.01244931761175394, + 0.027964958921074867, + 0.016672572121024132, + 0.015148891136050224, + 0.011932489462196827, + 0.014577224850654602, + 0.012687626294791698, + 0.003511192509904504, + 0.0031250191386789083, + 0.0038336466532200575, + 0.005110064521431923, + 0.0030485240276902914, + 0.003192445496097207, + 0.005909479223191738, + 0.007445179391652346, + 0.0077442885376513, + 0.01027316227555275, + 0.006977797020226717, + 0.009723113849759102, + 0.010223670862615108, + 0.008040145970880985, + 0.010036477819085121, + 0.007944153621792793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.4289189875125885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189241047948599 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.003189241047948599, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189241047948599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02798597514629364, + 0.29678237438201904, + 0.010623995214700699, + 0.010709911584854126, + 0.05333277955651283, + 0.037867821753025055, + 0.07423847168684006, + 0.020015254616737366, + 0.08862484991550446, + 0.3798186480998993 + ], + "credit_map": [ + 0.0042766788974404335, + 0.0049996948800981045, + 0.003437152598053217, + 0.0027527399361133575, + 0.0017464591655880213, + 0.002281500492244959, + 0.0039133294485509396, + 0.004578415770083666, + 0.11858271807432175, + 0.04172639548778534, + 0.014533515088260174, + 0.0049373251385986805, + 0.0031519997864961624, + 0.012271527200937271, + 0.008935172110795975, + 0.09264370054006577, + 0.0016104655805975199, + 0.0010398130398243666, + 0.0008297511958517134, + 0.0016104716341942549, + 0.0012199577176943421, + 0.0010314490646123886, + 0.0015672608278691769, + 0.0017148249316960573, + 0.0014598146080970764, + 0.0013275152305141091, + 0.0008676424622535706, + 0.0012170084519311786, + 0.0018674667226150632, + 0.0011231229873374104, + 0.0013489709235727787, + 0.001498370198532939, + 0.007669549901038408, + 0.004890847485512495, + 0.010648887604475021, + 0.004082733765244484, + 0.004807742778211832, + 0.0029664712492376566, + 0.004292707424610853, + 0.013973833061754704, + 0.002058406826108694, + 0.0028921670746058226, + 0.007977571338415146, + 0.0033182743936777115, + 0.0041154720820486546, + 0.012482967227697372, + 0.0023683784529566765, + 0.0026545817963778973, + 0.0074870167300105095, + 0.005108505021780729, + 0.01929134503006935, + 0.011727393604815006, + 0.006990635301917791, + 0.008752899244427681, + 0.006134459748864174, + 0.008746209554374218, + 0.0019778998102992773, + 0.001977504463866353, + 0.0034351691137999296, + 0.0026064026169478893, + 0.003095000283792615, + 0.002567453309893608, + 0.0022922034841030836, + 0.002063620602712035, + 0.007190515752881765, + 0.007340129930526018, + 0.008894035592675209, + 0.01261556800454855, + 0.007378656882792711, + 0.011238009668886662, + 0.016584312543272972, + 0.017383605241775513, + 0.03443310409784317, + 0.0411415696144104, + 0.04294661805033684, + 0.0494345985352993, + 0.05921485275030136, + 0.04437053203582764, + 0.057733774185180664, + 0.05054354667663574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.003189241047948599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined maglev freight train with a polished silver alloy and matte carbon fiber hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train feature glowing cyan-blue energy strips along its sides and two piercing white LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a dim, industrial-looking hangar where a massive mirror blocks the train's path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the train impacting the mirror, causing it to deform, crack, and explode into glass shards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle and does it rapidly switch between close-ups of the train's nose and glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04229786619544029, + 0.6067771911621094, + 0.03394857421517372, + 0.012127610854804516, + 0.0645095482468605, + 0.08711434155702591, + 0.019663866609334946, + 0.02800930291414261, + 0.02700752764940262, + 0.07854419201612473 + ], + "credit_map": [ + 0.004382463172078133, + 0.0068412949331104755, + 0.005243214778602123, + 0.004728926345705986, + 0.002889480907469988, + 0.0025176419876515865, + 0.009080848656594753, + 0.00661399494856596, + 0.24129021167755127, + 0.08091722428798676, + 0.03016272559762001, + 0.009527788497507572, + 0.006243609823286533, + 0.024787137284874916, + 0.018520629033446312, + 0.19532787799835205, + 0.0046036625280976295, + 0.0037072973791509867, + 0.00232923892326653, + 0.0046617123298347, + 0.004857734777033329, + 0.0027008664328604937, + 0.007343912962824106, + 0.0037441467866301537, + 0.001588064362294972, + 0.0016994682373479009, + 0.0007673638174310327, + 0.0012966408394277096, + 0.0029360384214669466, + 0.0009897238342091441, + 0.001467872061766684, + 0.0013824396301060915, + 0.005658500827848911, + 0.0074374196119606495, + 0.007410997990518808, + 0.005549211520701647, + 0.020029885694384575, + 0.0041424185037612915, + 0.003867135150358081, + 0.01041397638618946, + 0.006056990474462509, + 0.005468572489917278, + 0.01621938683092594, + 0.007775736507028341, + 0.010295640677213669, + 0.028333501890301704, + 0.006576802581548691, + 0.006387708242982626, + 0.0023558596149086952, + 0.002442056778818369, + 0.003382747294381261, + 0.002493390580639243, + 0.001784398453310132, + 0.0022749011404812336, + 0.0020725044887512922, + 0.0028580063953995705, + 0.002625565743073821, + 0.0029961811378598213, + 0.006018835119903088, + 0.003377815242856741, + 0.0035799446050077677, + 0.003398294560611248, + 0.0033626705408096313, + 0.002649996429681778, + 0.003025315934792161, + 0.0022083374205976725, + 0.0034960450138896704, + 0.0040420363657176495, + 0.0027390788309276104, + 0.004179018083959818, + 0.004068317357450724, + 0.003249378176406026, + 0.007206185720860958, + 0.00941279623657465, + 0.008166170679032803, + 0.010302760638296604, + 0.012319167144596577, + 0.009260210208594799, + 0.012052887119352818, + 0.009824017994105816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.00864158570766449 + } + ], + "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/0000125/group_5/index_tensor([5680, 5680])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_5/index_tensor([5680, 5680])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2a84b4f68bfc0f24593e8972c1210ff7d1ca52f7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_5/index_tensor([5680, 5680])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static, wide-angle shot captures a scene of apocalyptic chaos, rendered entirely in a chaotic splatter-art style. The sky is a turbulent canvas of exploding, abstracted colors, with violent splashes of fiery crimson, bruised violet, and acidic yellow colliding and bleeding into one another. From this maelstrom descends a massive tornado, not wind and dust, but a swirling vortex of thick, dark, moody paint; charcoal black and deep indigo form its funnel, while frantic, arcing white streaks hint at its terrifying speed. This massive form descends directly upon a magnificent mansion, its ornate, multi-towered structure being actively demolished. The structure itself is deconstructed into frantic, violent drips and splashes, becoming one with the storm, with entire sections of the roof rendered as shattered shards of black paint, hurled upward into the swirling vortex. The scene is a symphony of destruction, the solid form of the ornate building dissolving under the assault, creating an overwhelming atmosphere of pure, unrestrained power and expressionistic terror.", + "seed": 11050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_5/index_5680_seed_11050_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": -0.03242011964321136, + "vlm_logprob": -0.03242011964321136, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03242011964321136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0956861823797226 + ], + "opposite_logprob_tokens": [ + -2.720686197280884 + ], + "logprob_reward": -0.0956861823797226, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0956861823797226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4617535471916199, + 0.08998840302228928, + 0.03293227404356003, + 0.09231103956699371, + 0.10428246110677719, + 0.01958470791578293, + 0.06632685661315918, + 0.059157174080610275, + 0.04729637876152992, + 0.02636713534593582 + ], + "credit_map": [ + 0.07378173619508743, + 0.03313914313912392, + 0.033425841480493546, + 0.02354939468204975, + 0.060754530131816864, + 0.035848867148160934, + 0.17621809244155884, + 0.025035925209522247, + 0.01045968011021614, + 0.005663084797561169, + 0.012610474601387978, + 0.012016532942652702, + 0.011850288137793541, + 0.014113765209913254, + 0.014342843554913998, + 0.008931738324463367, + 0.0037853033281862736, + 0.0023025856353342533, + 0.00271746888756752, + 0.001757349818944931, + 0.0020587013568729162, + 0.00808322336524725, + 0.0073370300233364105, + 0.00489061139523983, + 0.010380280204117298, + 0.0053710537031292915, + 0.008543252013623714, + 0.006290869787335396, + 0.012670832686126232, + 0.00792043935507536, + 0.03276452049612999, + 0.008369790390133858, + 0.011635564267635345, + 0.0057367971166968346, + 0.014765097759664059, + 0.00944521464407444, + 0.020133985206484795, + 0.011423666030168533, + 0.024073580279946327, + 0.007068553939461708, + 0.002207418903708458, + 0.0019020690815523267, + 0.0017007766291499138, + 0.0017575150122866035, + 0.0018010197672992945, + 0.0021062742453068495, + 0.0062252553179860115, + 0.001884376979433, + 0.006715433672070503, + 0.003860906697809696, + 0.009995423257350922, + 0.0062920874916017056, + 0.009240245446562767, + 0.008626440539956093, + 0.01570487767457962, + 0.0058914353139698505, + 0.01434893999248743, + 0.004707284737378359, + 0.005800946149975061, + 0.005596914328634739, + 0.005610079970210791, + 0.00913991965353489, + 0.00857686996459961, + 0.005376218352466822, + 0.006387327332049608, + 0.0029892604798078537, + 0.004405362531542778, + 0.0033499617129564285, + 0.004896141123026609, + 0.004281992092728615, + 0.017788168042898178, + 0.0031981647480279207, + 0.00340819638222456, + 0.00158686563372612, + 0.001810479792766273, + 0.002662374172359705, + 0.004720309283584356, + 0.0024493252858519554, + 0.007938995957374573, + 0.001790589070878923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0956861823797226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108936034142971 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.01108936034142971, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108936034142971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2570425570011139, + 0.1524238884449005, + 0.05153905972838402, + 0.028674805536866188, + 0.03436177596449852, + 0.04448497295379639, + 0.3030831515789032, + 0.02966327778995037, + 0.06997781246900558, + 0.028748644515872 + ], + "credit_map": [ + 0.028634414076805115, + 0.021814433857798576, + 0.022001300007104874, + 0.017464792355895042, + 0.03126434236764908, + 0.02528691291809082, + 0.09499766677618027, + 0.0155786769464612, + 0.0168664138764143, + 0.008737809024751186, + 0.02144748345017433, + 0.014456027187407017, + 0.020504238083958626, + 0.019010009244084358, + 0.0381564125418663, + 0.013245481997728348, + 0.007462022826075554, + 0.00363025045953691, + 0.004097401164472103, + 0.004105472471565008, + 0.0030956068076193333, + 0.014063872396945953, + 0.01069923210889101, + 0.004385202191770077, + 0.00366549682803452, + 0.002060749102383852, + 0.0037067781668156385, + 0.002458377042785287, + 0.0036104246973991394, + 0.0028819164726883173, + 0.007961995899677277, + 0.002329067327082157, + 0.0032073690090328455, + 0.002313697012141347, + 0.0033737316261976957, + 0.0038471254520118237, + 0.00754909124225378, + 0.003062727628275752, + 0.00845384132117033, + 0.002554193837568164, + 0.007458742242306471, + 0.003944902215152979, + 0.0036957990378141403, + 0.002707360079512, + 0.003660717746242881, + 0.0032189064659178257, + 0.016639750450849533, + 0.0031587951816618443, + 0.031352587044239044, + 0.018318206071853638, + 0.0501180924475193, + 0.029034916311502457, + 0.03865324705839157, + 0.04079818353056908, + 0.0704122856259346, + 0.024395640939474106, + 0.00738370418548584, + 0.003146382514387369, + 0.003045025747269392, + 0.0026175545062869787, + 0.0030363118276000023, + 0.004302969668060541, + 0.0033070144709199667, + 0.0028243144042789936, + 0.012247136794030666, + 0.004506212193518877, + 0.006898434832692146, + 0.004979338031262159, + 0.006963914725929499, + 0.005604329518973827, + 0.02357768639922142, + 0.005200755782425404, + 0.003071034327149391, + 0.0014758375473320484, + 0.0018115791026502848, + 0.002331368625164032, + 0.005968017503619194, + 0.0032228378113359213, + 0.009262059815227985, + 0.0016059110639616847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.01108936034142971 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249396450817585 + ], + "opposite_logprob_tokens": [ + -6.002493858337402 + ], + "logprob_reward": -0.00249396450817585, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249396450817585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3231220841407776, + 0.3193916380405426, + 0.04774395748972893, + 0.012039244174957275, + 0.07549604773521423, + 0.07575211673974991, + 0.07070569694042206, + 0.010040123015642166, + 0.029769036918878555, + 0.035940032452344894 + ], + "credit_map": [ + 0.036951690912246704, + 0.025815702974796295, + 0.02107410319149494, + 0.01961865834891796, + 0.0412210077047348, + 0.030907470732927322, + 0.1284148097038269, + 0.019118646159768105, + 0.032373715192079544, + 0.02262972481548786, + 0.0428621806204319, + 0.0436713770031929, + 0.05251365900039673, + 0.04598920792341232, + 0.050716832280159, + 0.028634926304221153, + 0.006587722338736057, + 0.003906314726918936, + 0.003568191546946764, + 0.0033591000828891993, + 0.003957897424697876, + 0.004849100019782782, + 0.01700682006776333, + 0.004508814308792353, + 0.0011864148546010256, + 0.0012620422057807446, + 0.0013544992543756962, + 0.0013972606975585222, + 0.002384953899309039, + 0.0011784023372456431, + 0.002346171997487545, + 0.0009294985793530941, + 0.007385566830635071, + 0.005041041877120733, + 0.007486848626285791, + 0.009475411847233772, + 0.015748655423521996, + 0.005910880863666534, + 0.01776558719575405, + 0.00668205413967371, + 0.01100557204335928, + 0.007254657801240683, + 0.006374570541083813, + 0.004813068546354771, + 0.006312462035566568, + 0.005243294406682253, + 0.02949528768658638, + 0.005253202747553587, + 0.007409027311950922, + 0.004386265296489, + 0.010740249417722225, + 0.007231931202113628, + 0.009601407684385777, + 0.008068486116826534, + 0.01751774735748768, + 0.005750580225139856, + 0.0012695381883531809, + 0.000839038984850049, + 0.0008725080988369882, + 0.0013651561457663774, + 0.0014563009608536959, + 0.0012001324212178588, + 0.0019403983606025577, + 0.0010970501461997628, + 0.0055208969861269, + 0.0018994323909282684, + 0.0027508139610290527, + 0.0020281558390706778, + 0.0030243168585002422, + 0.0024260880891233683, + 0.0095016835257411, + 0.0026176515966653824, + 0.0036195567809045315, + 0.0020514517091214657, + 0.0024214384611696005, + 0.002794309053570032, + 0.007370050065219402, + 0.0039232089184224606, + 0.011739419773221016, + 0.002020599553361535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.00249396450817585 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043053578585386276 + ], + "opposite_logprob_tokens": [ + -3.16805362701416 + ], + "logprob_reward": -0.043053578585386276, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043053578585386276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15596400201320648, + 0.4894427955150604, + 0.027203606441617012, + 0.02414611726999283, + 0.046432122588157654, + 0.035574015229940414, + 0.045669011771678925, + 0.12383861094713211, + 0.04497510567307472, + 0.006754635367542505 + ], + "credit_map": [ + 0.017681842669844627, + 0.013295555487275124, + 0.010236869566142559, + 0.011310412548482418, + 0.021496538072824478, + 0.013979806564748287, + 0.059346601366996765, + 0.008616387844085693, + 0.04870995134115219, + 0.03019721992313862, + 0.08691635727882385, + 0.06174551323056221, + 0.054063331335783005, + 0.0823344811797142, + 0.08409862220287323, + 0.04137730598449707, + 0.004845936316996813, + 0.0024132293183356524, + 0.0023202341981232166, + 0.0022919343318790197, + 0.002464377786964178, + 0.0025408267974853516, + 0.008654679171741009, + 0.0016723884036764503, + 0.0030285862740129232, + 0.0016672646161168814, + 0.00196419982239604, + 0.0018012935761362314, + 0.0031988706905394793, + 0.002233817707747221, + 0.007696737535297871, + 0.0025553444866091013, + 0.004687439650297165, + 0.002575380727648735, + 0.005497828591614962, + 0.004196506459265947, + 0.009677455760538578, + 0.0043455613777041435, + 0.012387038208544254, + 0.003064909251406789, + 0.004601528402417898, + 0.003340598428621888, + 0.0030055935494601727, + 0.0025548748672008514, + 0.0029482010286301374, + 0.0028899048920720816, + 0.013010217808187008, + 0.0032230967190116644, + 0.00495641864836216, + 0.002801812021061778, + 0.007303192280232906, + 0.004381105769425631, + 0.00634462246671319, + 0.005469175521284342, + 0.01098206453025341, + 0.003430618904531002, + 0.02851029299199581, + 0.009922697208821774, + 0.012781416065990925, + 0.01415590662509203, + 0.01020473800599575, + 0.019790207967162132, + 0.01617000252008438, + 0.012303361669182777, + 0.006159238982945681, + 0.0030038401018828154, + 0.004165632650256157, + 0.0030324391555041075, + 0.004599666688591242, + 0.004115873482078314, + 0.016887549310922623, + 0.0030108680948615074, + 0.0009237462654709816, + 0.0005976330721750855, + 0.0006261445232667029, + 0.00072255585109815, + 0.0012452572118490934, + 0.0008256933651864529, + 0.0012940294109284878, + 0.0005195751436986029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.043053578585386276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777512401342392 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777512401342392, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777512401342392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06317885965108871, + 0.2969016432762146, + 0.13859325647354126, + 0.027679063379764557, + 0.042505424469709396, + 0.06863249093294144, + 0.19002510607242584, + 0.06669002026319504, + 0.09013313055038452, + 0.015660930424928665 + ], + "credit_map": [ + 0.00822998397052288, + 0.009118497371673584, + 0.004992300178855658, + 0.005409135948866606, + 0.010588129051029682, + 0.006329587660729885, + 0.012331373989582062, + 0.006179853342473507, + 0.030805325135588646, + 0.01869034953415394, + 0.054623041301965714, + 0.03381689637899399, + 0.03260495513677597, + 0.05122873932123184, + 0.04974738508462906, + 0.025384970009326935, + 0.020371243357658386, + 0.010224184952676296, + 0.009054353460669518, + 0.008891788311302662, + 0.009256318211555481, + 0.02114969678223133, + 0.043515510857105255, + 0.01613016240298748, + 0.0027801666874438524, + 0.0018158197635784745, + 0.0031591930892318487, + 0.003111079568043351, + 0.003945290111005306, + 0.002203577198088169, + 0.00889533944427967, + 0.0017685993807390332, + 0.0036309405695647, + 0.0036417830269783735, + 0.003979946952313185, + 0.006985639687627554, + 0.008482693694531918, + 0.003338787704706192, + 0.008568882942199707, + 0.0038767470978200436, + 0.009810352697968483, + 0.006170016713440418, + 0.00664830906316638, + 0.005419749766588211, + 0.0058715310879051685, + 0.005592415574938059, + 0.022714586928486824, + 0.006405530031770468, + 0.019367272034287453, + 0.013067508116364479, + 0.031151751056313515, + 0.01955001801252365, + 0.02334602363407612, + 0.024748757481575012, + 0.04352426156401634, + 0.015269526280462742, + 0.012579577974975109, + 0.004883792717009783, + 0.006425035651773214, + 0.009287316352128983, + 0.00738705275580287, + 0.008206216618418694, + 0.011182376183569431, + 0.006738647818565369, + 0.014826035127043724, + 0.0056038484908640385, + 0.008767830207943916, + 0.00563552463427186, + 0.009007097221910954, + 0.007601545657962561, + 0.0330984964966774, + 0.005592757370322943, + 0.0020671177189797163, + 0.0015059523284435272, + 0.0017597321420907974, + 0.001525824423879385, + 0.002550247823819518, + 0.0017288466915488243, + 0.003225743304938078, + 0.0012974657583981752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.009777512401342392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide-angle shot captures a scene of apocalyptic chaos, rendered entirely in a chaotic splatter-art style. The sky is a turbulent canvas of exploding, abstracted colors, with violent splashes of fiery crimson, bruised violet, and acidic yellow colliding and bleeding into one another. From this maelstrom descends a massive tornado, not wind and dust, but a swirling vortex of thick, dark, moody paint; charcoal black and deep indigo form its funnel, while frantic, arcing white streaks hint at its terrifying speed. This massive form descends directly upon a magnificent mansion, its ornate, multi-towered structure being actively demolished. The structure itself is deconstructed into frantic, violent drips and splashes, becoming one with the storm, with entire sections of the roof rendered as shattered shards of black paint, hurled upward into the swirling vortex. The scene is a symphony of destruction, the solid form of the ornate building dissolving under the assault, creating an overwhelming atmosphere of pure, unrestrained power and expressionistic terror.", + "seed": 11051, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_5/index_5680_seed_11051_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": -0.03660110330674797, + "vlm_logprob": -0.03660110330674797, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.03660110330674797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1592477262020111 + ], + "opposite_logprob_tokens": [ + -2.284247636795044 + ], + "logprob_reward": -0.1592477262020111, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1592477262020111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06127793341875076, + 0.053896814584732056, + 0.04265764355659485, + 0.05587317794561386, + 0.38747644424438477, + 0.14840248227119446, + 0.08567406982183456, + 0.03752918168902397, + 0.1041809469461441, + 0.023031311109662056 + ], + "credit_map": [ + 0.007438135799020529, + 0.009927180595695972, + 0.006703932769596577, + 0.007711614947766066, + 0.007898210547864437, + 0.006593732628971338, + 0.0062740035355091095, + 0.008731119334697723, + 0.0068999119102954865, + 0.005803050007671118, + 0.007299039512872696, + 0.00892345979809761, + 0.0065336283296346664, + 0.005016621667891741, + 0.008088021539151669, + 0.005333086475729942, + 0.005107982084155083, + 0.005515831056982279, + 0.00348458718508482, + 0.004278200678527355, + 0.007612884510308504, + 0.007035453803837299, + 0.004937177523970604, + 0.004685527645051479, + 0.007108569145202637, + 0.005258940625935793, + 0.005793499294668436, + 0.008632523939013481, + 0.0071052368730306625, + 0.008052016608417034, + 0.008066760376095772, + 0.005855628289282322, + 0.05419762060046196, + 0.0181096363812685, + 0.021211693063378334, + 0.021048130467534065, + 0.0277815293520689, + 0.13997331261634827, + 0.07474488765001297, + 0.03040963038802147, + 0.019137630239129066, + 0.013338742777705193, + 0.0156535767018795, + 0.019374124705791473, + 0.025946222245693207, + 0.01812073588371277, + 0.02168264612555504, + 0.015148811042308807, + 0.008594626560807228, + 0.008471105247735977, + 0.006187737453728914, + 0.010824698954820633, + 0.011053533293306828, + 0.01662314124405384, + 0.015506799332797527, + 0.008412420749664307, + 0.004125348292291164, + 0.0028919794131070375, + 0.005859742406755686, + 0.006593346130102873, + 0.004039195366203785, + 0.005365679040551186, + 0.004663428291678429, + 0.003990462515503168, + 0.007662171497941017, + 0.0062538255006074905, + 0.005310106091201305, + 0.0071030715480446815, + 0.011207262985408306, + 0.027653217315673828, + 0.03329981490969658, + 0.00569147989153862, + 0.0029241819866001606, + 0.0016126667615026236, + 0.0035546161234378815, + 0.0032605810556560755, + 0.003409120487049222, + 0.003260072786360979, + 0.0027339488733559847, + 0.002276123035699129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.1592477262020111 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865045003592968 + ], + "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": [ + 0.11349450796842575, + 0.07202252000570297, + 0.15144458413124084, + 0.08296040445566177, + 0.16176146268844604, + 0.15788689255714417, + 0.060813821852207184, + 0.10489273071289062, + 0.049841493368148804, + 0.044881559908390045 + ], + "credit_map": [ + 0.015441996976733208, + 0.016244694590568542, + 0.015604394488036633, + 0.013202897273004055, + 0.014036979526281357, + 0.015615453012287617, + 0.011259833350777626, + 0.012088267132639885, + 0.010270301252603531, + 0.011617445386946201, + 0.0037871275562793016, + 0.004435566253960133, + 0.013884206302464008, + 0.010438138619065285, + 0.005414610728621483, + 0.012175117619335651, + 0.023113630712032318, + 0.02062312327325344, + 0.012797740288078785, + 0.010426947847008705, + 0.022743942216038704, + 0.025802724063396454, + 0.013509264215826988, + 0.022427214309573174, + 0.0154693853110075, + 0.010077481158077717, + 0.00533325457945466, + 0.007432617712765932, + 0.011585764586925507, + 0.013108517043292522, + 0.007321394979953766, + 0.01263198908418417, + 0.019846593961119652, + 0.008131763897836208, + 0.009029112756252289, + 0.009471300058066845, + 0.013959089294075966, + 0.05043431743979454, + 0.03862103074789047, + 0.012268251739442348, + 0.02368740364909172, + 0.014898613095283508, + 0.016165567561984062, + 0.019354423508048058, + 0.02092696540057659, + 0.015505578368902206, + 0.029395416378974915, + 0.01795293763279915, + 0.007239446509629488, + 0.005691298749297857, + 0.007465853355824947, + 0.011098348535597324, + 0.007433128543198109, + 0.006503400858491659, + 0.008653179742395878, + 0.006729163695126772, + 0.011668366380035877, + 0.010577589273452759, + 0.013013583607971668, + 0.016062919050455093, + 0.016060305759310722, + 0.014100558124482632, + 0.01441569160670042, + 0.008993713185191154, + 0.0033666728995740414, + 0.003291929606348276, + 0.0023792721331119537, + 0.003285124432295561, + 0.005541385617107153, + 0.01348142884671688, + 0.015175084583461285, + 0.0033205943182110786, + 0.008477545343339443, + 0.0038779100868850946, + 0.00578169384971261, + 0.006472392473369837, + 0.004303594119846821, + 0.004404363222420216, + 0.00640576146543026, + 0.005158300511538982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "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 camera angle low, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001716570113785565 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.001716570113785565, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001716570113785565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03582688048481941, + 0.026755530387163162, + 0.05970713868737221, + 0.07356826215982437, + 0.16276001930236816, + 0.21774093806743622, + 0.03763089329004288, + 0.02657497487962246, + 0.301415354013443, + 0.058020077645778656 + ], + "credit_map": [ + 0.004665154032409191, + 0.005596205126494169, + 0.0034711428452283144, + 0.005835156887769699, + 0.004275527782738209, + 0.003979784902185202, + 0.004052089061588049, + 0.0039518135599792, + 0.003192542353644967, + 0.005249048583209515, + 0.0020491487812250853, + 0.0020145191811025143, + 0.006201631389558315, + 0.002769686747342348, + 0.0023477179929614067, + 0.002931233262643218, + 0.008347481489181519, + 0.00982477143406868, + 0.004498883616179228, + 0.0047858948819339275, + 0.009310864843428135, + 0.008719896897673607, + 0.005579589866101742, + 0.008639753796160221, + 0.00968078151345253, + 0.005177417770028114, + 0.01059018261730671, + 0.011401780880987644, + 0.009883963502943516, + 0.011270176619291306, + 0.008163990452885628, + 0.007399964611977339, + 0.01426821667701006, + 0.010592136532068253, + 0.010293404571712017, + 0.01372551266103983, + 0.015527212992310524, + 0.04425724968314171, + 0.0413658507168293, + 0.012730429880321026, + 0.023631663993000984, + 0.019865060225129128, + 0.02153077721595764, + 0.025222113355994225, + 0.03270551189780235, + 0.02293391525745392, + 0.04685705900192261, + 0.024994811043143272, + 0.004268891643732786, + 0.0036156598944216967, + 0.004041608888655901, + 0.00520943570882082, + 0.004514213185757399, + 0.004993580747395754, + 0.0070835319347679615, + 0.0039039698895066977, + 0.0023310494143515825, + 0.002859549131244421, + 0.002799150999635458, + 0.002818357665091753, + 0.005551823880523443, + 0.004638252314180136, + 0.003555743722245097, + 0.002021045656874776, + 0.021194085478782654, + 0.01770360767841339, + 0.0139911575242877, + 0.01867830567061901, + 0.03718949109315872, + 0.08888140320777893, + 0.08697406947612762, + 0.0168031994253397, + 0.006105316802859306, + 0.0035504756961017847, + 0.00630580447614193, + 0.007283393759280443, + 0.010426772758364677, + 0.013207687065005302, + 0.007061648648232222, + 0.004078975412994623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.001716570113785565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613848239183426 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.003613848239183426, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613848239183426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08080386370420456, + 0.04863420128822327, + 0.06420893967151642, + 0.13256630301475525, + 0.10777033865451813, + 0.12808796763420105, + 0.14513637125492096, + 0.08173800259828568, + 0.13175933063030243, + 0.07929464429616928 + ], + "credit_map": [ + 0.011624743230640888, + 0.013085337355732918, + 0.010874196887016296, + 0.011901063844561577, + 0.008188521489501, + 0.007487619295716286, + 0.011331849731504917, + 0.0063105300068855286, + 0.004970134235918522, + 0.007335747592151165, + 0.004947949200868607, + 0.006608283147215843, + 0.007541674189269543, + 0.005067034158855677, + 0.00665990961715579, + 0.005503468681126833, + 0.00926748476922512, + 0.01116990577429533, + 0.0051305098459124565, + 0.007795128971338272, + 0.010906517505645752, + 0.007082522381097078, + 0.006344108376652002, + 0.006512756925076246, + 0.016942868009209633, + 0.01144877914339304, + 0.015653876587748528, + 0.02354476787149906, + 0.02070966176688671, + 0.012805960141122341, + 0.021143624559044838, + 0.010316770523786545, + 0.010686706751585007, + 0.007963229902088642, + 0.011948885396122932, + 0.01291984599083662, + 0.013793673366308212, + 0.0167580246925354, + 0.020222555845975876, + 0.013477416709065437, + 0.0120689757168293, + 0.014411154203116894, + 0.015193664468824863, + 0.01808917708694935, + 0.016791174188256264, + 0.01832585036754608, + 0.017366740852594376, + 0.01584124006330967, + 0.01530251931399107, + 0.012161897495388985, + 0.0194239504635334, + 0.019932415336370468, + 0.01729539968073368, + 0.018727798014879227, + 0.023803099989891052, + 0.0184892937541008, + 0.009759039618074894, + 0.008260528557002544, + 0.009365200996398926, + 0.009955968707799911, + 0.01418393850326538, + 0.0076674483716487885, + 0.013198134489357471, + 0.00934773962944746, + 0.01024690642952919, + 0.006558861583471298, + 0.00613653426989913, + 0.007648846600204706, + 0.013035167008638382, + 0.04113788530230522, + 0.040430426597595215, + 0.006564691662788391, + 0.006265255156904459, + 0.009534474462270737, + 0.009055767208337784, + 0.011102140881121159, + 0.012381313368678093, + 0.00934414379298687, + 0.013487706892192364, + 0.008123837411403656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.003613848239183426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776921942830086 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009776921942830086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776921942830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08560144156217575, + 0.06735344231128693, + 0.0950632393360138, + 0.07058539241552353, + 0.3541710078716278, + 0.09127961844205856, + 0.08240696042776108, + 0.04462810605764389, + 0.03973715007305145, + 0.06917363405227661 + ], + "credit_map": [ + 0.009781040251255035, + 0.013756719417870045, + 0.008105807937681675, + 0.011467834934592247, + 0.013644496910274029, + 0.008695060387253761, + 0.009788054041564465, + 0.010362420231103897, + 0.00958835519850254, + 0.010182065889239311, + 0.00530181685462594, + 0.006837401073426008, + 0.015848062932491302, + 0.006540702655911446, + 0.006398023571819067, + 0.0066570015624165535, + 0.013415028341114521, + 0.017682118341326714, + 0.011057300493121147, + 0.010086894035339355, + 0.013057864271104336, + 0.011430054903030396, + 0.008636726066470146, + 0.009697242639958858, + 0.008709857240319252, + 0.006797214969992638, + 0.007416058797389269, + 0.009412089362740517, + 0.009118790738284588, + 0.011659013107419014, + 0.007700612302869558, + 0.009771750308573246, + 0.03830850496888161, + 0.016393670812249184, + 0.0192001610994339, + 0.022727346047759056, + 0.030473588034510612, + 0.11064091324806213, + 0.0904335156083107, + 0.025993280112743378, + 0.008203030563890934, + 0.01000706572085619, + 0.008798125199973583, + 0.012126876041293144, + 0.01585906744003296, + 0.01342032104730606, + 0.014488961547613144, + 0.008376163430511951, + 0.00927501730620861, + 0.011538510210812092, + 0.00801028497517109, + 0.015160969458520412, + 0.009650660678744316, + 0.008532599546015263, + 0.012807064689695835, + 0.007431849371641874, + 0.005064430180937052, + 0.004628676455467939, + 0.005208675283938646, + 0.007013986352831125, + 0.006211596075445414, + 0.00519462488591671, + 0.0065020364709198475, + 0.004804076161235571, + 0.004337796941399574, + 0.0038619874976575375, + 0.003496448276564479, + 0.004905708599835634, + 0.006425291765481234, + 0.006926043424755335, + 0.006828374229371548, + 0.0029554942157119513, + 0.008342280983924866, + 0.0067345015704631805, + 0.012906570918858051, + 0.011459420435130596, + 0.007566716056317091, + 0.007545277010649443, + 0.00749586895108223, + 0.007122990675270557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.009776921942830086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide-angle shot captures a scene of apocalyptic chaos, rendered entirely in a chaotic splatter-art style. The sky is a turbulent canvas of exploding, abstracted colors, with violent splashes of fiery crimson, bruised violet, and acidic yellow colliding and bleeding into one another. From this maelstrom descends a massive tornado, not wind and dust, but a swirling vortex of thick, dark, moody paint; charcoal black and deep indigo form its funnel, while frantic, arcing white streaks hint at its terrifying speed. This massive form descends directly upon a magnificent mansion, its ornate, multi-towered structure being actively demolished. The structure itself is deconstructed into frantic, violent drips and splashes, becoming one with the storm, with entire sections of the roof rendered as shattered shards of black paint, hurled upward into the swirling vortex. The scene is a symphony of destruction, the solid form of the ornate building dissolving under the assault, creating an overwhelming atmosphere of pure, unrestrained power and expressionistic terror.", + "seed": 11052, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_5/index_5680_seed_11052_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": -0.03277038717642426, + "vlm_logprob": -0.03277038717642426, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.03277038717642426, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13278378546237946 + ], + "opposite_logprob_tokens": [ + -2.382783889770508 + ], + "logprob_reward": -0.13278378546237946, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.13278378546237946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17219381034374237, + 0.054934170097112656, + 0.05199054628610611, + 0.14039385318756104, + 0.07314546406269073, + 0.27915218472480774, + 0.024763986468315125, + 0.025572946295142174, + 0.11298513412475586, + 0.06486788392066956 + ], + "credit_map": [ + 0.01515343226492405, + 0.023658664897084236, + 0.022097526118159294, + 0.03826535865664482, + 0.01455336157232523, + 0.016894446685910225, + 0.024097934365272522, + 0.017473097890615463, + 0.0062407879158854485, + 0.007789310533553362, + 0.0055668954737484455, + 0.0066254641860723495, + 0.007482974790036678, + 0.00816831924021244, + 0.007758118212223053, + 0.005302301608026028, + 0.0058734118938446045, + 0.0071954019367694855, + 0.00497062224894762, + 0.008081787265837193, + 0.005629031918942928, + 0.005215069279074669, + 0.009887535125017166, + 0.005137683358043432, + 0.019893700256943703, + 0.015220733359456062, + 0.008369065821170807, + 0.009856187738478184, + 0.017289167270064354, + 0.02143283747136593, + 0.013181736692786217, + 0.035150427371263504, + 0.01369835902005434, + 0.006849910132586956, + 0.006152628920972347, + 0.004449475556612015, + 0.010552356950938702, + 0.01771382987499237, + 0.004831342026591301, + 0.008897554129362106, + 0.058615341782569885, + 0.03064933605492115, + 0.017847510054707527, + 0.022943101823329926, + 0.07429039478302002, + 0.028205016627907753, + 0.025750696659088135, + 0.020850805565714836, + 0.004840562120079994, + 0.002359305042773485, + 0.0023513201158493757, + 0.003283935831859708, + 0.0019860665779560804, + 0.005337362177670002, + 0.0020504591520875692, + 0.0025549745187163353, + 0.004845021292567253, + 0.004131362773478031, + 0.002094636671245098, + 0.002723120152950287, + 0.0028984572272747755, + 0.002858430380001664, + 0.003280902747064829, + 0.0027410152833908796, + 0.011380903422832489, + 0.008335225284099579, + 0.009047364816069603, + 0.008649643510580063, + 0.008612969890236855, + 0.031528640538454056, + 0.010276561602950096, + 0.02515382505953312, + 0.014721925370395184, + 0.004820705857127905, + 0.003745119087398052, + 0.004262806847691536, + 0.005984713789075613, + 0.02022060751914978, + 0.005115277599543333, + 0.005996727850288153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.13278378546237946 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004653813783079386 + ], + "opposite_logprob_tokens": [ + -5.3796539306640625 + ], + "logprob_reward": -0.004653813783079386, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004653813783079386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36918145418167114, + 0.1098974421620369, + 0.07078001648187637, + 0.04121684655547142, + 0.032610468566417694, + 0.21574412286281586, + 0.01696149632334709, + 0.056249476969242096, + 0.05469977855682373, + 0.032658860087394714 + ], + "credit_map": [ + 0.03648323938250542, + 0.06711093336343765, + 0.03489113971590996, + 0.06202612444758415, + 0.033396363258361816, + 0.0397920235991478, + 0.053139809519052505, + 0.042341794818639755, + 0.011657586321234703, + 0.009004834108054638, + 0.005651353858411312, + 0.009605370461940765, + 0.010507685132324696, + 0.028256572782993317, + 0.015102444216609001, + 0.02011159062385559, + 0.005811737384647131, + 0.010256242007017136, + 0.004087388981133699, + 0.006042610388249159, + 0.01227105874568224, + 0.00921064242720604, + 0.015753425657749176, + 0.0073469169437885284, + 0.009927568025887012, + 0.006194557994604111, + 0.0031353605445474386, + 0.002396167255938053, + 0.0046149203553795815, + 0.006524278782308102, + 0.0031701477710157633, + 0.005253845825791359, + 0.003031605388969183, + 0.004312378354370594, + 0.002639166312292218, + 0.004840466193854809, + 0.005517725367099047, + 0.003731237258762121, + 0.004577229265123606, + 0.003960662055760622, + 0.03051692061126232, + 0.021617181599140167, + 0.010559015907347202, + 0.009654548950493336, + 0.05688142031431198, + 0.05838935822248459, + 0.013581372797489166, + 0.014544323086738586, + 0.0023466700222343206, + 0.0028127385303378105, + 0.0014588653575628996, + 0.002376857213675976, + 0.0021393741481006145, + 0.0027029109187424183, + 0.0015088963555172086, + 0.0016151838935911655, + 0.010192983783781528, + 0.005047403275966644, + 0.003757024882361293, + 0.004259165376424789, + 0.006894426885992289, + 0.00838731974363327, + 0.0053220693953335285, + 0.012389082461595535, + 0.006612783297896385, + 0.006519953720271587, + 0.00794776901602745, + 0.006342111621052027, + 0.008073336444795132, + 0.005156512372195721, + 0.008623999543488026, + 0.005423309281468391, + 0.00411309115588665, + 0.001927178236655891, + 0.0026124147698283195, + 0.0021940958686172962, + 0.0038604973815381527, + 0.006213626358658075, + 0.003149415133520961, + 0.008588540367782116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004653813783079386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619549795985222 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619549795985222, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619549795985222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23055019974708557, + 0.092307448387146, + 0.14200358092784882, + 0.047324538230895996, + 0.09622988104820251, + 0.15577425062656403, + 0.041596412658691406, + 0.07873808592557907, + 0.044816821813583374, + 0.07065879553556442 + ], + "credit_map": [ + 0.026341458782553673, + 0.04936946555972099, + 0.01862245425581932, + 0.0398203544318676, + 0.02680222876369953, + 0.02271704189479351, + 0.01903645135462284, + 0.027840744704008102, + 0.014175748452544212, + 0.010809121653437614, + 0.0070960610173642635, + 0.010167104192078114, + 0.006833573337644339, + 0.012911488302052021, + 0.015509570017457008, + 0.014804774895310402, + 0.014493734575808048, + 0.022402523085474968, + 0.007439689245074987, + 0.022196708247065544, + 0.025974296033382416, + 0.01736285910010338, + 0.019025634974241257, + 0.013108138926327229, + 0.008449528366327286, + 0.004454033449292183, + 0.00317657133564353, + 0.005427646916359663, + 0.0057531241327524185, + 0.008588574826717377, + 0.0060514225624501705, + 0.005423636641353369, + 0.015767065808176994, + 0.00794963352382183, + 0.008593144826591015, + 0.00979984924197197, + 0.011819513514637947, + 0.01802757941186428, + 0.01123120915144682, + 0.013041887432336807, + 0.03654114156961441, + 0.018614836037158966, + 0.01275140605866909, + 0.008055662736296654, + 0.01407889649271965, + 0.04397059231996536, + 0.008852824568748474, + 0.01290888525545597, + 0.007873520255088806, + 0.006459670141339302, + 0.0032339810859411955, + 0.0030521985609084368, + 0.004543592222034931, + 0.006201461888849735, + 0.004213337786495686, + 0.006018652580678463, + 0.015035397373139858, + 0.008366122841835022, + 0.004944469779729843, + 0.004930503200739622, + 0.010471721179783344, + 0.01913299970328808, + 0.00599769689142704, + 0.009859173558652401, + 0.004308100789785385, + 0.0041345106437802315, + 0.003539459779858589, + 0.0035371927078813314, + 0.004102405626326799, + 0.01202011201530695, + 0.0037826024927198887, + 0.009392432868480682, + 0.015077576972544193, + 0.00634989608079195, + 0.005561611615121365, + 0.004801702219992876, + 0.007532574702054262, + 0.01848411001265049, + 0.0056368280202150345, + 0.0072144982405006886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.003619549795985222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606214977800846 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.01606214977800846, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606214977800846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22315792739391327, + 0.2356032133102417, + 0.06151317432522774, + 0.05634729564189911, + 0.05401758477091789, + 0.11835654824972153, + 0.013046144507825375, + 0.03024178184568882, + 0.17321887612342834, + 0.03449750691652298 + ], + "credit_map": [ + 0.01875191368162632, + 0.027518192306160927, + 0.016457635909318924, + 0.0377802848815918, + 0.02016075700521469, + 0.02519528940320015, + 0.05337507650256157, + 0.023918762803077698, + 0.019467703998088837, + 0.02005327306687832, + 0.013596244156360626, + 0.026811040937900543, + 0.02433447539806366, + 0.03267328441143036, + 0.06167607754468918, + 0.03699107840657234, + 0.004071560688316822, + 0.007642569951713085, + 0.003984896000474691, + 0.006620864849537611, + 0.009857098571956158, + 0.00758634926751256, + 0.014967497438192368, + 0.006782333832234144, + 0.012055455707013607, + 0.006993836723268032, + 0.0031027377117425203, + 0.004885220900177956, + 0.006598313804715872, + 0.012455534189939499, + 0.005817118566483259, + 0.004439071286469698, + 0.009841825813055038, + 0.0051888637244701385, + 0.005539622623473406, + 0.006497832015156746, + 0.005255859810858965, + 0.010620813816785812, + 0.004885077942162752, + 0.006187679711729288, + 0.01714717596769333, + 0.011487447656691074, + 0.006904500536620617, + 0.008830036036670208, + 0.023751407861709595, + 0.03504164516925812, + 0.006638872437179089, + 0.008555454201996326, + 0.0020373843144625425, + 0.002230183221399784, + 0.0008913573110476136, + 0.0014983514556661248, + 0.001411861157976091, + 0.001858565490692854, + 0.0015865699388086796, + 0.001531871734187007, + 0.003670474048703909, + 0.0028143408708274364, + 0.0017306908266618848, + 0.003591648070141673, + 0.0037984673399478197, + 0.008143852464854717, + 0.0031889518722891808, + 0.003303355537354946, + 0.02027519978582859, + 0.015274827368557453, + 0.00888771004974842, + 0.008429685607552528, + 0.01378668937832117, + 0.049072787165641785, + 0.014007379300892353, + 0.043484579771757126, + 0.0037359504494816065, + 0.005068359896540642, + 0.0023222838062793016, + 0.00436669634655118, + 0.00397723400965333, + 0.006399150006473064, + 0.00279436307027936, + 0.005833467002958059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.01606214977800846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067326370626688 + ], + "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": [ + 0.23785611987113953, + 0.10078807920217514, + 0.0743379294872284, + 0.08936528861522675, + 0.061258941888809204, + 0.20452608168125153, + 0.034302063286304474, + 0.06510110944509506, + 0.06682459264993668, + 0.06563985347747803 + ], + "credit_map": [ + 0.024162260815501213, + 0.03834395110607147, + 0.016343673691153526, + 0.027935802936553955, + 0.02493884786963463, + 0.02715243212878704, + 0.04803123325109482, + 0.03094789944589138, + 0.010530997067689896, + 0.008896183222532272, + 0.010108334943652153, + 0.008909872733056545, + 0.009448390454053879, + 0.025087909772992134, + 0.010147962719202042, + 0.017658432945609093, + 0.006473691202700138, + 0.01132059283554554, + 0.0050755999982357025, + 0.006642836146056652, + 0.014076362363994122, + 0.012149101123213768, + 0.010365627706050873, + 0.008234119042754173, + 0.009594307281076908, + 0.00537687074393034, + 0.005078472197055817, + 0.00593377323821187, + 0.009087612852454185, + 0.01965305581688881, + 0.009617073461413383, + 0.025024116039276123, + 0.01217588596045971, + 0.00667743431404233, + 0.006550131365656853, + 0.004241697955876589, + 0.007781279273331165, + 0.011654396541416645, + 0.006156824063509703, + 0.006021293345838785, + 0.01907009445130825, + 0.019034767523407936, + 0.007073662243783474, + 0.012687238864600658, + 0.07450447231531143, + 0.03851521387696266, + 0.020262280479073524, + 0.013378355652093887, + 0.006289690732955933, + 0.005896155256778002, + 0.0027870095800608397, + 0.00528709078207612, + 0.0028529444243758917, + 0.004856469109654427, + 0.003393733175471425, + 0.002938970923423767, + 0.0104684978723526, + 0.00735572911798954, + 0.004166266415268183, + 0.006218410097062588, + 0.008580300956964493, + 0.017763258889317513, + 0.005738037638366222, + 0.004810609854757786, + 0.008879515342414379, + 0.00984492339193821, + 0.008906357921659946, + 0.007773109711706638, + 0.007060118485242128, + 0.00920927058905363, + 0.006045053713023663, + 0.009106244891881943, + 0.007688197772949934, + 0.0111696170642972, + 0.004667432047426701, + 0.009779395535588264, + 0.007789234630763531, + 0.009555884636938572, + 0.008094881661236286, + 0.0068952059373259544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0067326370626688 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide-angle shot captures a scene of apocalyptic chaos, rendered entirely in a chaotic splatter-art style. The sky is a turbulent canvas of exploding, abstracted colors, with violent splashes of fiery crimson, bruised violet, and acidic yellow colliding and bleeding into one another. From this maelstrom descends a massive tornado, not wind and dust, but a swirling vortex of thick, dark, moody paint; charcoal black and deep indigo form its funnel, while frantic, arcing white streaks hint at its terrifying speed. This massive form descends directly upon a magnificent mansion, its ornate, multi-towered structure being actively demolished. The structure itself is deconstructed into frantic, violent drips and splashes, becoming one with the storm, with entire sections of the roof rendered as shattered shards of black paint, hurled upward into the swirling vortex. The scene is a symphony of destruction, the solid form of the ornate building dissolving under the assault, creating an overwhelming atmosphere of pure, unrestrained power and expressionistic terror.", + "seed": 11053, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_5/index_5680_seed_11053_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": -0.05741104008629918, + "vlm_logprob": -0.05741104008629918, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.05741104008629918, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20529629290103912 + ], + "opposite_logprob_tokens": [ + -1.955296277999878 + ], + "logprob_reward": -0.20529629290103912, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.20529629290103912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08449800312519073, + 0.04612959548830986, + 0.07748259603977203, + 0.08125106245279312, + 0.166045680642128, + 0.14093011617660522, + 0.058251798152923584, + 0.18084540963172913, + 0.12802517414093018, + 0.036540623754262924 + ], + "credit_map": [ + 0.010808870196342468, + 0.010414025746285915, + 0.012926925905048847, + 0.008834068663418293, + 0.0096140680834651, + 0.010168835520744324, + 0.011015089228749275, + 0.010716108605265617, + 0.005740505643188953, + 0.005119329784065485, + 0.00607274379581213, + 0.0065424744971096516, + 0.004935580305755138, + 0.004280766472220421, + 0.008845079690217972, + 0.00459311343729496, + 0.01085951179265976, + 0.007526774425059557, + 0.011266988702118397, + 0.008882858790457249, + 0.006841067224740982, + 0.006333968136459589, + 0.01762930490076542, + 0.008142111822962761, + 0.010837775655090809, + 0.02979428507387638, + 0.007698648609220982, + 0.003961603622883558, + 0.010193565860390663, + 0.005778396036475897, + 0.005631604231894016, + 0.00735517218708992, + 0.01961640827357769, + 0.037519801408052444, + 0.024048568680882454, + 0.02551756240427494, + 0.011937310919165611, + 0.02048431523144245, + 0.014431274496018887, + 0.012490419670939445, + 0.019989892840385437, + 0.04733426868915558, + 0.01778271235525608, + 0.011089125648140907, + 0.01172367949038744, + 0.013312316499650478, + 0.009726335294544697, + 0.00997177604585886, + 0.007001830730587244, + 0.012325538322329521, + 0.007613230496644974, + 0.010369274765253067, + 0.00677457544952631, + 0.003462801221758127, + 0.004966145846992731, + 0.005738395731896162, + 0.014580460265278816, + 0.06565108895301819, + 0.010728356428444386, + 0.014498955570161343, + 0.04740244150161743, + 0.009935467503964901, + 0.009058934636414051, + 0.008989700116217136, + 0.012121426872909069, + 0.02186579257249832, + 0.010054565966129303, + 0.01060309074819088, + 0.01642153412103653, + 0.012999698519706726, + 0.024599218741059303, + 0.019359836354851723, + 0.005423367489129305, + 0.004315828904509544, + 0.003680257825180888, + 0.006045693531632423, + 0.0049523282796144485, + 0.003663992742076516, + 0.00352418958209455, + 0.004934962838888168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.20529629290103912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205030158162117 + ], + "opposite_logprob_tokens": [ + -2.8120503425598145 + ], + "logprob_reward": -0.06205030158162117, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205030158162117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09983683377504349, + 0.20723696053028107, + 0.12680009007453918, + 0.1315852850675583, + 0.04562116786837578, + 0.04547784850001335, + 0.05136776342988014, + 0.1756519228219986, + 0.048851128667593, + 0.06757105886936188 + ], + "credit_map": [ + 0.012726730667054653, + 0.012413998134434223, + 0.01251315325498581, + 0.009389289654791355, + 0.010458996519446373, + 0.010868997313082218, + 0.010157030075788498, + 0.02130863070487976, + 0.03648276999592781, + 0.02635512873530388, + 0.0393078438937664, + 0.01802539825439453, + 0.023230360820889473, + 0.01570981554687023, + 0.013490989804267883, + 0.034634627401828766, + 0.016908060759305954, + 0.013845794834196568, + 0.012660340406000614, + 0.01576702669262886, + 0.011120772920548916, + 0.011937189847230911, + 0.031109262257814407, + 0.013451640494167805, + 0.016627870500087738, + 0.05199287831783295, + 0.010077457875013351, + 0.006677334662526846, + 0.018538102507591248, + 0.006650599185377359, + 0.008936837315559387, + 0.012084197252988815, + 0.005221724044531584, + 0.010312371887266636, + 0.004684808198362589, + 0.008312913589179516, + 0.003830833127722144, + 0.0057415650226175785, + 0.004379337187856436, + 0.003137611784040928, + 0.00609720591455698, + 0.010098559781908989, + 0.007308986037969589, + 0.0038919267244637012, + 0.0036895121447741985, + 0.004729162435978651, + 0.0046713389456272125, + 0.004991155117750168, + 0.005794613156467676, + 0.01381785236299038, + 0.004720701370388269, + 0.004389489535242319, + 0.006643042434006929, + 0.006662752479314804, + 0.0043602692894637585, + 0.004979039542376995, + 0.01566123776137829, + 0.05819288268685341, + 0.015228885225951672, + 0.02071268856525421, + 0.03947610780596733, + 0.009473822079598904, + 0.009585599415004253, + 0.007320695091038942, + 0.005168646574020386, + 0.006637849844992161, + 0.0038780388422310352, + 0.003742689499631524, + 0.005721450783312321, + 0.006029111798852682, + 0.009378474205732346, + 0.008294865489006042, + 0.007090585771948099, + 0.0079584289342165, + 0.005163971800357103, + 0.006634474731981754, + 0.014928380027413368, + 0.007320474833250046, + 0.009121542796492577, + 0.009353199042379856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.06205030158162117 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785539470613003 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785539470613003, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785539470613003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1516566127538681, + 0.09501953423023224, + 0.20541614294052124, + 0.1719093918800354, + 0.09224142134189606, + 0.03401299566030502, + 0.027927180752158165, + 0.06244596093893051, + 0.05689325928688049, + 0.1024775356054306 + ], + "credit_map": [ + 0.016564683988690376, + 0.023174310103058815, + 0.016028571873903275, + 0.015625104308128357, + 0.01616549864411354, + 0.020698444917798042, + 0.016778450459241867, + 0.026621561497449875, + 0.016055213287472725, + 0.011677135713398457, + 0.018461260944604874, + 0.00901398342102766, + 0.009599223732948303, + 0.007160527165979147, + 0.009693841449916363, + 0.013358351774513721, + 0.023070188239216805, + 0.020414382219314575, + 0.018286114558577538, + 0.025089485570788383, + 0.02558140829205513, + 0.019087009131908417, + 0.05232793092727661, + 0.021559620276093483, + 0.02212536334991455, + 0.07043982297182083, + 0.010305766947567463, + 0.008826802484691143, + 0.025948548689484596, + 0.008732521906495094, + 0.011714184656739235, + 0.01381638366729021, + 0.01245090365409851, + 0.023080801591277122, + 0.015063718892633915, + 0.011548085138201714, + 0.007871617563068867, + 0.007983522489666939, + 0.00697333412244916, + 0.007269440218806267, + 0.004177023656666279, + 0.008213869296014309, + 0.0029278313741087914, + 0.0046825832687318325, + 0.0032593647483736277, + 0.0037344694137573242, + 0.002816656604409218, + 0.0042011975310742855, + 0.00301846070215106, + 0.005108342505991459, + 0.0020637756679207087, + 0.0023702112957835197, + 0.005740483291447163, + 0.0028346895705908537, + 0.0033337159547954798, + 0.003457501530647278, + 0.004865436814725399, + 0.022270886227488518, + 0.003798650112003088, + 0.003297504736110568, + 0.016225505620241165, + 0.003641269402578473, + 0.004390789195895195, + 0.0039559160359203815, + 0.005777025129646063, + 0.009974513202905655, + 0.004166640341281891, + 0.005124952644109726, + 0.007425019517540932, + 0.006461320444941521, + 0.009794224053621292, + 0.00816956628113985, + 0.008803660981357098, + 0.012505973689258099, + 0.010264674201607704, + 0.009520390070974827, + 0.020082537084817886, + 0.01111507322639227, + 0.01385557558387518, + 0.016329647973179817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.009785539470613003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190429415553808 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190429415553808, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190429415553808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20824509859085083, + 0.06971324235200882, + 0.052416153252124786, + 0.03153500333428383, + 0.06327042728662491, + 0.06867840141057968, + 0.026125455275177956, + 0.22043471038341522, + 0.15597687661647797, + 0.10360468178987503 + ], + "credit_map": [ + 0.03061281144618988, + 0.026752453297376633, + 0.02974710240960121, + 0.020926201716065407, + 0.01988842524588108, + 0.030239244922995567, + 0.023881368339061737, + 0.02619747444987297, + 0.01012021116912365, + 0.007178280968219042, + 0.010588864795863628, + 0.008817333728075027, + 0.007347292732447386, + 0.006886228919029236, + 0.009525676257908344, + 0.009249347262084484, + 0.006572301499545574, + 0.008409446105360985, + 0.004708149470388889, + 0.005320579744875431, + 0.007140203379094601, + 0.004819898400455713, + 0.0067967623472213745, + 0.008648809045553207, + 0.003811151022091508, + 0.009840457700192928, + 0.0021748156286776066, + 0.0020168612245470285, + 0.0031293563079088926, + 0.0019154351903125644, + 0.005082137882709503, + 0.0035647847689688206, + 0.00763955432921648, + 0.01840178668498993, + 0.006448270287364721, + 0.008320805616676807, + 0.005849003791809082, + 0.007060124073177576, + 0.00476789241656661, + 0.0047829835675656796, + 0.009635979309678078, + 0.021374471485614777, + 0.007114861626178026, + 0.005165985319763422, + 0.005593699868768454, + 0.00581328384578228, + 0.006018552463501692, + 0.007961565628647804, + 0.002717952709645033, + 0.004935301374644041, + 0.0018294601468369365, + 0.0024874326772987843, + 0.004072512034326792, + 0.002354136435315013, + 0.003436221508309245, + 0.0042924354784190655, + 0.018799610435962677, + 0.07184284180402756, + 0.012884258292615414, + 0.025868430733680725, + 0.05059037730097771, + 0.014952125959098339, + 0.01216593012213707, + 0.013331123627722263, + 0.012230626307427883, + 0.025444839149713516, + 0.00907871313393116, + 0.015213493257761002, + 0.018383435904979706, + 0.014479659497737885, + 0.03369099646806717, + 0.027455106377601624, + 0.010213270783424377, + 0.01264001801609993, + 0.01103285700082779, + 0.010167942382395267, + 0.018803197890520096, + 0.010352879762649536, + 0.011040430516004562, + 0.01935407519340515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.003190429415553808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, static, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a chaotic splatter-art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tornado composed of thick dark paint (charcoal black and deep indigo) descending from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ornate multi-towered mansion visible being demolished by the tornado?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sections of the mansion's roof rendered as shattered shards of black paint hurled upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067326370626688 + ], + "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": [ + 0.053825244307518005, + 0.043957509100437164, + 0.05900414288043976, + 0.10248208791017532, + 0.0847071036696434, + 0.07321471720933914, + 0.07883908599615097, + 0.1749749332666397, + 0.08285169303417206, + 0.24614344537258148 + ], + "credit_map": [ + 0.008708326146006584, + 0.007337547373026609, + 0.006218988914042711, + 0.006204706616699696, + 0.005288512445986271, + 0.006582539062947035, + 0.006976066157221794, + 0.006508556194603443, + 0.006059390027076006, + 0.005518627353012562, + 0.005148559343069792, + 0.005625029560178518, + 0.005124971736222506, + 0.004218637943267822, + 0.0073730540461838245, + 0.004889240022748709, + 0.01032000221312046, + 0.005425501149147749, + 0.013635347597301006, + 0.005044476129114628, + 0.0058019827120006084, + 0.005042861681431532, + 0.004639224614948034, + 0.009094745852053165, + 0.012918327935039997, + 0.0393710620701313, + 0.008708409033715725, + 0.005273144226521254, + 0.01411348395049572, + 0.006436028052121401, + 0.00806013960391283, + 0.007601488847285509, + 0.011656424961984158, + 0.022342737764120102, + 0.014163519255816936, + 0.010056545957922935, + 0.005733477883040905, + 0.00861269049346447, + 0.006112557370215654, + 0.006029150448739529, + 0.009240514598786831, + 0.016036629676818848, + 0.00622699735686183, + 0.00549498712643981, + 0.0063640046864748, + 0.007689469028264284, + 0.00798072200268507, + 0.014181389473378658, + 0.00967958103865385, + 0.013860808685421944, + 0.007804911583662033, + 0.006819609552621841, + 0.015327022410929203, + 0.006421756464987993, + 0.011290469206869602, + 0.0076349289156496525, + 0.015321094542741776, + 0.05937438830733299, + 0.00914474856108427, + 0.016876693814992905, + 0.04290884733200073, + 0.01209588348865509, + 0.009430375881493092, + 0.009822892025113106, + 0.00800074078142643, + 0.015171848237514496, + 0.00583282858133316, + 0.010204444639384747, + 0.009724311530590057, + 0.008216842077672482, + 0.013702275231480598, + 0.011998401954770088, + 0.021722618490457535, + 0.028387337923049927, + 0.020074080675840378, + 0.021901484578847885, + 0.054150331765413284, + 0.024954646825790405, + 0.03412974625825882, + 0.04082319885492325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.0067326370626688 + } + ], + "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/0000125/group_7/index_tensor([1262, 1262])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_tensor([1262, 1262])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7aaeebc70dc54c8011670737e999f5ebe16ebcf4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_tensor([1262, 1262])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static shot, shot vertically from above, captures a classical woman in white, with a soft-focus, vintage film quality. Her long, lustrous black hair is meticulously tied into an elegant bun, secured with a simple wooden hairpin. A few strands brush against her cheeks as she moves. She is nestled deep within a tranquil, dense bamboo forest, her feet blanketed in a thick carpet of dry, yellowing bamboo leaves. She wears a flowing white, wide-sleeved kilt, the light silk material gleaming softly in the dappled light. She twirls effortlessly in place, her steps light and effortless, her toes tapping the fallen leaves almost silently. Her wide white sleeves unfurl as she spins, drawing a perfect circle on the ground, like a slowly blooming white lotus. Dappled sunlight filters through the lush green of the bamboo leaves overhead, casting ever-changing specks of light across her pure white dress and the leaf-strewn ground, intertwining with the shadows of the bamboo forest. The painting, set against a backdrop of emerald green and withered yellow, highlights the pure white of the central figure, creating a tranquil, ethereal, and poetic atmosphere of classical aesthetics.", + "seed": 33350, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_1262_seed_33350_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": -0.09490619320422411, + "vlm_logprob": -0.09490619320422411, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.09490619320422411, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.36250388622283936 + ], + "opposite_logprob_tokens": [ + -1.2375038862228394 + ], + "logprob_reward": -0.36250388622283936, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36250388622283936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26103875041007996, + 0.08911184966564178, + 0.054263100028038025, + 0.05616799369454384, + 0.03497679904103279, + 0.020505407825112343, + 0.04107058420777321, + 0.07500650733709335, + 0.03630664572119713, + 0.3315523564815521 + ], + "credit_map": [ + 0.017870640382170677, + 0.016453487798571587, + 0.08004444092512131, + 0.020669417455792427, + 0.02065269462764263, + 0.05682837963104248, + 0.017571810632944107, + 0.030947908759117126, + 0.01484548207372427, + 0.006197582930326462, + 0.026655664667487144, + 0.006312314886599779, + 0.006733206566423178, + 0.008814987726509571, + 0.012812984175980091, + 0.006739629432559013, + 0.005907412152737379, + 0.011376498267054558, + 0.007475980557501316, + 0.006687173154205084, + 0.006418106611818075, + 0.007021530997008085, + 0.004825138952583075, + 0.004551256075501442, + 0.01131659559905529, + 0.00364772230386734, + 0.01100078783929348, + 0.0036745050456374884, + 0.005900162272155285, + 0.007178827188909054, + 0.007017537020146847, + 0.006431856658309698, + 0.00413939543068409, + 0.0037894854322075844, + 0.0032924767583608627, + 0.003103654133155942, + 0.008260940201580524, + 0.005029776133596897, + 0.004148037638515234, + 0.003213034477084875, + 0.004190865904092789, + 0.0014747560489922762, + 0.0027369000017642975, + 0.001507401466369629, + 0.0011555235832929611, + 0.0021522038150578737, + 0.0013705063611268997, + 0.005917251575738192, + 0.006328875664621592, + 0.0024846794549375772, + 0.006600892171263695, + 0.0030245075467973948, + 0.00314714596606791, + 0.008234811015427113, + 0.004100820980966091, + 0.007148852106183767, + 0.007976510562002659, + 0.00597401661798358, + 0.018636435270309448, + 0.0033866711892187595, + 0.010491393506526947, + 0.014360329136252403, + 0.004953527823090553, + 0.009227619506418705, + 0.005067907739430666, + 0.004159539006650448, + 0.0038609367329627275, + 0.0030713211745023727, + 0.0033191244583576918, + 0.0098379235714674, + 0.0020739000756293535, + 0.004915995057672262, + 0.07327733933925629, + 0.008759936317801476, + 0.023731106892228127, + 0.024583110585808754, + 0.014826495200395584, + 0.13259966671466827, + 0.009172193706035614, + 0.04460248723626137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.36250388622283936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.026366863399744034 + ], + "opposite_logprob_tokens": [ + -3.651366949081421 + ], + "logprob_reward": -0.026366863399744034, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366863399744034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22863928973674774, + 0.1770976483821869, + 0.1103905662894249, + 0.08689449727535248, + 0.02865018881857395, + 0.03843783214688301, + 0.03257260099053383, + 0.08793407678604126, + 0.04174394905567169, + 0.1676393449306488 + ], + "credit_map": [ + 0.029276665300130844, + 0.03818223625421524, + 0.04315711930394173, + 0.02804313786327839, + 0.023784566670656204, + 0.029915401712059975, + 0.01684270240366459, + 0.019437450915575027, + 0.02640148624777794, + 0.01563415862619877, + 0.052780888974666595, + 0.012980853207409382, + 0.010663057677447796, + 0.01875060610473156, + 0.025811536237597466, + 0.014075055718421936, + 0.013257620856165886, + 0.024543171748518944, + 0.013188065029680729, + 0.011275297962129116, + 0.01670657843351364, + 0.012100103311240673, + 0.010006261989474297, + 0.009313465096056461, + 0.022740438580513, + 0.005979409907013178, + 0.02411949262022972, + 0.004997824784368277, + 0.00764343561604619, + 0.006377240177243948, + 0.006934078875929117, + 0.008102570660412312, + 0.004333739634603262, + 0.0035303933545947075, + 0.003354885848239064, + 0.0020791387651115656, + 0.0050711724907159805, + 0.0028159930370748043, + 0.002870267489925027, + 0.0045945956371724606, + 0.005017228424549103, + 0.004965857602655888, + 0.004624760244041681, + 0.00433797063305974, + 0.006294416729360819, + 0.005145549308508635, + 0.004017658066004515, + 0.004034389741718769, + 0.0049984874203801155, + 0.0025354574900120497, + 0.004205881617963314, + 0.002709729829803109, + 0.00272887060418725, + 0.006560345645993948, + 0.004209290258586407, + 0.004624536726623774, + 0.014297951012849808, + 0.010560854338109493, + 0.023148924112319946, + 0.004996027331799269, + 0.007905885577201843, + 0.010419781319797039, + 0.005507098976522684, + 0.011097551323473454, + 0.00552644208073616, + 0.0066807991825044155, + 0.0037760762497782707, + 0.004821460228413343, + 0.004300934728235006, + 0.008222732692956924, + 0.003398172790184617, + 0.0050173308700323105, + 0.02803984470665455, + 0.0057496679946780205, + 0.015794117003679276, + 0.012204033322632313, + 0.011314154602587223, + 0.05882280319929123, + 0.0062319315038621426, + 0.02948276698589325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.026366863399744034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.06204458698630333 + ], + "opposite_logprob_tokens": [ + -2.812044620513916 + ], + "logprob_reward": -0.06204458698630333, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204458698630333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30867305397987366, + 0.07872622460126877, + 0.04739798232913017, + 0.1565675288438797, + 0.02112438529729843, + 0.03660539537668228, + 0.06847380846738815, + 0.12064914405345917, + 0.023782314732670784, + 0.13800007104873657 + ], + "credit_map": [ + 0.025281578302383423, + 0.042696449905633926, + 0.08689388632774353, + 0.016265012323856354, + 0.029396912083029747, + 0.058042190968990326, + 0.015646714717149734, + 0.034450333565473557, + 0.011141717433929443, + 0.006978414487093687, + 0.019593026489019394, + 0.006406445521861315, + 0.005908947438001633, + 0.01099474262446165, + 0.01034423429518938, + 0.0073587000370025635, + 0.005786187946796417, + 0.008766758255660534, + 0.005503614898771048, + 0.005620414391160011, + 0.006239648908376694, + 0.006663462612777948, + 0.00414104713127017, + 0.004676850978285074, + 0.024287816137075424, + 0.010471702553331852, + 0.038492679595947266, + 0.011218166910111904, + 0.019135713577270508, + 0.00996046606451273, + 0.02906309999525547, + 0.013937903568148613, + 0.0031276149675250053, + 0.003228645771741867, + 0.0025277421809732914, + 0.001967602176591754, + 0.0034747151657938957, + 0.0018184925429522991, + 0.0025654558558017015, + 0.0024141184985637665, + 0.0045649646781384945, + 0.004654433112591505, + 0.005226645618677139, + 0.003790927119553089, + 0.005428357049822807, + 0.004951768089085817, + 0.003727551084011793, + 0.004260750953108072, + 0.010444425977766514, + 0.005445983726531267, + 0.01026914268732071, + 0.005901510361582041, + 0.006143068429082632, + 0.014006000943481922, + 0.008860037662088871, + 0.007403641007840633, + 0.02495276927947998, + 0.011625781655311584, + 0.027733445167541504, + 0.007140515837818384, + 0.010293844155967236, + 0.014887175522744656, + 0.0077641489915549755, + 0.016251467168331146, + 0.0036479299888014793, + 0.002586554968729615, + 0.0017864847322925925, + 0.0024743929971009493, + 0.0021816391963511705, + 0.0061795893125236034, + 0.0024966304190456867, + 0.0024290962610393763, + 0.018375057727098465, + 0.007610940840095282, + 0.012193702161312103, + 0.010065581649541855, + 0.00783191341906786, + 0.04681757092475891, + 0.005088674370199442, + 0.030016642063856125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.06204458698630333 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.011075802147388458 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075802147388458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075802147388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43209022283554077, + 0.032945387065410614, + 0.022007940337061882, + 0.06571127474308014, + 0.050168655812740326, + 0.03042319416999817, + 0.02999328263103962, + 0.04039813205599785, + 0.006686311215162277, + 0.28957557678222656 + ], + "credit_map": [ + 0.045304011553525925, + 0.08446543663740158, + 0.10469821840524673, + 0.023798108100891113, + 0.039366938173770905, + 0.07720804214477539, + 0.012699097394943237, + 0.04455035179853439, + 0.004782546311616898, + 0.0030727293342351913, + 0.006282103713601828, + 0.002647487446665764, + 0.0029605149757117033, + 0.005090112332254648, + 0.0035228270571678877, + 0.004587063565850258, + 0.002308923052623868, + 0.0036724840756505728, + 0.0030849173199385405, + 0.002703632926568389, + 0.0032021687366068363, + 0.002656195778399706, + 0.0019783482421189547, + 0.0024012704379856586, + 0.020761482417583466, + 0.003274159738793969, + 0.018415778875350952, + 0.002186907920986414, + 0.0045465221628546715, + 0.007234684657305479, + 0.0033272174187004566, + 0.005964524578303099, + 0.009730162099003792, + 0.007458521053195, + 0.0051613361574709415, + 0.0020315300207585096, + 0.004352978430688381, + 0.012837992049753666, + 0.002735224785283208, + 0.005860909353941679, + 0.004389952402561903, + 0.0027557688299566507, + 0.004954550415277481, + 0.002476169029250741, + 0.0032713725231587887, + 0.004490796942263842, + 0.0023353302385658026, + 0.005749253556132317, + 0.0047098184004426, + 0.0023943865671753883, + 0.0049460167065262794, + 0.002021467313170433, + 0.0024672094732522964, + 0.006140153855085373, + 0.002751941094174981, + 0.004562285728752613, + 0.004478915594518185, + 0.0037751903291791677, + 0.005061185918748379, + 0.002719196258112788, + 0.006703309249132872, + 0.007826073095202446, + 0.003750286065042019, + 0.006083977874368429, + 0.0008884257986210287, + 0.0008391709998250008, + 0.0008505082805640996, + 0.0009049071231856942, + 0.0006908663199283183, + 0.0009769683238118887, + 0.000822271395009011, + 0.0007131925667636096, + 0.05943514406681061, + 0.01600431464612484, + 0.029638590291142464, + 0.012094730511307716, + 0.029685519635677338, + 0.050171371549367905, + 0.013752329163253307, + 0.07879360765218735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.011075802147388458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.012539827264845371 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539827264845371, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539827264845371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20338518917560577, + 0.07184997946023941, + 0.03996863588690758, + 0.18222884833812714, + 0.02684250846505165, + 0.04670239984989166, + 0.045384861528873444, + 0.08624467998743057, + 0.029301589354872704, + 0.2680913209915161 + ], + "credit_map": [ + 0.027381233870983124, + 0.022314129397273064, + 0.02252686582505703, + 0.02739155851304531, + 0.029428748413920403, + 0.0272577665746212, + 0.026599830016493797, + 0.02048506960272789, + 0.010816145688295364, + 0.006886264309287071, + 0.020711950957775116, + 0.0045600272715091705, + 0.005507790017873049, + 0.00746115343645215, + 0.010285323485732079, + 0.005621323827654123, + 0.005041635129600763, + 0.007096599787473679, + 0.0051398673094809055, + 0.005126739852130413, + 0.005384990479797125, + 0.004268349148333073, + 0.0036125294864177704, + 0.00429792795330286, + 0.04079572856426239, + 0.008730010129511356, + 0.050970375537872314, + 0.011027461849153042, + 0.015564355067908764, + 0.013074452057480812, + 0.022882314398884773, + 0.01918414793908596, + 0.0030783081892877817, + 0.0030601981561630964, + 0.0033391015604138374, + 0.0030646859668195248, + 0.0047761499881744385, + 0.0022247277665883303, + 0.003973457496613264, + 0.0033258781768381596, + 0.005842822138220072, + 0.005195075646042824, + 0.004400450736284256, + 0.0074508883990347385, + 0.006708867847919464, + 0.005105013493448496, + 0.007020064629614353, + 0.004979213699698448, + 0.007746838498860598, + 0.004939347971230745, + 0.006760329008102417, + 0.003864366328343749, + 0.004538293927907944, + 0.007308525964617729, + 0.0036708274856209755, + 0.0065563335083425045, + 0.012822603806853294, + 0.012426688335835934, + 0.009878387674689293, + 0.007976357825100422, + 0.010429427959024906, + 0.015216724015772343, + 0.006331068463623524, + 0.011163423769176006, + 0.0034760108683258295, + 0.0035867972765117884, + 0.004755414556711912, + 0.004154613241553307, + 0.0030242265202105045, + 0.0037932833656668663, + 0.0026884099934250116, + 0.0038228342309594154, + 0.06454184651374817, + 0.007061904761940241, + 0.019004767760634422, + 0.02006377838551998, + 0.009620572440326214, + 0.10930243134498596, + 0.007233422249555588, + 0.03126261383295059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.012539827264845371 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot vertically from above, captures a classical woman in white, with a soft-focus, vintage film quality. Her long, lustrous black hair is meticulously tied into an elegant bun, secured with a simple wooden hairpin. A few strands brush against her cheeks as she moves. She is nestled deep within a tranquil, dense bamboo forest, her feet blanketed in a thick carpet of dry, yellowing bamboo leaves. She wears a flowing white, wide-sleeved kilt, the light silk material gleaming softly in the dappled light. She twirls effortlessly in place, her steps light and effortless, her toes tapping the fallen leaves almost silently. Her wide white sleeves unfurl as she spins, drawing a perfect circle on the ground, like a slowly blooming white lotus. Dappled sunlight filters through the lush green of the bamboo leaves overhead, casting ever-changing specks of light across her pure white dress and the leaf-strewn ground, intertwining with the shadows of the bamboo forest. The painting, set against a backdrop of emerald green and withered yellow, highlights the pure white of the central figure, creating a tranquil, ethereal, and poetic atmosphere of classical aesthetics.", + "seed": 33351, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_1262_seed_33351_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": -0.16182439252734185, + "vlm_logprob": -0.16182439252734185, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.16182439252734185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.6518176794052124 + ], + "opposite_logprob_tokens": [ + -0.7768176794052124 + ], + "logprob_reward": -0.6518176794052124, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6518176794052124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06180395558476448, + 0.07746763527393341, + 0.018641632050275803, + 0.0204690583050251, + 0.150993213057518, + 0.5093252062797546, + 0.04694647714495659, + 0.06385695934295654, + 0.03603557497262955, + 0.014460302889347076 + ], + "credit_map": [ + 0.005044307094067335, + 0.003298122202977538, + 0.016114987432956696, + 0.0035170933697372675, + 0.005589115433394909, + 0.004685582593083382, + 0.004290638491511345, + 0.019264109432697296, + 0.00738659780472517, + 0.007780007552355528, + 0.01035768911242485, + 0.012355494312942028, + 0.013873971067368984, + 0.008409836329519749, + 0.0061421021819114685, + 0.011161946691572666, + 0.0022728994954377413, + 0.0014310975093394518, + 0.0023288524243980646, + 0.00264173885807395, + 0.003031488973647356, + 0.0026963346172124147, + 0.002083571394905448, + 0.0021556501742452383, + 0.0039806002750992775, + 0.0018385120201855898, + 0.0018401466077193618, + 0.0027691565919667482, + 0.0018216421594843268, + 0.0027640066109597683, + 0.002670539543032646, + 0.0027844568248838186, + 0.023995166644454002, + 0.006725657731294632, + 0.03848443552851677, + 0.015467610210180283, + 0.007306461688131094, + 0.0053149620071053505, + 0.043291207402944565, + 0.010407715104520321, + 0.05735982954502106, + 0.11360117048025131, + 0.04840203747153282, + 0.054058387875556946, + 0.03839781880378723, + 0.0872354581952095, + 0.0638275071978569, + 0.04644301161170006, + 0.008329812437295914, + 0.0057009560987353325, + 0.00560594629496336, + 0.0075687929056584835, + 0.0038174097426235676, + 0.005186190363019705, + 0.004507994279265404, + 0.0062293787486851215, + 0.01781652681529522, + 0.008677316829562187, + 0.01104620099067688, + 0.004361869767308235, + 0.003907311707735062, + 0.008914140984416008, + 0.003494714619591832, + 0.0056388783268630505, + 0.005035489797592163, + 0.003094231942668557, + 0.004481364041566849, + 0.004008505027741194, + 0.0044277021661400795, + 0.005207632668316364, + 0.00484232883900404, + 0.0049383253790438175, + 0.0015337065560743213, + 0.0013003514613956213, + 0.001578977215103805, + 0.002540361136198044, + 0.0012119937455281615, + 0.0015743939438834786, + 0.0018398016691207886, + 0.002880718559026718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.6518176794052124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.029831761494278908 + ], + "opposite_logprob_tokens": [ + -3.529831647872925 + ], + "logprob_reward": -0.029831761494278908, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029831761494278908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06361034512519836, + 0.06008860468864441, + 0.06808838993310928, + 0.027155539020895958, + 0.04499630630016327, + 0.6067662239074707, + 0.046196773648262024, + 0.046158116310834885, + 0.021179407835006714, + 0.015760404989123344 + ], + "credit_map": [ + 0.008566678501665592, + 0.006102817598730326, + 0.012955059297382832, + 0.005593929905444384, + 0.0071265515871346, + 0.008200697600841522, + 0.004925915505737066, + 0.01013869233429432, + 0.00551943713799119, + 0.00542965903878212, + 0.00836232677102089, + 0.010543863289058208, + 0.009401693008840084, + 0.007828119210898876, + 0.004948558285832405, + 0.008054947480559349, + 0.006732257083058357, + 0.007276365999132395, + 0.007981684058904648, + 0.012836546637117863, + 0.011306490749120712, + 0.00558642391115427, + 0.008098719641566277, + 0.008269903250038624, + 0.004813811741769314, + 0.0025018982123583555, + 0.0031628438737243414, + 0.004187049809843302, + 0.003447749884799123, + 0.0037768529728055, + 0.002896602964028716, + 0.002368731191381812, + 0.006663576699793339, + 0.0049412925727665424, + 0.006373527459800243, + 0.005809568800032139, + 0.006400802172720432, + 0.0036721948999911547, + 0.005040793213993311, + 0.0060945479199290276, + 0.06901232153177261, + 0.13202962279319763, + 0.060339245945215225, + 0.06554767489433289, + 0.04702657461166382, + 0.10050545632839203, + 0.07787422090768814, + 0.05443107336759567, + 0.005798406898975372, + 0.0054364134557545185, + 0.0056543247774243355, + 0.007923989556729794, + 0.005036946851760149, + 0.00572008453309536, + 0.0047170426696538925, + 0.005909564904868603, + 0.009970307350158691, + 0.0049867150373756886, + 0.006604612339287996, + 0.00413131108507514, + 0.005922704469412565, + 0.006337394937872887, + 0.0030425763688981533, + 0.0051624970510602, + 0.0038318363949656487, + 0.002057349309325218, + 0.0026334249414503574, + 0.0021352048497647047, + 0.00202344567514956, + 0.0031173075549304485, + 0.0021491055376827717, + 0.0032317331060767174, + 0.002257927320897579, + 0.0011091281194239855, + 0.0025435134302824736, + 0.0032231586519628763, + 0.0008607528288848698, + 0.0017256666906177998, + 0.001202677027322352, + 0.0028375815600156784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.029831761494278908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.1003202423453331 + ], + "opposite_logprob_tokens": [ + -2.350320339202881 + ], + "logprob_reward": -0.1003202423453331, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003202423453331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12053882330656052, + 0.26798364520072937, + 0.03512836620211601, + 0.023540198802947998, + 0.12496260553598404, + 0.14502321183681488, + 0.07780525833368301, + 0.0685136616230011, + 0.06127062439918518, + 0.07523361593484879 + ], + "credit_map": [ + 0.012573424726724625, + 0.009103335440158844, + 0.01843685284256935, + 0.01400222722440958, + 0.011523889377713203, + 0.014549508690834045, + 0.011024700477719307, + 0.029324878007173538, + 0.02553640305995941, + 0.026189805939793587, + 0.03637685626745224, + 0.047459643334150314, + 0.04815729707479477, + 0.024993836879730225, + 0.02431230992078781, + 0.034957483410835266, + 0.006208618637174368, + 0.002729160012677312, + 0.004210189450532198, + 0.004213572945445776, + 0.003970755264163017, + 0.006948935333639383, + 0.0036749979481101036, + 0.003172140335664153, + 0.0049101123586297035, + 0.0022990668658167124, + 0.002476850990206003, + 0.0023780418559908867, + 0.0028491562698036432, + 0.003825301304459572, + 0.0025027496740221977, + 0.0022989208810031414, + 0.019643256440758705, + 0.005947311408817768, + 0.03178144991397858, + 0.011632833629846573, + 0.007665027864277363, + 0.006616168189793825, + 0.03210487961769104, + 0.009571676142513752, + 0.018439881503582, + 0.023571887984871864, + 0.014212074689567089, + 0.015024491585791111, + 0.012429073452949524, + 0.025214672088623047, + 0.021256789565086365, + 0.01487434096634388, + 0.015946675091981888, + 0.009282893501222134, + 0.011372368782758713, + 0.008333095349371433, + 0.007914675399661064, + 0.007328558713197708, + 0.007931415922939777, + 0.009695581160485744, + 0.013270996510982513, + 0.007156580686569214, + 0.010014734230935574, + 0.005347119178622961, + 0.013857998885214329, + 0.008004548959434032, + 0.005396423861384392, + 0.005465261172503233, + 0.008614301681518555, + 0.004717338364571333, + 0.007736884523183107, + 0.005414546001702547, + 0.008730266243219376, + 0.00984371080994606, + 0.006227129139006138, + 0.009986448101699352, + 0.007283051032572985, + 0.007525353692471981, + 0.008626146242022514, + 0.012151297181844711, + 0.01265174150466919, + 0.007122459355741739, + 0.008584699593484402, + 0.011288860812783241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.1003202423453331 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.01607329398393631 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607329398393631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607329398393631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027841541916131973, + 0.4807140827178955, + 0.008575436659157276, + 0.008461475372314453, + 0.05558770149946213, + 0.3286888301372528, + 0.02065175026655197, + 0.03795835003256798, + 0.016746608540415764, + 0.01477421447634697 + ], + "credit_map": [ + 0.0028183329850435257, + 0.0019979544449597597, + 0.006076217163354158, + 0.002030838979408145, + 0.0027812959160655737, + 0.0025905128568410873, + 0.002037956379354, + 0.007508435752242804, + 0.04552336409687996, + 0.04519527778029442, + 0.0647900179028511, + 0.08647070825099945, + 0.08562476187944412, + 0.04623221233487129, + 0.04321514070034027, + 0.06366261839866638, + 0.001153139746747911, + 0.0006290576420724392, + 0.0009169088443741202, + 0.00107942137401551, + 0.0012649368727579713, + 0.0014631105586886406, + 0.0010671814670786262, + 0.0010016805026680231, + 0.0016519257333129644, + 0.0008474882924929261, + 0.0010400171158835292, + 0.0010197099763900042, + 0.0008198638679459691, + 0.001146818744018674, + 0.0009196610772050917, + 0.0010159905068576336, + 0.008772550150752068, + 0.002614492317661643, + 0.01397608406841755, + 0.005615753587335348, + 0.0028714428190141916, + 0.0022107278928160667, + 0.015655050054192543, + 0.003871596185490489, + 0.036848679184913635, + 0.07014181464910507, + 0.03145143762230873, + 0.03612056374549866, + 0.025584358721971512, + 0.05564675107598305, + 0.04296845942735672, + 0.02992674708366394, + 0.003064897609874606, + 0.001857386203482747, + 0.003160465508699417, + 0.0028666234575212, + 0.002900100313127041, + 0.001987508498132229, + 0.0024666718672960997, + 0.0023480975069105625, + 0.010593381710350513, + 0.005378238391131163, + 0.006913776509463787, + 0.0024670797865837812, + 0.0021344360429793596, + 0.005082511808723211, + 0.0020677982829511166, + 0.0033211284317076206, + 0.003705315524712205, + 0.0016458864556625485, + 0.0029551193583756685, + 0.001196100958622992, + 0.0011190377408638597, + 0.002317802980542183, + 0.0011496207443997264, + 0.0026577243115752935, + 0.001561331911943853, + 0.0012971584219485521, + 0.0018401460256427526, + 0.0030158234294503927, + 0.001049095531925559, + 0.0016581059899181128, + 0.0011706181103363633, + 0.0031819345895200968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.01607329398393631 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.011078985407948494 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078985407948494, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078985407948494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01336861215531826, + 0.05731528624892235, + 0.008778680115938187, + 0.006099166814237833, + 0.02282574772834778, + 0.8143459558486938, + 0.009619885124266148, + 0.02469370886683464, + 0.019040854647755623, + 0.023912062868475914 + ], + "credit_map": [ + 0.001614987151697278, + 0.0013098587514832616, + 0.002471789252012968, + 0.0018072944367304444, + 0.0015810605837032199, + 0.0015484964242205024, + 0.0009021242149174213, + 0.0021330013405531645, + 0.005546191241592169, + 0.0052267685532569885, + 0.007990583777427673, + 0.01003219559788704, + 0.009396114386618137, + 0.006703201215714216, + 0.0045900968834757805, + 0.007830132730305195, + 0.0010318629210814834, + 0.0007176215294748545, + 0.0009346859296783805, + 0.0015192553400993347, + 0.0015180821064859629, + 0.0010735595133155584, + 0.0010161162354052067, + 0.0009674967732280493, + 0.0005620695883408189, + 0.0004848810494877398, + 0.0006619669729843736, + 0.0011168888304382563, + 0.0008038821397349238, + 0.0005158621352165937, + 0.0008839795482344925, + 0.0010696363169699907, + 0.00338509539142251, + 0.001235572388395667, + 0.0052000791765749454, + 0.0024509651120752096, + 0.0016023247735574841, + 0.0007385132485069335, + 0.006475086789578199, + 0.0017381105571985245, + 0.09305180609226227, + 0.1752694845199585, + 0.08058622479438782, + 0.08798103779554367, + 0.06400059908628464, + 0.13321299850940704, + 0.10632890462875366, + 0.07391496002674103, + 0.0013771152589470148, + 0.0018254403257742524, + 0.0010810915846377611, + 0.0008071124320849776, + 0.001073159626685083, + 0.0015633890870958567, + 0.0008991615613922477, + 0.000993414898402989, + 0.006739385891705751, + 0.0038708390202373266, + 0.0041276332922279835, + 0.001578458584845066, + 0.00163289625197649, + 0.0033535249531269073, + 0.0013251148629933596, + 0.0020658564753830433, + 0.0022228567395359278, + 0.0018885141471400857, + 0.0025924642104655504, + 0.0023047197610139847, + 0.002479906426742673, + 0.0020283523481339216, + 0.0030305327381938696, + 0.0024935081601142883, + 0.0023765210062265396, + 0.00240010442212224, + 0.003955913707613945, + 0.003793031442910433, + 0.003196942387148738, + 0.0027940429281443357, + 0.002435050904750824, + 0.002960457233712077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.011078985407948494 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot vertically from above, captures a classical woman in white, with a soft-focus, vintage film quality. Her long, lustrous black hair is meticulously tied into an elegant bun, secured with a simple wooden hairpin. A few strands brush against her cheeks as she moves. She is nestled deep within a tranquil, dense bamboo forest, her feet blanketed in a thick carpet of dry, yellowing bamboo leaves. She wears a flowing white, wide-sleeved kilt, the light silk material gleaming softly in the dappled light. She twirls effortlessly in place, her steps light and effortless, her toes tapping the fallen leaves almost silently. Her wide white sleeves unfurl as she spins, drawing a perfect circle on the ground, like a slowly blooming white lotus. Dappled sunlight filters through the lush green of the bamboo leaves overhead, casting ever-changing specks of light across her pure white dress and the leaf-strewn ground, intertwining with the shadows of the bamboo forest. The painting, set against a backdrop of emerald green and withered yellow, highlights the pure white of the central figure, creating a tranquil, ethereal, and poetic atmosphere of classical aesthetics.", + "seed": 33352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_1262_seed_33352_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": -0.2003326727077365, + "vlm_logprob": -0.2003326727077365, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.2003326727077365, + "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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8552727699279785 + ], + "opposite_logprob_tokens": [ + -0.6052727699279785 + ], + "logprob_reward": -0.8552727699279785, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8552727699279785, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10434744507074356, + 0.17067332565784454, + 0.20597423613071442, + 0.03163006529211998, + 0.06940361857414246, + 0.06811390817165375, + 0.0570109486579895, + 0.08287615329027176, + 0.06514306366443634, + 0.144827201962471 + ], + "credit_map": [ + 0.011966251768171787, + 0.010530466213822365, + 0.01671968773007393, + 0.011751057580113411, + 0.021112654358148575, + 0.008973179385066032, + 0.012292636558413506, + 0.011001513339579105, + 0.013232599943876266, + 0.01442854292690754, + 0.05022524669766426, + 0.008532135747373104, + 0.0031004338525235653, + 0.00795451458543539, + 0.004630806390196085, + 0.06856905668973923, + 0.01889209635555744, + 0.007712800521403551, + 0.06805694103240967, + 0.011846392415463924, + 0.006118312478065491, + 0.01060769148170948, + 0.008082996122539043, + 0.07465701550245285, + 0.003596365451812744, + 0.002720072166994214, + 0.00363719929009676, + 0.0036605221685022116, + 0.005397091619670391, + 0.003183990251272917, + 0.0036690570414066315, + 0.005765770561993122, + 0.010711654089391232, + 0.009978769347071648, + 0.015613830648362637, + 0.0033399825915694237, + 0.0036657110322266817, + 0.0060194069519639015, + 0.003084493102505803, + 0.016989771276712418, + 0.007109546568244696, + 0.004928880836814642, + 0.020479651167988777, + 0.0026834241580218077, + 0.002631551818922162, + 0.004814681131392717, + 0.0038360878825187683, + 0.021630091592669487, + 0.009715387597680092, + 0.006191369146108627, + 0.009713792242109776, + 0.006977800279855728, + 0.007185523398220539, + 0.005256853066384792, + 0.004288543947041035, + 0.007681681774556637, + 0.009216605685651302, + 0.009828130714595318, + 0.007478718645870686, + 0.007853985764086246, + 0.012720389291644096, + 0.009695376269519329, + 0.009785400703549385, + 0.016297558322548866, + 0.010065092705190182, + 0.00908159650862217, + 0.0061891996301710606, + 0.005273832008242607, + 0.010771985165774822, + 0.009367838501930237, + 0.007542019244283438, + 0.00685150409117341, + 0.017865832895040512, + 0.013653927482664585, + 0.041669704020023346, + 0.008189878426492214, + 0.007577722892165184, + 0.02731996588408947, + 0.016704723238945007, + 0.011845453642308712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.8552727699279785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.03813574090600014 + ], + "opposite_logprob_tokens": [ + -3.2881357669830322 + ], + "logprob_reward": -0.03813574090600014, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813574090600014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09279096871614456, + 0.3264407515525818, + 0.29145240783691406, + 0.03659672662615776, + 0.0410294234752655, + 0.02639489620923996, + 0.07633180916309357, + 0.05487138777971268, + 0.02691841498017311, + 0.02717319317162037 + ], + "credit_map": [ + 0.008218402974307537, + 0.007757607381790876, + 0.013414988294243813, + 0.016624348238110542, + 0.007392150815576315, + 0.006854699924588203, + 0.009915531612932682, + 0.02261323854327202, + 0.024056270718574524, + 0.024758288636803627, + 0.0990975946187973, + 0.015395639464259148, + 0.005056027323007584, + 0.014246298000216484, + 0.00824092235416174, + 0.13558973371982574, + 0.027226394042372704, + 0.010389885865151882, + 0.0990065336227417, + 0.015980392694473267, + 0.005893000867217779, + 0.01411564089357853, + 0.00874798558652401, + 0.11009255796670914, + 0.004747801925987005, + 0.002980345394462347, + 0.006049632094800472, + 0.0038945490960031748, + 0.004517456516623497, + 0.002302391454577446, + 0.0037674754858016968, + 0.00833707395941019, + 0.006999528966844082, + 0.0064458963461220264, + 0.006481535732746124, + 0.0028908534441143274, + 0.0032200338318943977, + 0.0038349744863808155, + 0.00262815342284739, + 0.008528451435267925, + 0.0028896909207105637, + 0.0038399191107600927, + 0.0028537348844110966, + 0.0037804325111210346, + 0.0030725952237844467, + 0.003297518938779831, + 0.0022476252634078264, + 0.00441338075324893, + 0.012635722756385803, + 0.011767820455133915, + 0.008693775162100792, + 0.0054200179874897, + 0.01689945161342621, + 0.007393409963697195, + 0.0065382057800889015, + 0.006983404513448477, + 0.0052167861722409725, + 0.0061448486521840096, + 0.004500455688685179, + 0.0043894085101783276, + 0.008536314591765404, + 0.01006896048784256, + 0.004908626899123192, + 0.011105986312031746, + 0.0033931622747331858, + 0.0038009556010365486, + 0.005085053853690624, + 0.0026023252867162228, + 0.002173837972804904, + 0.003719493979588151, + 0.0032079468946903944, + 0.0029356384184211493, + 0.004171141888946295, + 0.004030464682728052, + 0.00328184780664742, + 0.002957099350169301, + 0.0021861575078219175, + 0.00334480544552207, + 0.0030931050423532724, + 0.004108572844415903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.03813574090600014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.0789848193526268 + ], + "opposite_logprob_tokens": [ + -2.5789847373962402 + ], + "logprob_reward": -0.0789848193526268, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789848193526268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06666682660579681, + 0.18136735260486603, + 0.15678004920482635, + 0.047395024448633194, + 0.04581725224852562, + 0.06708552688360214, + 0.12802833318710327, + 0.07747136056423187, + 0.047001440078020096, + 0.1823868602514267 + ], + "credit_map": [ + 0.006481812801212072, + 0.007080560550093651, + 0.011181999929249287, + 0.007388392463326454, + 0.00796433538198471, + 0.006267376709729433, + 0.007858021184802055, + 0.012444325722754002, + 0.013909481465816498, + 0.014368569478392601, + 0.0521484836935997, + 0.00833122618496418, + 0.004513629246503115, + 0.00778360478579998, + 0.006372333969920874, + 0.07394002377986908, + 0.014892155304551125, + 0.006187980994582176, + 0.04558150842785835, + 0.009140430018305779, + 0.007218805607408285, + 0.007450996898114681, + 0.01111047063022852, + 0.05519769340753555, + 0.003951790742576122, + 0.002853326266631484, + 0.01362830400466919, + 0.004088204819709063, + 0.0032011920120567083, + 0.004280129913240671, + 0.002614116296172142, + 0.01277796272188425, + 0.009293493814766407, + 0.008140036836266518, + 0.00579350721091032, + 0.00450719241052866, + 0.003965257201343775, + 0.005387251731008291, + 0.00389271043241024, + 0.0048378026112914085, + 0.007331786677241325, + 0.004405207000672817, + 0.01648867130279541, + 0.005388056393712759, + 0.004434346687048674, + 0.005591296590864658, + 0.004419336561113596, + 0.019026823341846466, + 0.015247865580022335, + 0.015687013044953346, + 0.028887126594781876, + 0.010519205592572689, + 0.01637227088212967, + 0.012806270271539688, + 0.010368341580033302, + 0.01814025081694126, + 0.008207960054278374, + 0.009620830416679382, + 0.006999967619776726, + 0.008291434496641159, + 0.013960760086774826, + 0.007342865690588951, + 0.010105708613991737, + 0.012941835448145866, + 0.008763208985328674, + 0.008278521709144115, + 0.0071425242349505424, + 0.002720347372815013, + 0.00439388258382678, + 0.008549140766263008, + 0.003860240336507559, + 0.00329357641749084, + 0.028499161824584007, + 0.01654222048819065, + 0.050316423177719116, + 0.010365533642470837, + 0.008863775990903378, + 0.035405196249485016, + 0.019919076934456825, + 0.012475461699068546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.0789848193526268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.018191400915384293 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.018191400915384293, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191400915384293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03615180402994156, + 0.22559389472007751, + 0.04151881858706474, + 0.014411786571145058, + 0.03941234201192856, + 0.02687842957675457, + 0.0277140773832798, + 0.07795032113790512, + 0.05881890282034874, + 0.4515496790409088 + ], + "credit_map": [ + 0.00413023354485631, + 0.0045442585833370686, + 0.003800208680331707, + 0.005009121261537075, + 0.004576615523546934, + 0.0032193714287132025, + 0.003378537716343999, + 0.007493455894291401, + 0.016882793977856636, + 0.016834106296300888, + 0.06849856674671173, + 0.010259934701025486, + 0.003205531043931842, + 0.009512459859251976, + 0.005056486930698156, + 0.09534400701522827, + 0.004587856121361256, + 0.004213391337543726, + 0.009815098717808723, + 0.00471872091293335, + 0.0015353993512690067, + 0.0027871730271726847, + 0.002159934490919113, + 0.011701242998242378, + 0.001730627380311489, + 0.0017221353482455015, + 0.0028080190531909466, + 0.0012838720576837659, + 0.0016472168499603868, + 0.0010429881513118744, + 0.001108593656681478, + 0.0030683339573442936, + 0.00363787985406816, + 0.0024087224155664444, + 0.012851598672568798, + 0.0041225384920835495, + 0.0014318404719233513, + 0.0020783424843102694, + 0.0023954191710799932, + 0.010486000217497349, + 0.0038526970893144608, + 0.004318071063607931, + 0.004637180361896753, + 0.001379100140184164, + 0.0014997400576248765, + 0.0037816979456692934, + 0.0011327253887429833, + 0.006277215667068958, + 0.004702327772974968, + 0.004178980831056833, + 0.0034477598965168, + 0.003072962863370776, + 0.0030870470218360424, + 0.0029533887282013893, + 0.0020485396962612867, + 0.00422306964173913, + 0.00750189833343029, + 0.008542987518012524, + 0.008016135543584824, + 0.005953336600214243, + 0.0086361113935709, + 0.015356051735579967, + 0.0041358754970133305, + 0.01980791985988617, + 0.01040203683078289, + 0.011000591330230236, + 0.010310370475053787, + 0.0037649672012776136, + 0.0026152085047215223, + 0.013043411076068878, + 0.0033126212656497955, + 0.004369692876935005, + 0.07246513664722443, + 0.03665492311120033, + 0.12889975309371948, + 0.02679242193698883, + 0.020153120160102844, + 0.0866004228591919, + 0.05023914575576782, + 0.029744736850261688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.018191400915384293 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.011078632436692715 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078632436692715, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078632436692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10992585122585297, + 0.16495391726493835, + 0.06982795149087906, + 0.07012936472892761, + 0.09681717306375504, + 0.08302000910043716, + 0.07514207810163498, + 0.08557020872831345, + 0.08365131169557571, + 0.16096213459968567 + ], + "credit_map": [ + 0.01371174305677414, + 0.01499219797551632, + 0.010991637594997883, + 0.012575753033161163, + 0.017758410423994064, + 0.010554272681474686, + 0.014650272205471992, + 0.014691568911075592, + 0.012393240816891193, + 0.012303430587053299, + 0.04779129475355148, + 0.008074512705206871, + 0.004189412575215101, + 0.007031301502138376, + 0.007776801008731127, + 0.06539393961429596, + 0.0058554233983159065, + 0.0031390830408781767, + 0.01912291906774044, + 0.0061652930453419685, + 0.007406597025692463, + 0.003107052529230714, + 0.0060211909003555775, + 0.0190103929489851, + 0.011778349988162518, + 0.010839889757335186, + 0.010270982049405575, + 0.009177199564874172, + 0.007282238453626633, + 0.006423652172088623, + 0.006674428470432758, + 0.007682625204324722, + 0.022614674642682076, + 0.019884832203388214, + 0.010036000050604343, + 0.00946399848908186, + 0.00836495403200388, + 0.011799178086221218, + 0.006462507881224155, + 0.008191022090613842, + 0.008670749142765999, + 0.004028861876577139, + 0.026082703843712807, + 0.003771040355786681, + 0.002250615507364273, + 0.005615133326500654, + 0.004763862118124962, + 0.02783704362809658, + 0.013916350901126862, + 0.012926214374601841, + 0.007752475328743458, + 0.006409470457583666, + 0.014566099271178246, + 0.007034622598439455, + 0.005331782624125481, + 0.007205063942819834, + 0.009835204109549522, + 0.010927782393991947, + 0.007431198377162218, + 0.008840732276439667, + 0.01470794528722763, + 0.008163021877408028, + 0.010893341153860092, + 0.014770986512303352, + 0.011817088350653648, + 0.010915151797235012, + 0.009433289989829063, + 0.006455984432250261, + 0.012834274210035801, + 0.014778615906834602, + 0.008315835148096085, + 0.009101072326302528, + 0.02045239508152008, + 0.011720151640474796, + 0.04844382777810097, + 0.010017788968980312, + 0.007956212386488914, + 0.032144058495759964, + 0.018771354109048843, + 0.011456347070634365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.011078632436692715 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot vertically from above, captures a classical woman in white, with a soft-focus, vintage film quality. Her long, lustrous black hair is meticulously tied into an elegant bun, secured with a simple wooden hairpin. A few strands brush against her cheeks as she moves. She is nestled deep within a tranquil, dense bamboo forest, her feet blanketed in a thick carpet of dry, yellowing bamboo leaves. She wears a flowing white, wide-sleeved kilt, the light silk material gleaming softly in the dappled light. She twirls effortlessly in place, her steps light and effortless, her toes tapping the fallen leaves almost silently. Her wide white sleeves unfurl as she spins, drawing a perfect circle on the ground, like a slowly blooming white lotus. Dappled sunlight filters through the lush green of the bamboo leaves overhead, casting ever-changing specks of light across her pure white dress and the leaf-strewn ground, intertwining with the shadows of the bamboo forest. The painting, set against a backdrop of emerald green and withered yellow, highlights the pure white of the central figure, creating a tranquil, ethereal, and poetic atmosphere of classical aesthetics.", + "seed": 33353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_1262_seed_33353_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": -0.1847893189638853, + "vlm_logprob": -0.1847893189638853, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.1847893189638853, + "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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7776942253112793 + ], + "opposite_logprob_tokens": [ + -0.6526942253112793 + ], + "logprob_reward": -0.7776942253112793, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7776942253112793, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12278636544942856, + 0.09279873967170715, + 0.03940826654434204, + 0.029678884893655777, + 0.020993351936340332, + 0.04133521392941475, + 0.18766550719738007, + 0.303760826587677, + 0.11700907349586487, + 0.04456380382180214 + ], + "credit_map": [ + 0.025008639320731163, + 0.014943959191441536, + 0.019344400614500046, + 0.013213111087679863, + 0.01375674270093441, + 0.00974655244499445, + 0.012006490491330624, + 0.01476646400988102, + 0.007233844138681889, + 0.0071191745810210705, + 0.02295704372227192, + 0.0053875683806836605, + 0.006992206443101168, + 0.006831509992480278, + 0.009477038867771626, + 0.026800353080034256, + 0.0038572235498577356, + 0.007262421306222677, + 0.006327264942228794, + 0.0032458205241709948, + 0.0044794646091759205, + 0.00627490971237421, + 0.0024812163319438696, + 0.005479942541569471, + 0.0026183798909187317, + 0.002384501975029707, + 0.004041657317429781, + 0.00312482169829309, + 0.00521430280059576, + 0.004014170728623867, + 0.003924246411770582, + 0.004356803838163614, + 0.002328746486455202, + 0.0017096770461648703, + 0.0027829809114336967, + 0.002874347846955061, + 0.0022716561798006296, + 0.0025578532367944717, + 0.0025957280304282904, + 0.003872363595291972, + 0.00422315439209342, + 0.004468307364732027, + 0.005311222281306982, + 0.005701478570699692, + 0.0047805472277104855, + 0.005359596107155085, + 0.0060330284759402275, + 0.0054578762501478195, + 0.02012140303850174, + 0.018708955496549606, + 0.02523096650838852, + 0.020343052223324776, + 0.01712043397128582, + 0.045478466898202896, + 0.014798249118030071, + 0.025863967835903168, + 0.04133091866970062, + 0.019513580948114395, + 0.04035767540335655, + 0.025731928646564484, + 0.027904769405722618, + 0.056444890797138214, + 0.03387356549501419, + 0.058603495359420776, + 0.01703769527375698, + 0.008732708171010017, + 0.022527603432536125, + 0.010815349407494068, + 0.011810783296823502, + 0.014224051497876644, + 0.02150828391313553, + 0.010352603159844875, + 0.0055671739391982555, + 0.005587253253906965, + 0.005382193718105555, + 0.005275032948702574, + 0.0101647162809968, + 0.0050062029622495174, + 0.0036948788911104202, + 0.0038863534573465586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.7776942253112793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.02983083575963974 + ], + "opposite_logprob_tokens": [ + -3.5298309326171875 + ], + "logprob_reward": -0.02983083575963974, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983083575963974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09575492888689041, + 0.033097755163908005, + 0.2980966866016388, + 0.019979232922196388, + 0.023866593837738037, + 0.01545824483036995, + 0.12793083488941193, + 0.22081488370895386, + 0.1370401233434677, + 0.02796073630452156 + ], + "credit_map": [ + 0.011804981157183647, + 0.006706222426146269, + 0.020660080015659332, + 0.008324363268911839, + 0.009561926126480103, + 0.006025498267263174, + 0.010141607373952866, + 0.022530240938067436, + 0.0029592816717922688, + 0.002518658759072423, + 0.009171507321298122, + 0.0034144294913858175, + 0.0031285749282687902, + 0.0025192226748913527, + 0.003829422639682889, + 0.005556656513363123, + 0.02121168002486229, + 0.01723182387650013, + 0.10551184415817261, + 0.009367918595671654, + 0.006990253925323486, + 0.03260582685470581, + 0.01077496912330389, + 0.09440234303474426, + 0.0021177702583372593, + 0.0025057035963982344, + 0.0027867776807397604, + 0.0023976597003638744, + 0.0020127228926867247, + 0.0031184100080281496, + 0.002368711633607745, + 0.0026714755222201347, + 0.002107368316501379, + 0.001827642903663218, + 0.0027931942604482174, + 0.0040148161351680756, + 0.0027232065331190825, + 0.003080119611695409, + 0.0034053167328238487, + 0.003914929460734129, + 0.0019133517052978277, + 0.0011977053945884109, + 0.0021921631414443254, + 0.001886497950181365, + 0.0020368231926113367, + 0.0017040931852534413, + 0.002305550267919898, + 0.0022220592945814133, + 0.009734689258038998, + 0.00932806171476841, + 0.01878313347697258, + 0.009381050243973732, + 0.017093008384108543, + 0.03687037155032158, + 0.010386787354946136, + 0.016353728249669075, + 0.034378353506326675, + 0.01038926187902689, + 0.05280712991952896, + 0.016142230480909348, + 0.013904127292335033, + 0.04281444475054741, + 0.019814658910036087, + 0.030564675107598305, + 0.020345626398921013, + 0.007859306409955025, + 0.03200564160943031, + 0.010973154567182064, + 0.011692671105265617, + 0.015551673248410225, + 0.020511982962489128, + 0.0181000754237175, + 0.00293362932279706, + 0.0034944762010127306, + 0.0034543084912002087, + 0.0037698072846978903, + 0.005278587806969881, + 0.003241699654608965, + 0.0025855458807200193, + 0.003202682128176093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.02983083575963974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.07899417728185654 + ], + "opposite_logprob_tokens": [ + -2.5789942741394043 + ], + "logprob_reward": -0.07899417728185654, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899417728185654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1443730592727661, + 0.037132758647203445, + 0.30914339423179626, + 0.014764511026442051, + 0.02695256844162941, + 0.013027633540332317, + 0.05658963695168495, + 0.29403674602508545, + 0.05448276922106743, + 0.04949692264199257 + ], + "credit_map": [ + 0.014050333760678768, + 0.009826297871768475, + 0.04066600650548935, + 0.008971268311142921, + 0.007078357972204685, + 0.008730149827897549, + 0.010264027863740921, + 0.04478663206100464, + 0.002443223260343075, + 0.0027520768344402313, + 0.009886122308671474, + 0.0026965553406625986, + 0.003525467123836279, + 0.0027342550456523895, + 0.0033629233948886395, + 0.009732142090797424, + 0.021314406767487526, + 0.017473604530096054, + 0.10874313861131668, + 0.009661830961704254, + 0.005791415460407734, + 0.032765086740255356, + 0.011209540069103241, + 0.10218439251184464, + 0.0014593807281926274, + 0.001396412611939013, + 0.001992993988096714, + 0.0018219135235995054, + 0.001723935129120946, + 0.0018308385042473674, + 0.002752509666606784, + 0.001786528737284243, + 0.0017625686014071107, + 0.0017044838750734925, + 0.003926558420062065, + 0.004020041320472956, + 0.003719508182257414, + 0.00394725427031517, + 0.0034345616586506367, + 0.0044375937432050705, + 0.0012292180908843875, + 0.0011001378297805786, + 0.0015312335453927517, + 0.0017749725375324488, + 0.002700110897421837, + 0.0011909755412489176, + 0.0018688795389607549, + 0.0016321068396791816, + 0.009148052893579006, + 0.006921410094946623, + 0.011115381494164467, + 0.006826619617640972, + 0.005658374167978764, + 0.005225205793976784, + 0.004902414977550507, + 0.006792181171476841, + 0.05844217166304588, + 0.019165411591529846, + 0.05898444727063179, + 0.023553144186735153, + 0.0225199107080698, + 0.06644520908594131, + 0.022124767303466797, + 0.022801682353019714, + 0.007101281080394983, + 0.0023799689952284098, + 0.014518612064421177, + 0.004825156647711992, + 0.004037446342408657, + 0.004512426909059286, + 0.011974585242569447, + 0.005133297760039568, + 0.006356766913086176, + 0.006503314711153507, + 0.006063053384423256, + 0.00728919031098485, + 0.008567756041884422, + 0.005959044676274061, + 0.0039085038006305695, + 0.0048492974601686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.07899417728185654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.02635037526488304 + ], + "opposite_logprob_tokens": [ + -3.651350259780884 + ], + "logprob_reward": -0.02635037526488304, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635037526488304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13385261595249176, + 0.02214980684220791, + 0.7390738129615784, + 0.011110792867839336, + 0.007088461425155401, + 0.004211344290524721, + 0.04258265346288681, + 0.019473552703857422, + 0.01017831265926361, + 0.01027857605367899 + ], + "credit_map": [ + 0.020801108330488205, + 0.013968675397336483, + 0.0332481786608696, + 0.010876476764678955, + 0.006482764612883329, + 0.00678327726200223, + 0.010175369679927826, + 0.031516775488853455, + 0.0027760937809944153, + 0.0034465466160327196, + 0.0049641854129731655, + 0.0013786210911348462, + 0.0012990918476134539, + 0.0029601817950606346, + 0.001704823225736618, + 0.003620265983045101, + 0.051718659698963165, + 0.043093759566545486, + 0.2600633203983307, + 0.022522097453475, + 0.014414171688258648, + 0.08019718527793884, + 0.026474980637431145, + 0.24058973789215088, + 0.0022571312729269266, + 0.0018950523808598518, + 0.0017847351264208555, + 0.001054230029694736, + 0.0009439681307412684, + 0.0008250887622125447, + 0.0009932841639965773, + 0.0013573041651397943, + 0.0010004963260143995, + 0.000663572340272367, + 0.000980017357505858, + 0.0007391477120108902, + 0.0005573419039137661, + 0.0008250612299889326, + 0.000619302736595273, + 0.0017035224009305239, + 0.0005598195712082088, + 0.0005086242454126477, + 0.0006630672723986208, + 0.0004518190398812294, + 0.0003804992011282593, + 0.000419894844526425, + 0.0004011972341686487, + 0.0008264235802926123, + 0.005550792906433344, + 0.0044265990145504475, + 0.0063431342132389545, + 0.0030508977361023426, + 0.003158936044201255, + 0.011390953324735165, + 0.0020451461896300316, + 0.006616197992116213, + 0.0032087836880236864, + 0.0010878201574087143, + 0.0037518860772252083, + 0.001514206174761057, + 0.0015774376224726439, + 0.0035645763855427504, + 0.001706844661384821, + 0.00306199979968369, + 0.002250684890896082, + 0.0007286382606253028, + 0.0016732367221266031, + 0.0009014891693368554, + 0.0009281524107791483, + 0.0016936801839619875, + 0.0010600194800645113, + 0.0009424121817573905, + 0.001743623404763639, + 0.0010942372027784586, + 0.0013897574972361326, + 0.0010574363404884934, + 0.0013583081308752298, + 0.001704465365037322, + 0.0006436905241571367, + 0.0012870586942881346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.02635037526488304 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vertical view from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure a woman with long black hair tied in a bun secured by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a flowing white wide-sleeved kilt made of light silk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a dense bamboo forest with dry yellowing leaves on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman twirl in place while her sleeves unfurl in a 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": [ + -0.011076981201767921 + ], + "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": [ + 0.37444084882736206, + 0.03427225351333618, + 0.194520965218544, + 0.02332479879260063, + 0.050242673605680466, + 0.017214754596352577, + 0.025681868195533752, + 0.14863845705986023, + 0.08928806334733963, + 0.04237527772784233 + ], + "credit_map": [ + 0.04380826652050018, + 0.025829942896962166, + 0.111601322889328, + 0.02155461348593235, + 0.010709760710597038, + 0.019722873345017433, + 0.02731461450457573, + 0.11389948427677155, + 0.0030093768145889044, + 0.004488738253712654, + 0.003772295545786619, + 0.0032488706056028605, + 0.005589609034359455, + 0.0033777852077037096, + 0.005573383532464504, + 0.0052121952176094055, + 0.013857176527380943, + 0.01185196079313755, + 0.06819723546504974, + 0.006300363689661026, + 0.003996803425252438, + 0.021344665437936783, + 0.006939917802810669, + 0.06203284114599228, + 0.0019073125440627337, + 0.0018895607208833098, + 0.002429281361401081, + 0.002266528084874153, + 0.005625426769256592, + 0.003445061855018139, + 0.002945968881249428, + 0.0028156600892543793, + 0.0047746156342327595, + 0.0033121721353381872, + 0.006255871616303921, + 0.00854479894042015, + 0.005733783356845379, + 0.007758863735944033, + 0.006057193502783775, + 0.007805374450981617, + 0.001294219633564353, + 0.0016642115078866482, + 0.002313439967110753, + 0.0024161653127521276, + 0.002344034379348159, + 0.002107002306729555, + 0.00239813351072371, + 0.0026775465812534094, + 0.0021737825591117144, + 0.0021455204114317894, + 0.004403304308652878, + 0.0031038306187838316, + 0.0029691909439861774, + 0.004803035408258438, + 0.0021099941805005074, + 0.003973207902163267, + 0.021229205653071404, + 0.008137224242091179, + 0.038724012672901154, + 0.009475884027779102, + 0.009252465330064297, + 0.03420763835310936, + 0.010457303375005722, + 0.017154717817902565, + 0.01381741464138031, + 0.005341004580259323, + 0.020714979618787766, + 0.007556010968983173, + 0.00596949178725481, + 0.010595684871077538, + 0.014106327667832375, + 0.011187154799699783, + 0.005484717432409525, + 0.0038033605087548494, + 0.004876195453107357, + 0.004919532686471939, + 0.00479467399418354, + 0.007976597175002098, + 0.004617512226104736, + 0.005902682896703482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.011076981201767921 + } + ], + "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/0000125/group_8/index_tensor([7372, 7372])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_8/index_tensor([7372, 7372])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..505d2de299516b0a123faebf36c5e60493bd1805 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_8/index_tensor([7372, 7372])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up triggers a slow, deliberate downward tilt, followed by a continuous shot of the action. The frame initially fills with a long, dagger-shaped icicle, its translucent crystal surface riddled with imperfections and containing tiny air bubbles, suspended from the weathered edge of a black asphalt shingle roof. Bright, low-angle winter sunlight bathes the scene, causing the icicle to glitter as it melts at its connection point. The camera meticulously captures the formation, expansion, and shedding of a droplet as the solid base visibly thins, transforming into a fragile, watery thread. As the connection inevitably weakens, a hairline crack suddenly appears at the top. With a silent, decisive snap, the icicle breaks under its own weight and separates from the roof. Pulled downward by gravity, it plummets, the camera tilting steadily downward to follow its descent against a blurred background. The icicle's fall reaches its peak as it makes contact with the frozen ground below, violently shattering into countless glittering crystal fragments and a mist of fine ice powder that catches the light, creating a brief, shimmering explosion that conveys a tranquil yet dramatic atmosphere of natural transition.", + "seed": 1338, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_8/index_7372_seed_1338_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": -0.12847256939858198, + "vlm_logprob": -0.12847256939858198, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.12847256939858198, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.009352208115160465 + ], + "opposite_logprob_tokens": [ + -6.384352207183838 + ], + "logprob_reward": -0.009352208115160465, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.009352208115160465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10053041577339172, + 0.06381814926862717, + 0.06848157197237015, + 0.09520163387060165, + 0.0861133486032486, + 0.06762215495109558, + 0.06074434891343117, + 0.0620298907160759, + 0.12885035574436188, + 0.2666080892086029 + ], + "credit_map": [ + 0.006546676624566317, + 0.009223169647157192, + 0.018117297440767288, + 0.012687770649790764, + 0.010668646544218063, + 0.018656697124242783, + 0.007568493485450745, + 0.017061669379472733, + 0.004425759427249432, + 0.0069488100707530975, + 0.012733670882880688, + 0.006222274620085955, + 0.009912031702697277, + 0.01152707263827324, + 0.0036777087952941656, + 0.008370829746127129, + 0.004374569281935692, + 0.007485521025955677, + 0.016858018934726715, + 0.0038714066613465548, + 0.007739826571196318, + 0.014746306464076042, + 0.00562549801543355, + 0.007780428975820541, + 0.0068275234661996365, + 0.007062425836920738, + 0.01471713837236166, + 0.004687508102506399, + 0.0070037334226071835, + 0.016873903572559357, + 0.018628636375069618, + 0.019400769844651222, + 0.009759055450558662, + 0.0064363013952970505, + 0.020743630826473236, + 0.00519243348389864, + 0.008321822620928288, + 0.014728919602930546, + 0.013996412046253681, + 0.006934778764843941, + 0.004654931835830212, + 0.003463584231212735, + 0.009098437614738941, + 0.002487439662218094, + 0.007403657305985689, + 0.018990958109498024, + 0.006438651587814093, + 0.015084496699273586, + 0.008233530446887016, + 0.0034307113382965326, + 0.008532471023499966, + 0.00334527506493032, + 0.016104983165860176, + 0.009194866754114628, + 0.006373555399477482, + 0.005528963170945644, + 0.006170921493321657, + 0.009832015261054039, + 0.005062301177531481, + 0.00796582829207182, + 0.010196581482887268, + 0.010224009864032269, + 0.005699503235518932, + 0.006878735963255167, + 0.007014583796262741, + 0.008094284683465958, + 0.04339537024497986, + 0.012568766251206398, + 0.0060229431837797165, + 0.006442131008952856, + 0.010861863382160664, + 0.03445042297244072, + 0.02057524211704731, + 0.015943041071295738, + 0.013721877709031105, + 0.0551178902387619, + 0.015980755910277367, + 0.020046669989824295, + 0.032652415335178375, + 0.09257019311189651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.009352208115160465 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.007647404447197914 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.007647404447197914, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647404447197914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1178271695971489, + 0.06328955292701721, + 0.09622137993574142, + 0.15195311605930328, + 0.053235650062561035, + 0.08951273560523987, + 0.08352932333946228, + 0.10132535547018051, + 0.18245097994804382, + 0.06065470352768898 + ], + "credit_map": [ + 0.008746951818466187, + 0.016643082723021507, + 0.020693626254796982, + 0.020053230226039886, + 0.012965064495801926, + 0.00950772874057293, + 0.01062640268355608, + 0.018591077998280525, + 0.003967237658798695, + 0.007935616187751293, + 0.012328780256211758, + 0.008743420243263245, + 0.0075456080958247185, + 0.008484487421810627, + 0.0052129607647657394, + 0.009071441367268562, + 0.006167084444314241, + 0.00905549991875887, + 0.0243717972189188, + 0.0059202671982347965, + 0.013354317285120487, + 0.020134877413511276, + 0.005558710545301437, + 0.011658822186291218, + 0.012658954598009586, + 0.012853236868977547, + 0.024517346173524857, + 0.008728999644517899, + 0.01135145965963602, + 0.026710033416748047, + 0.023768030107021332, + 0.0313650444149971, + 0.0054985941387712955, + 0.0038629404734820127, + 0.013817458413541317, + 0.0032045356929302216, + 0.00504954531788826, + 0.009687933139503002, + 0.0071694692596793175, + 0.004945175256580114, + 0.0070824832655489445, + 0.00854587648063898, + 0.015983501449227333, + 0.006315857172012329, + 0.01075698807835579, + 0.012109932489693165, + 0.013917531818151474, + 0.014800571836531162, + 0.013283340260386467, + 0.004020276013761759, + 0.012395167723298073, + 0.0072054630145430565, + 0.014359181746840477, + 0.013432353734970093, + 0.008641731925308704, + 0.010191804729402065, + 0.008766580373048782, + 0.01870579831302166, + 0.009654314257204533, + 0.014627453871071339, + 0.014645330607891083, + 0.01780969835817814, + 0.008105787448585033, + 0.009010388515889645, + 0.014129654504358768, + 0.020788975059986115, + 0.04754072427749634, + 0.02725485898554325, + 0.009396548382937908, + 0.01105569675564766, + 0.018742304295301437, + 0.03354222699999809, + 0.0044420380145311356, + 0.005801071412861347, + 0.010896981693804264, + 0.006883374880999327, + 0.007303302641957998, + 0.0077140675857663155, + 0.0053873793222010136, + 0.012226484715938568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.007647404447197914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.03809339925646782 + ], + "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": [ + 0.12212155759334564, + 0.06492061913013458, + 0.061260707676410675, + 0.07711929082870483, + 0.08797735720872879, + 0.06731245666742325, + 0.13239209353923798, + 0.12538020312786102, + 0.12098689377307892, + 0.1405288577079773 + ], + "credit_map": [ + 0.007324683479964733, + 0.015623838640749454, + 0.021522043272852898, + 0.01919381134212017, + 0.012954780831933022, + 0.015498588792979717, + 0.010956548154354095, + 0.019047247245907784, + 0.003982636146247387, + 0.00879431702196598, + 0.00942927785217762, + 0.008193952031433582, + 0.009492779150605202, + 0.008109156973659992, + 0.007461005821824074, + 0.009457487612962723, + 0.004011932294815779, + 0.008830911479890347, + 0.01157057099044323, + 0.004367099609225988, + 0.007567842956632376, + 0.012609842233359814, + 0.005185479298233986, + 0.007117022294551134, + 0.0071112471632659435, + 0.012691328302025795, + 0.012821759097278118, + 0.00759859150275588, + 0.007110958453267813, + 0.010084162466228008, + 0.008211895823478699, + 0.011489346623420715, + 0.010639051906764507, + 0.0061066835187375546, + 0.018983395770192146, + 0.005240458529442549, + 0.008520619943737984, + 0.013660809025168419, + 0.01570563204586506, + 0.009120693430304527, + 0.006772767286747694, + 0.00677894102409482, + 0.01044787559658289, + 0.004426567815244198, + 0.008134396746754646, + 0.012161975726485252, + 0.008526979014277458, + 0.010062942281365395, + 0.019629398360848427, + 0.007248471491038799, + 0.024023471400141716, + 0.007917829789221287, + 0.02423880808055401, + 0.018693339079618454, + 0.01859675534069538, + 0.01204400509595871, + 0.014921700581908226, + 0.022933846339583397, + 0.012312667444348335, + 0.018111011013388634, + 0.018302276730537415, + 0.01921171508729458, + 0.009556982666254044, + 0.010029996745288372, + 0.00995703972876072, + 0.010715329088270664, + 0.02993054874241352, + 0.021389639005064964, + 0.006253709085285664, + 0.00873985979706049, + 0.013391789980232716, + 0.020608972758054733, + 0.011810707859694958, + 0.007080260664224625, + 0.007674835622310638, + 0.020158471539616585, + 0.012144474312663078, + 0.008597181178629398, + 0.02925771474838257, + 0.043805208057165146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.03809339925646782 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.5761873126029968 + ], + "opposite_logprob_tokens": [ + -0.8261873126029968 + ], + "logprob_reward": -0.5761873126029968, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761873126029968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06679458916187286, + 0.06620418280363083, + 0.03203964978456497, + 0.07864174246788025, + 0.05475567653775215, + 0.07019772380590439, + 0.0508234016597271, + 0.03275143727660179, + 0.2864687442779541, + 0.26132285594940186 + ], + "credit_map": [ + 0.004633921664208174, + 0.008507062681019306, + 0.012483818456530571, + 0.008111629635095596, + 0.009764539077877998, + 0.007241870742291212, + 0.0061314222402870655, + 0.009920326061546803, + 0.004542986396700144, + 0.007408616598695517, + 0.015523644164204597, + 0.007252694107592106, + 0.005176347680389881, + 0.008527404628694057, + 0.006043740548193455, + 0.011728751473128796, + 0.0018563029589131474, + 0.0018040643772110343, + 0.00561315193772316, + 0.0013268657494336367, + 0.0035569481551647186, + 0.010115180164575577, + 0.002211164217442274, + 0.0055559719912707806, + 0.007253298070281744, + 0.010478928685188293, + 0.01327439770102501, + 0.005822708364576101, + 0.006050904281437397, + 0.011590488255023956, + 0.009625019505620003, + 0.014546000398695469, + 0.007078930735588074, + 0.002840107772499323, + 0.012788579799234867, + 0.002342985011637211, + 0.004016548860818148, + 0.009503749199211597, + 0.009110845625400543, + 0.007073931861668825, + 0.005681267939507961, + 0.005868065170943737, + 0.011670134961605072, + 0.00412740046158433, + 0.007983990013599396, + 0.011796805076301098, + 0.012474081479012966, + 0.01059598010033369, + 0.006502384785562754, + 0.003355738939717412, + 0.010168156586587429, + 0.0053171031177043915, + 0.007406312972307205, + 0.00703843729570508, + 0.006000361405313015, + 0.005034907255321741, + 0.003418434876948595, + 0.004923045635223389, + 0.0044794646091759205, + 0.005035580601543188, + 0.004655049182474613, + 0.004792626481503248, + 0.0029339632019400597, + 0.0025132710579782724, + 0.018982931971549988, + 0.015288305468857288, + 0.08779311925172806, + 0.04292335361242294, + 0.010080532170832157, + 0.014843069016933441, + 0.02963872440159321, + 0.06691873073577881, + 0.02272261120378971, + 0.012931679375469685, + 0.015758337453007698, + 0.04520198330283165, + 0.02153927832841873, + 0.014675251208245754, + 0.03959885984659195, + 0.08889483660459518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.5761873126029968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.011082522571086884 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082522571086884, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082522571086884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07625214755535126, + 0.03857845813035965, + 0.059104032814502716, + 0.1432507485151291, + 0.07291264832019806, + 0.1804903745651245, + 0.052897859364748, + 0.0448027104139328, + 0.06832385808229446, + 0.2633872330188751 + ], + "credit_map": [ + 0.005101438146084547, + 0.00932906474918127, + 0.011486123315989971, + 0.009793177247047424, + 0.009124740026891232, + 0.013115378096699715, + 0.0065698884427547455, + 0.01173233613371849, + 0.0033672316931188107, + 0.003507726127281785, + 0.008282321505248547, + 0.0028410139493644238, + 0.004506565630435944, + 0.007696435786783695, + 0.0025500168558210135, + 0.005827150773257017, + 0.004292510915547609, + 0.004264873918145895, + 0.010410181246697903, + 0.003488173708319664, + 0.007349432446062565, + 0.016164027154445648, + 0.004437115043401718, + 0.00869771558791399, + 0.010254924185574055, + 0.010023712180554867, + 0.024838780984282494, + 0.006319950334727764, + 0.010255727916955948, + 0.025935225188732147, + 0.023236054927110672, + 0.03238636627793312, + 0.007814108394086361, + 0.004811982158571482, + 0.016525916755199432, + 0.0047641415148973465, + 0.006710270885378122, + 0.012141365557909012, + 0.012951022945344448, + 0.00719383405521512, + 0.01560699287801981, + 0.005813829600811005, + 0.021477727219462395, + 0.005483763292431831, + 0.020509827882051468, + 0.05325166881084442, + 0.015350540168583393, + 0.042996011674404144, + 0.005931946914643049, + 0.0028428242076188326, + 0.010038734413683414, + 0.0036689687985926867, + 0.009100312367081642, + 0.008374212309718132, + 0.007127413991838694, + 0.005813443101942539, + 0.005862811580300331, + 0.006882533896714449, + 0.005041129887104034, + 0.004987341817468405, + 0.006756186485290527, + 0.006336824968457222, + 0.004013159312307835, + 0.0049227215349674225, + 0.0037347793113440275, + 0.0055034467950463295, + 0.020290330052375793, + 0.008976403623819351, + 0.00239352835342288, + 0.003570140339434147, + 0.00696194265037775, + 0.016893284395337105, + 0.01898237317800522, + 0.014071241021156311, + 0.01951267197728157, + 0.04567420855164528, + 0.01604059711098671, + 0.0180104561150074, + 0.034799035638570786, + 0.09629663825035095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.011082522571086884 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up triggers a slow, deliberate downward tilt, followed by a continuous shot of the action. The frame initially fills with a long, dagger-shaped icicle, its translucent crystal surface riddled with imperfections and containing tiny air bubbles, suspended from the weathered edge of a black asphalt shingle roof. Bright, low-angle winter sunlight bathes the scene, causing the icicle to glitter as it melts at its connection point. The camera meticulously captures the formation, expansion, and shedding of a droplet as the solid base visibly thins, transforming into a fragile, watery thread. As the connection inevitably weakens, a hairline crack suddenly appears at the top. With a silent, decisive snap, the icicle breaks under its own weight and separates from the roof. Pulled downward by gravity, it plummets, the camera tilting steadily downward to follow its descent against a blurred background. The icicle's fall reaches its peak as it makes contact with the frozen ground below, violently shattering into countless glittering crystal fragments and a mist of fine ice powder that catches the light, creating a brief, shimmering explosion that conveys a tranquil yet dramatic atmosphere of natural transition.", + "seed": 1339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_8/index_7372_seed_1339_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": -0.05395034030079841, + "vlm_logprob": -0.05395034030079841, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.05395034030079841, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.011505795642733574 + ], + "opposite_logprob_tokens": [ + -6.386505603790283 + ], + "logprob_reward": -0.011505795642733574, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.011505795642733574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12080173194408417, + 0.12296181172132492, + 0.06918670237064362, + 0.07111085206270218, + 0.06946177035570145, + 0.11175723373889923, + 0.10055031627416611, + 0.16875489056110382, + 0.07412371784448624, + 0.09129095077514648 + ], + "credit_map": [ + 0.007292334455996752, + 0.009095647372305393, + 0.026425693184137344, + 0.01752941682934761, + 0.011073282919824123, + 0.02615155652165413, + 0.011356432922184467, + 0.011877372860908508, + 0.01039351150393486, + 0.010889110155403614, + 0.0234055295586586, + 0.015370030887424946, + 0.021591681987047195, + 0.017323173582553864, + 0.013361998833715916, + 0.010626771487295628, + 0.005493246018886566, + 0.010949079878628254, + 0.01154408510774374, + 0.008377408608794212, + 0.011954879388213158, + 0.008988941088318825, + 0.006034651305526495, + 0.005844410974532366, + 0.006807850208133459, + 0.007919135503470898, + 0.012312258593738079, + 0.01031089760363102, + 0.005561298690736294, + 0.014913852326571941, + 0.004011140204966068, + 0.009274418465793133, + 0.006126597989350557, + 0.00657080952078104, + 0.012480550445616245, + 0.013464359566569328, + 0.0063085658475756645, + 0.012776772491633892, + 0.00552723603323102, + 0.00620688172057271, + 0.014866333454847336, + 0.013129845261573792, + 0.015627222135663033, + 0.01391657069325447, + 0.01342417486011982, + 0.015278955921530724, + 0.007084869779646397, + 0.018429266288876534, + 0.01004182081669569, + 0.013839817605912685, + 0.018029222264885902, + 0.012448679655790329, + 0.0042501697316765785, + 0.024058755487203598, + 0.008401000872254372, + 0.0094808554276824, + 0.013976114802062511, + 0.017697099596261978, + 0.020265614613890648, + 0.017930539324879646, + 0.043176885694265366, + 0.024718357250094414, + 0.014744069427251816, + 0.016246212646365166, + 0.009726352989673615, + 0.009109441190958023, + 0.010616072453558445, + 0.009874614886939526, + 0.010355694219470024, + 0.007122006267309189, + 0.009890954941511154, + 0.007428580429404974, + 0.008765614591538906, + 0.007999088615179062, + 0.012756508775055408, + 0.013391734100878239, + 0.011308998800814152, + 0.009318075142800808, + 0.0174030140042305, + 0.010347924195230007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.011505795642733574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.009805017150938511 + ], + "opposite_logprob_tokens": [ + -4.634805202484131 + ], + "logprob_reward": -0.009805017150938511, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805017150938511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0666627585887909, + 0.16272501647472382, + 0.13842900097370148, + 0.16126123070716858, + 0.06732306629419327, + 0.08004593104124069, + 0.06189688667654991, + 0.14811447262763977, + 0.06679117679595947, + 0.04675047844648361 + ], + "credit_map": [ + 0.005338122136890888, + 0.005725499242544174, + 0.013055549003183842, + 0.009732694365084171, + 0.005061035510152578, + 0.01715679094195366, + 0.002821636153385043, + 0.007771431002765894, + 0.01370764710009098, + 0.023041535168886185, + 0.028149545192718506, + 0.02144700475037098, + 0.02430061809718609, + 0.023552820086479187, + 0.01409346703439951, + 0.014432379975914955, + 0.009229395538568497, + 0.018325703218579292, + 0.02090735360980034, + 0.015223688445985317, + 0.017162110656499863, + 0.03171588480472565, + 0.009183434769511223, + 0.016681428998708725, + 0.012723559513688087, + 0.017616799101233482, + 0.02344682812690735, + 0.017184894531965256, + 0.01689571887254715, + 0.03883533552289009, + 0.010024101473391056, + 0.02453399822115898, + 0.0066906302236020565, + 0.008929211646318436, + 0.012150384485721588, + 0.010863907635211945, + 0.005804927553981543, + 0.013443391770124435, + 0.003280661068856716, + 0.006159949116408825, + 0.013764590956270695, + 0.01108392421156168, + 0.011510198935866356, + 0.009512541815638542, + 0.006520336028188467, + 0.009785670787096024, + 0.003785876091569662, + 0.014082789421081543, + 0.005610779393464327, + 0.007670664228498936, + 0.013731559738516808, + 0.009170466102659702, + 0.003937393892556429, + 0.008699010126292706, + 0.007029579486697912, + 0.006047435104846954, + 0.012090216390788555, + 0.021995238959789276, + 0.012944024056196213, + 0.014761785045266151, + 0.036665137857198715, + 0.023875301703810692, + 0.016013484448194504, + 0.009769282303750515, + 0.00649484246969223, + 0.0070021189749240875, + 0.009790604002773762, + 0.009161489084362984, + 0.01003412064164877, + 0.007350392173975706, + 0.010478995740413666, + 0.0064786141738295555, + 0.005045975558459759, + 0.004930537194013596, + 0.00638612499460578, + 0.011312881484627724, + 0.004987547639757395, + 0.003745855065062642, + 0.007333819288760424, + 0.003007733030244708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.009805017150938511 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.014193067327141762 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193067327141762, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193067327141762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0709185004234314, + 0.17071962356567383, + 0.08885643631219864, + 0.057141076773405075, + 0.06692226976156235, + 0.16248872876167297, + 0.12637878954410553, + 0.11697123199701309, + 0.06861162185668945, + 0.07099169492721558 + ], + "credit_map": [ + 0.005088360980153084, + 0.006633378565311432, + 0.010466503910720348, + 0.007696806453168392, + 0.010362712666392326, + 0.011718949303030968, + 0.007580868899822235, + 0.011370926164090633, + 0.016221465542912483, + 0.019333839416503906, + 0.034309398382902145, + 0.02063983678817749, + 0.024179650470614433, + 0.02543633058667183, + 0.013782641850411892, + 0.016816481947898865, + 0.005491856951266527, + 0.010950066149234772, + 0.0143427150323987, + 0.009224874898791313, + 0.012216171249747276, + 0.020204836502671242, + 0.005140057764947414, + 0.011285865679383278, + 0.005682252813130617, + 0.008021051064133644, + 0.009902586229145527, + 0.00789501890540123, + 0.007069842889904976, + 0.008689885027706623, + 0.004563638474792242, + 0.005316801834851503, + 0.007857274264097214, + 0.00894448533654213, + 0.01322918850928545, + 0.007755782455205917, + 0.007605012971907854, + 0.00882617849856615, + 0.00468445522710681, + 0.008019899018108845, + 0.021358352154493332, + 0.016886338591575623, + 0.019752902910113335, + 0.01996372453868389, + 0.020636649802327156, + 0.02403031475841999, + 0.013255767524242401, + 0.026604680344462395, + 0.010992459952831268, + 0.013529962860047817, + 0.02448262646794319, + 0.016917334869503975, + 0.007518773898482323, + 0.031864240765571594, + 0.010318906046450138, + 0.010754483751952648, + 0.010598604567348957, + 0.013122846372425556, + 0.01594581827521324, + 0.01389002613723278, + 0.02047772891819477, + 0.015858659520745277, + 0.010880534537136555, + 0.016197025775909424, + 0.009283504448831081, + 0.007877951487898827, + 0.007073048036545515, + 0.006678029429167509, + 0.013059831224381924, + 0.007609924301505089, + 0.008162662386894226, + 0.008866669610142708, + 0.0060071493498981, + 0.005574823822826147, + 0.01650950126349926, + 0.016489196568727493, + 0.0048818644136190414, + 0.003535061376169324, + 0.012765830382704735, + 0.005228269845247269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.014193067327141762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.22559796273708344 + ], + "opposite_logprob_tokens": [ + -1.6005979776382446 + ], + "logprob_reward": -0.22559796273708344, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559796273708344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07548772543668747, + 0.04550877958536148, + 0.04756801202893257, + 0.05162409693002701, + 0.06757276505231857, + 0.08457893133163452, + 0.14558018743991852, + 0.25197669863700867, + 0.15243519842624664, + 0.0776677057147026 + ], + "credit_map": [ + 0.004635881166905165, + 0.0066122994758188725, + 0.016524650156497955, + 0.01018829271197319, + 0.009145409800112247, + 0.013747085817158222, + 0.00789337232708931, + 0.00674072653055191, + 0.00334545667283237, + 0.009481806308031082, + 0.006475754082202911, + 0.0049744863063097, + 0.007586028892546892, + 0.005793245043605566, + 0.003536568721756339, + 0.004315429367125034, + 0.003355972468852997, + 0.007885735481977463, + 0.00827780831605196, + 0.005510010756552219, + 0.0063043930567801, + 0.008755285292863846, + 0.0032867928966879845, + 0.004192008171230555, + 0.005976138170808554, + 0.0058538350276649, + 0.007794475182890892, + 0.005000740755349398, + 0.004369099624454975, + 0.011623634025454521, + 0.0030762890819460154, + 0.007929879240691662, + 0.006920448504388332, + 0.00797323789447546, + 0.016787301748991013, + 0.008521746844053268, + 0.006697484292089939, + 0.010460144840180874, + 0.0048858895897865295, + 0.005326504819095135, + 0.01348066795617342, + 0.009867909364402294, + 0.01768544875085354, + 0.012385126203298569, + 0.004763550590723753, + 0.011389571242034435, + 0.003028740407899022, + 0.0119779072701931, + 0.01402182038873434, + 0.0174405500292778, + 0.025124186649918556, + 0.019051048904657364, + 0.009714413434267044, + 0.03272794187068939, + 0.015344414860010147, + 0.012155803851783276, + 0.025282783433794975, + 0.03698423504829407, + 0.02106078527867794, + 0.023316266015172005, + 0.0637957751750946, + 0.03373211994767189, + 0.026123905554413795, + 0.021680807694792747, + 0.019117729738354683, + 0.013570008799433708, + 0.022104421630501747, + 0.020843857899308205, + 0.018696701154112816, + 0.017904983833432198, + 0.020133662968873978, + 0.020063815638422966, + 0.006389603950083256, + 0.007002060301601887, + 0.010636751540005207, + 0.015514378435909748, + 0.008149618282914162, + 0.008397051133215427, + 0.011476974934339523, + 0.010101263411343098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.22559796273708344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.008649858646094799 + ], + "opposite_logprob_tokens": [ + -4.758649826049805 + ], + "logprob_reward": -0.008649858646094799, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649858646094799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0665305033326149, + 0.10908141732215881, + 0.14575162529945374, + 0.06553983688354492, + 0.12006524950265884, + 0.0565403513610363, + 0.13311780989170074, + 0.10345742851495743, + 0.06815213710069656, + 0.13176368176937103 + ], + "credit_map": [ + 0.003941201139241457, + 0.006860756780952215, + 0.010772963054478168, + 0.007002304773777723, + 0.0091936606913805, + 0.013924432918429375, + 0.006130647379904985, + 0.008704537525773048, + 0.008682037703692913, + 0.026098918169736862, + 0.01572931371629238, + 0.011552342213690281, + 0.015420903451740742, + 0.014734542928636074, + 0.008369384333491325, + 0.008493969216942787, + 0.011845650151371956, + 0.02117173932492733, + 0.028973830863833427, + 0.017542362213134766, + 0.013096103444695473, + 0.027800224721431732, + 0.004881351720541716, + 0.020440367981791496, + 0.005272114649415016, + 0.00739736994728446, + 0.011438833549618721, + 0.007413237821310759, + 0.003976146690547466, + 0.016579780727624893, + 0.002568176481872797, + 0.010894179344177246, + 0.021054133772850037, + 0.018756534904241562, + 0.019740168005228043, + 0.015256956219673157, + 0.0078066508285701275, + 0.01973201520740986, + 0.004869441501796246, + 0.012849352322518826, + 0.007741258013993502, + 0.006948760244995356, + 0.007747489959001541, + 0.007054679561406374, + 0.006556349340826273, + 0.008026513271033764, + 0.00401644129306078, + 0.008448858745396137, + 0.014631558209657669, + 0.0183100588619709, + 0.02540949173271656, + 0.012007676996290684, + 0.004680274520069361, + 0.03848610818386078, + 0.007419305387884378, + 0.012173336930572987, + 0.009264261461794376, + 0.01258520781993866, + 0.008827710524201393, + 0.008615713566541672, + 0.030018066987395287, + 0.015465758740901947, + 0.009380471892654896, + 0.009300234727561474, + 0.010460695251822472, + 0.008226525969803333, + 0.00858579482883215, + 0.007921641692519188, + 0.009531510062515736, + 0.006805177312344313, + 0.008355071768164635, + 0.008265716023743153, + 0.01367761567234993, + 0.012453601695597172, + 0.021833494305610657, + 0.021618515253067017, + 0.01271437294781208, + 0.013090169988572598, + 0.02089606411755085, + 0.015479846857488155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.008649858646094799 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up triggers a slow, deliberate downward tilt, followed by a continuous shot of the action. The frame initially fills with a long, dagger-shaped icicle, its translucent crystal surface riddled with imperfections and containing tiny air bubbles, suspended from the weathered edge of a black asphalt shingle roof. Bright, low-angle winter sunlight bathes the scene, causing the icicle to glitter as it melts at its connection point. The camera meticulously captures the formation, expansion, and shedding of a droplet as the solid base visibly thins, transforming into a fragile, watery thread. As the connection inevitably weakens, a hairline crack suddenly appears at the top. With a silent, decisive snap, the icicle breaks under its own weight and separates from the roof. Pulled downward by gravity, it plummets, the camera tilting steadily downward to follow its descent against a blurred background. The icicle's fall reaches its peak as it makes contact with the frozen ground below, violently shattering into countless glittering crystal fragments and a mist of fine ice powder that catches the light, creating a brief, shimmering explosion that conveys a tranquil yet dramatic atmosphere of natural transition.", + "seed": 1340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_8/index_7372_seed_1340_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": -0.03723334455862641, + "vlm_logprob": -0.03723334455862641, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.03723334455862641, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.006836713291704655 + ], + "opposite_logprob_tokens": [ + -6.131836891174316 + ], + "logprob_reward": -0.006836713291704655, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.006836713291704655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3810212016105652, + 0.0665903389453888, + 0.061705563217401505, + 0.05463742837309837, + 0.049321260303258896, + 0.07998618483543396, + 0.12314718961715698, + 0.04170424863696098, + 0.0771649032831192, + 0.06472169607877731 + ], + "credit_map": [ + 0.04006657376885414, + 0.025225721299648285, + 0.05940883234143257, + 0.01767583377659321, + 0.03869696334004402, + 0.033688776195049286, + 0.029462506994605064, + 0.1367959976196289, + 0.004466579761356115, + 0.005916989874094725, + 0.009150281548500061, + 0.011708189733326435, + 0.007165513467043638, + 0.008863430470228195, + 0.009057711809873581, + 0.01026164274662733, + 0.009869230911135674, + 0.005214207340031862, + 0.011205770075321198, + 0.003663061885163188, + 0.006977708078920841, + 0.011202013120055199, + 0.004260673187673092, + 0.009312898851931095, + 0.00836390070617199, + 0.00685522286221385, + 0.007874456234276295, + 0.00903378427028656, + 0.005245855078101158, + 0.006854122504591942, + 0.004709186963737011, + 0.005700899753719568, + 0.0065166521817445755, + 0.0036759241484105587, + 0.012300041504204273, + 0.003404376795515418, + 0.0040823533199727535, + 0.00936173927038908, + 0.005342239513993263, + 0.00463793333619833, + 0.014362775720655918, + 0.010037633590400219, + 0.010033637285232544, + 0.00901931244879961, + 0.011960850097239017, + 0.008791301399469376, + 0.008721194230020046, + 0.007059477269649506, + 0.01313190720975399, + 0.013344661332666874, + 0.017941204831004143, + 0.019324976950883865, + 0.015841957181692123, + 0.01647344045341015, + 0.015153747983276844, + 0.011935297399759293, + 0.0029731907416135073, + 0.0038844835944473743, + 0.005308930296450853, + 0.00974933709949255, + 0.0050368052907288074, + 0.004692372400313616, + 0.0059142750687897205, + 0.004144852515310049, + 0.008433159440755844, + 0.011438488960266113, + 0.00898564700037241, + 0.01321220863610506, + 0.00970531813800335, + 0.0076165152713656425, + 0.009899397380650043, + 0.007874163798987865, + 0.00895001832395792, + 0.009740099310874939, + 0.008190598338842392, + 0.009959896095097065, + 0.006251087412238121, + 0.005756911356002092, + 0.006140767131000757, + 0.009732311591506004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.006836713291704655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.005262214224785566 + ], + "opposite_logprob_tokens": [ + -5.25526237487793 + ], + "logprob_reward": -0.005262214224785566, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262214224785566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5646427273750305, + 0.037252768874168396, + 0.014635882340371609, + 0.020410872995853424, + 0.07516731321811676, + 0.04492605850100517, + 0.029420511797070503, + 0.02744094841182232, + 0.1476079225540161, + 0.0384950265288353 + ], + "credit_map": [ + 0.07439249008893967, + 0.021218016743659973, + 0.09752524644136429, + 0.013679414056241512, + 0.051352355629205704, + 0.03000386245548725, + 0.04057944938540459, + 0.2358919084072113, + 0.0024061519652605057, + 0.0031371309887617826, + 0.0057279812172055244, + 0.006159323733299971, + 0.003756545251235366, + 0.005499784369021654, + 0.00441319914534688, + 0.006152650807052851, + 0.001871926593594253, + 0.001601034076884389, + 0.002789245918393135, + 0.0010506317485123873, + 0.0015318129444494843, + 0.0028913645073771477, + 0.0008573094964958727, + 0.0020425578113645315, + 0.003061050781980157, + 0.002272589597851038, + 0.002931184135377407, + 0.0037383437156677246, + 0.0023768849205225706, + 0.002185218734666705, + 0.001894569955766201, + 0.0019510319689288735, + 0.01066343579441309, + 0.0064126597717404366, + 0.017790498211979866, + 0.007029434200376272, + 0.004713622387498617, + 0.013583041727542877, + 0.007613292429596186, + 0.007361321244388819, + 0.00833049975335598, + 0.006115894299000502, + 0.005657727364450693, + 0.004657785873860121, + 0.00663329241797328, + 0.0045443191193044186, + 0.005009440239518881, + 0.003977096173912287, + 0.003240040736272931, + 0.0029929904267191887, + 0.003999896813184023, + 0.0036422000266611576, + 0.003957508131861687, + 0.004481486976146698, + 0.004276700783520937, + 0.0028296879027038813, + 0.0021149422973394394, + 0.002231681253761053, + 0.002541978843510151, + 0.0033877319656312466, + 0.0036938756238669157, + 0.005104830022901297, + 0.004007481038570404, + 0.004358427133411169, + 0.019714223220944405, + 0.0241229347884655, + 0.018717098981142044, + 0.025094565004110336, + 0.018256284296512604, + 0.01213928870856762, + 0.01631217822432518, + 0.013251343742012978, + 0.004672197159379721, + 0.00677201384678483, + 0.004895973484963179, + 0.005940872710198164, + 0.004343946930021048, + 0.003734408877789974, + 0.004240900278091431, + 0.003894711146131158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.005262214224785566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.012535824440419674 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535824440419674, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535824440419674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21746493875980377, + 0.07019930332899094, + 0.05893907696008682, + 0.10589775443077087, + 0.07704363018274307, + 0.10086183995008469, + 0.10214000940322876, + 0.0688806027173996, + 0.14623881876468658, + 0.05233404040336609 + ], + "credit_map": [ + 0.018909385427832603, + 0.01768413372337818, + 0.03179563954472542, + 0.011770147830247879, + 0.029703648760914803, + 0.016724102199077606, + 0.01962004229426384, + 0.0712578296661377, + 0.0055263894610106945, + 0.005495324265211821, + 0.01087170373648405, + 0.009226313792169094, + 0.0074175260961055756, + 0.01212705671787262, + 0.009126998484134674, + 0.010407986119389534, + 0.007399356923997402, + 0.005542319733649492, + 0.012379376217722893, + 0.0036910492926836014, + 0.007777306716889143, + 0.01126769371330738, + 0.003612021915614605, + 0.007269947789609432, + 0.016293326392769814, + 0.0154923927038908, + 0.01516485121101141, + 0.01870373822748661, + 0.013006193563342094, + 0.009076099842786789, + 0.007511482574045658, + 0.010649662464857101, + 0.01111588440835476, + 0.007092256564646959, + 0.018775565549731255, + 0.0071599120274186134, + 0.00399106927216053, + 0.014175078831613064, + 0.008013567887246609, + 0.006720284465700388, + 0.013307242654263973, + 0.008774494752287865, + 0.01827060990035534, + 0.01060278620570898, + 0.013483994640409946, + 0.012989901937544346, + 0.010409059002995491, + 0.013023740611970425, + 0.011298096738755703, + 0.013868400827050209, + 0.015264428220689297, + 0.013627597130835056, + 0.012800497934222221, + 0.01134037971496582, + 0.014289339073002338, + 0.00965126696974039, + 0.0056354752741754055, + 0.005699644796550274, + 0.008124937303364277, + 0.01438194140791893, + 0.007238512858748436, + 0.009049189276993275, + 0.009749997407197952, + 0.009000897407531738, + 0.01493197400122881, + 0.023843031376600266, + 0.01709965243935585, + 0.024839580059051514, + 0.01684906706213951, + 0.012554805725812912, + 0.01641971617937088, + 0.01970098167657852, + 0.006466164253652096, + 0.007307685911655426, + 0.006453270558267832, + 0.00754378829151392, + 0.006287950091063976, + 0.00535202119499445, + 0.006572300568223, + 0.006350853014737368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.012535824440419674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.1603495180606842 + ], + "opposite_logprob_tokens": [ + -1.9103494882583618 + ], + "logprob_reward": -0.1603495180606842, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603495180606842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5597274899482727, + 0.021175744011998177, + 0.011411336250603199, + 0.014307096600532532, + 0.011702779680490494, + 0.035513751208782196, + 0.044736482203006744, + 0.042464062571525574, + 0.22437964379787445, + 0.03458164259791374 + ], + "credit_map": [ + 0.07786683738231659, + 0.0204144399613142, + 0.09696628898382187, + 0.013545790687203407, + 0.05205298960208893, + 0.028991427272558212, + 0.03802947327494621, + 0.2318602055311203, + 0.0017278456361964345, + 0.0017505427822470665, + 0.0032312441617250443, + 0.0040336488746106625, + 0.0020815394818782806, + 0.0026155959349125624, + 0.0027451328933238983, + 0.002990195294842124, + 0.001903628115542233, + 0.0011703416239470243, + 0.0022294153459370136, + 0.0009732982725836337, + 0.0009014690876938403, + 0.0019156222697347403, + 0.0005521038547158241, + 0.0017654566327109933, + 0.0021230396814644337, + 0.0018460976425558329, + 0.002356884302571416, + 0.002163110300898552, + 0.0011928215390071273, + 0.0018655371386557817, + 0.0009741167887113988, + 0.0017854881007224321, + 0.0014549823245033622, + 0.0011077870149165392, + 0.0023181233555078506, + 0.0011194893158972263, + 0.0007623511482961476, + 0.001812369329854846, + 0.0015214961022138596, + 0.001606180565431714, + 0.006557456683367491, + 0.0036779826041311026, + 0.0047906325198709965, + 0.0034502174239605665, + 0.004950322210788727, + 0.0036125967744737864, + 0.004796529188752174, + 0.0036780096124857664, + 0.00461342791095376, + 0.004629024304449558, + 0.00614151731133461, + 0.00624559773132205, + 0.005755885038524866, + 0.007126787677407265, + 0.0060247983783483505, + 0.004199441988021135, + 0.003141269087791443, + 0.0033248420804739, + 0.003534699557349086, + 0.003995765931904316, + 0.006724982056766748, + 0.00871041789650917, + 0.0052301911637187, + 0.007801893632858992, + 0.02916327677667141, + 0.03881320357322693, + 0.029467353597283363, + 0.029085557907819748, + 0.028920074924826622, + 0.017694013193249702, + 0.026398882269859314, + 0.024837253615260124, + 0.003579151351004839, + 0.005701461341232061, + 0.004726087208837271, + 0.006132494658231735, + 0.0033763626124709845, + 0.0037333595100790262, + 0.004830435384064913, + 0.0025022870395332575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.1603495180606842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.0011824527755379677 + ], + "opposite_logprob_tokens": [ + -6.751182556152344 + ], + "logprob_reward": -0.0011824527755379677, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011824527755379677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6113124489784241, + 0.03695747256278992, + 0.030968694016337395, + 0.029127908870577812, + 0.06213352829217911, + 0.06457625329494476, + 0.03111460618674755, + 0.02448131889104843, + 0.05244959890842438, + 0.05687814950942993 + ], + "credit_map": [ + 0.08811543881893158, + 0.024426819756627083, + 0.10444251447916031, + 0.015799151733517647, + 0.05812894552946091, + 0.03188532590866089, + 0.03982510790228844, + 0.2486891746520996, + 0.003633225802332163, + 0.0036812841426581144, + 0.006693210918456316, + 0.0028133902233093977, + 0.003878365270793438, + 0.005777104292064905, + 0.0033596991561353207, + 0.007121192757040262, + 0.005012902431190014, + 0.0021947151981294155, + 0.0074776276014745235, + 0.001881531672552228, + 0.0022616933565586805, + 0.006100943312048912, + 0.0014339637709781528, + 0.004605316556990147, + 0.005169871728867292, + 0.0027880400884896517, + 0.00499446876347065, + 0.005068529862910509, + 0.00339451408945024, + 0.0025833561085164547, + 0.0021845283918082714, + 0.0029445996042340994, + 0.009088433347642422, + 0.0036709194537252188, + 0.014916871674358845, + 0.0037737267557531595, + 0.004390375688672066, + 0.013207442127168179, + 0.006386078428477049, + 0.006699684076011181, + 0.009897471405565739, + 0.00564596988260746, + 0.008349704556167126, + 0.003894202411174774, + 0.007657366339117289, + 0.0074846819043159485, + 0.011074791662395, + 0.01057206280529499, + 0.003863791236653924, + 0.002372441114857793, + 0.003469773568212986, + 0.005054709501564503, + 0.003939403221011162, + 0.0037688598968088627, + 0.004142136313021183, + 0.004503493197262287, + 0.0016507442342117429, + 0.002729359082877636, + 0.002509060548618436, + 0.003947170916944742, + 0.003371750470250845, + 0.00332858064211905, + 0.003835615934804082, + 0.0031090364791452885, + 0.007622481789439917, + 0.011447915807366371, + 0.007046144921332598, + 0.006889806594699621, + 0.006609261967241764, + 0.003912964835762978, + 0.004597397055476904, + 0.0043236264027655125, + 0.007692528888583183, + 0.006696886382997036, + 0.007408962119370699, + 0.007704013492912054, + 0.006830012891441584, + 0.006220074836164713, + 0.008333954028785229, + 0.005991719663143158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0011824527755379677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up triggers a slow, deliberate downward tilt, followed by a continuous shot of the action. The frame initially fills with a long, dagger-shaped icicle, its translucent crystal surface riddled with imperfections and containing tiny air bubbles, suspended from the weathered edge of a black asphalt shingle roof. Bright, low-angle winter sunlight bathes the scene, causing the icicle to glitter as it melts at its connection point. The camera meticulously captures the formation, expansion, and shedding of a droplet as the solid base visibly thins, transforming into a fragile, watery thread. As the connection inevitably weakens, a hairline crack suddenly appears at the top. With a silent, decisive snap, the icicle breaks under its own weight and separates from the roof. Pulled downward by gravity, it plummets, the camera tilting steadily downward to follow its descent against a blurred background. The icicle's fall reaches its peak as it makes contact with the frozen ground below, violently shattering into countless glittering crystal fragments and a mist of fine ice powder that catches the light, creating a brief, shimmering explosion that conveys a tranquil yet dramatic atmosphere of natural transition.", + "seed": 1341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_8/index_7372_seed_1341_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": -0.010047242883592844, + "vlm_logprob": -0.010047242883592844, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.010047242883592844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.006593733094632626 + ], + "opposite_logprob_tokens": [ + -6.006593704223633 + ], + "logprob_reward": -0.006593733094632626, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.006593733094632626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2984740436077118, + 0.0983695238828659, + 0.08588019758462906, + 0.13259288668632507, + 0.10226810723543167, + 0.06823328882455826, + 0.0970596969127655, + 0.03931064158678055, + 0.03322761505842209, + 0.04458395764231682 + ], + "credit_map": [ + 0.02358580380678177, + 0.02993033640086651, + 0.07852283120155334, + 0.020700285211205482, + 0.042306073009967804, + 0.026836112141609192, + 0.024208655580878258, + 0.05238394811749458, + 0.010566195473074913, + 0.009694754146039486, + 0.019069854170084, + 0.009699570015072823, + 0.01014009676873684, + 0.01726313680410385, + 0.007743311580270529, + 0.014192605391144753, + 0.026083197444677353, + 0.008399112150073051, + 0.013219255954027176, + 0.005485711619257927, + 0.0048424312844872475, + 0.011841916479170322, + 0.004189606290310621, + 0.011818962171673775, + 0.018215367570519447, + 0.007504391949623823, + 0.017314089462161064, + 0.011376023292541504, + 0.00885673426091671, + 0.021991051733493805, + 0.014606636017560959, + 0.03272860124707222, + 0.00781989749521017, + 0.009971492923796177, + 0.023747911676764488, + 0.009162123315036297, + 0.011742052622139454, + 0.017394281923770905, + 0.00807918980717659, + 0.014351154677569866, + 0.008261507377028465, + 0.0077154007740318775, + 0.011213433928787708, + 0.006406691391021013, + 0.004446601495146751, + 0.00752690713852644, + 0.0120646171271801, + 0.01059812679886818, + 0.010619508102536201, + 0.0065615675412118435, + 0.011021541431546211, + 0.0058561768382787704, + 0.00921198632568121, + 0.01540757343173027, + 0.015312288887798786, + 0.023069048300385475, + 0.006918675731867552, + 0.003030097112059593, + 0.005792534444481134, + 0.004121818114072084, + 0.005409634672105312, + 0.005676377564668655, + 0.003313349327072501, + 0.005048155784606934, + 0.0049798693507909775, + 0.0033500159624964, + 0.0032528622541576624, + 0.004511725623160601, + 0.0043588802218437195, + 0.0046389298513531685, + 0.004801445174962282, + 0.0033338863868266344, + 0.00766601599752903, + 0.005724281072616577, + 0.005980039481073618, + 0.005484539084136486, + 0.0033624127972871065, + 0.0043236929923295975, + 0.006098459474742413, + 0.005944520700722933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.006593733094632626 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.0052674319595098495 + ], + "opposite_logprob_tokens": [ + -5.25526762008667 + ], + "logprob_reward": -0.0052674319595098495, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052674319595098495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1474725306034088, + 0.12770424783229828, + 0.03471726179122925, + 0.10823003202676773, + 0.09514045715332031, + 0.09069301187992096, + 0.12546253204345703, + 0.1683998703956604, + 0.06850417703390121, + 0.03367588669061661 + ], + "credit_map": [ + 0.010144303552806377, + 0.008897966705262661, + 0.048206377774477005, + 0.009549641981720924, + 0.013734959997236729, + 0.025756048038601875, + 0.010257619433104992, + 0.020925622433423996, + 0.014069228433072567, + 0.012335858307778835, + 0.01673983410000801, + 0.010430990718305111, + 0.01938054896891117, + 0.023468559607863426, + 0.011355476453900337, + 0.019923744723200798, + 0.004518544767051935, + 0.0035743259359151125, + 0.0058821989223361015, + 0.002907564165070653, + 0.003371027996763587, + 0.006325855385512114, + 0.002623752923682332, + 0.0055139921605587006, + 0.01179557479918003, + 0.010667352937161922, + 0.017355987802147865, + 0.012289351783692837, + 0.009268992580473423, + 0.015339666046202183, + 0.017732864245772362, + 0.013780241832137108, + 0.006035033613443375, + 0.012180543504655361, + 0.017864329740405083, + 0.011172005906701088, + 0.016124693676829338, + 0.013109017163515091, + 0.008407930843532085, + 0.01024690829217434, + 0.007358565926551819, + 0.012717962265014648, + 0.010571488179266453, + 0.011460534296929836, + 0.013414799235761166, + 0.0083340834826231, + 0.013131257146596909, + 0.013704318553209305, + 0.009855088777840137, + 0.009501899592578411, + 0.013457038439810276, + 0.01757441647350788, + 0.011171991005539894, + 0.015932118520140648, + 0.024685226380825043, + 0.023284757509827614, + 0.012251758947968483, + 0.009626329876482487, + 0.016529081389307976, + 0.019655639305710793, + 0.021576320752501488, + 0.02508510835468769, + 0.024089237675070763, + 0.039586395025253296, + 0.009930060245096684, + 0.010313116945326328, + 0.006674725096672773, + 0.007210629992187023, + 0.01148981787264347, + 0.008704980835318565, + 0.009275903925299644, + 0.004904939327389002, + 0.0023323202040046453, + 0.0038250822108238935, + 0.004877904430031776, + 0.0042730835266411304, + 0.003252072026953101, + 0.003162238746881485, + 0.003941211383789778, + 0.008011973462998867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0052674319595098495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.020579271018505096 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579271018505096, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579271018505096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16453662514686584, + 0.09312208741903305, + 0.09515116363763809, + 0.10091708600521088, + 0.15766046941280365, + 0.07200603187084198, + 0.1946839988231659, + 0.05776407569646835, + 0.03419719636440277, + 0.029961280524730682 + ], + "credit_map": [ + 0.011883633211255074, + 0.012336928397417068, + 0.03565635159611702, + 0.013205070048570633, + 0.019669031724333763, + 0.016911398619413376, + 0.01937396079301834, + 0.03550024330615997, + 0.009299255907535553, + 0.007150065153837204, + 0.01623498648405075, + 0.006772398948669434, + 0.009520807303488255, + 0.022759273648262024, + 0.008095686323940754, + 0.013289612717926502, + 0.0183170884847641, + 0.010144413448870182, + 0.016751661896705627, + 0.007725108880549669, + 0.00663317134603858, + 0.013997502624988556, + 0.007264010142534971, + 0.014318210072815418, + 0.01110469363629818, + 0.00916988868266344, + 0.011332235299050808, + 0.011222820729017258, + 0.00927501730620861, + 0.015273328870534897, + 0.015478153713047504, + 0.018060950562357903, + 0.010139232501387596, + 0.01931544579565525, + 0.025072365999221802, + 0.024272389709949493, + 0.023324038833379745, + 0.02170783467590809, + 0.016803810372948647, + 0.017025349661707878, + 0.00871158204972744, + 0.0068658581003546715, + 0.009483338333666325, + 0.007942181080579758, + 0.007435273844748735, + 0.011062338016927242, + 0.007900327444076538, + 0.012605133466422558, + 0.012859579175710678, + 0.013869820162653923, + 0.019608967006206512, + 0.016567008569836617, + 0.021705862134695053, + 0.029343131929636, + 0.04059159383177757, + 0.04013804346323013, + 0.0042570424266159534, + 0.005667713470757008, + 0.005981290712952614, + 0.00678642513230443, + 0.00844762846827507, + 0.008553672581911087, + 0.00948558934032917, + 0.008584711700677872, + 0.006711588241159916, + 0.0037337441463023424, + 0.0035590073093771935, + 0.004859959706664085, + 0.0038393905851989985, + 0.004293393809348345, + 0.004236896056681871, + 0.0029632176738232374, + 0.0028739029075950384, + 0.004159250762313604, + 0.003228423884138465, + 0.003107869764789939, + 0.00569649925455451, + 0.004060556646436453, + 0.0033024500589817762, + 0.0035323279444128275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.020579271018505096 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.01253676600754261 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253676600754261, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253676600754261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06308427453041077, + 0.0566224530339241, + 0.047165192663669586, + 0.16997578740119934, + 0.14340034127235413, + 0.07269322872161865, + 0.10637936741113663, + 0.1456497758626938, + 0.09481935948133469, + 0.10021030902862549 + ], + "credit_map": [ + 0.004863737616688013, + 0.004868695978075266, + 0.013526537455618382, + 0.004994743037968874, + 0.005607639439404011, + 0.00786514114588499, + 0.006366854067891836, + 0.014990923926234245, + 0.007847732864320278, + 0.006408138200640678, + 0.010111085139214993, + 0.0037159800995141268, + 0.005670331418514252, + 0.008584869094192982, + 0.004639763385057449, + 0.009644553996622562, + 0.008926838636398315, + 0.005029888357967138, + 0.007648462895303965, + 0.003273520153015852, + 0.0034852272365242243, + 0.009304427541792393, + 0.002331489697098732, + 0.007165336050093174, + 0.020777031779289246, + 0.014039588160812855, + 0.02202020213007927, + 0.01500874012708664, + 0.01947135478258133, + 0.03382697328925133, + 0.022555114701390266, + 0.022276781499385834, + 0.009934703819453716, + 0.02009749971330166, + 0.01892741024494171, + 0.02233949862420559, + 0.022742239758372307, + 0.017847411334514618, + 0.017147690057754517, + 0.014363872818648815, + 0.005934460554271936, + 0.005365521647036076, + 0.01301892101764679, + 0.006848112214356661, + 0.004715657792985439, + 0.008648937568068504, + 0.013942367397248745, + 0.014219244942069054, + 0.009318644180893898, + 0.012143660336732864, + 0.013060021214187145, + 0.010008213110268116, + 0.011016003787517548, + 0.014090734533965588, + 0.02103850618004799, + 0.015703581273555756, + 0.011418996378779411, + 0.013059855438768864, + 0.011837716214358807, + 0.014056148007512093, + 0.019551200792193413, + 0.02138395607471466, + 0.02537321113049984, + 0.028968695551156998, + 0.009383642114698887, + 0.010509639978408813, + 0.010618476197123528, + 0.01331179030239582, + 0.013910026289522648, + 0.014449164271354675, + 0.015400664880871773, + 0.007235950790345669, + 0.01002733875066042, + 0.009499412029981613, + 0.014997336082160473, + 0.014805126003921032, + 0.008563869632780552, + 0.00918564386665821, + 0.011733945459127426, + 0.021397631615400314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.01253676600754261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 long, dagger-shaped icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the icicle translucent with visible imperfections and tiny air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate downward tilt to follow the icicle's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle break with a snap and shatter into fragments and ice powder upon hitting the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred during the camera's 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": [ + -0.005259012337774038 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005259012337774038, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259012337774038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17574238777160645, + 0.06893704831600189, + 0.0745399221777916, + 0.08386225253343582, + 0.18988054990768433, + 0.14399375021457672, + 0.12433553487062454, + 0.056946855038404465, + 0.053371477872133255, + 0.028390273451805115 + ], + "credit_map": [ + 0.013322493061423302, + 0.016731400042772293, + 0.04432885721325874, + 0.012937515042722225, + 0.024324720725417137, + 0.024348096922039986, + 0.017915448173880577, + 0.021833859384059906, + 0.007486716844141483, + 0.006323494017124176, + 0.01220284029841423, + 0.0047186692245304585, + 0.008485434576869011, + 0.013019123114645481, + 0.007637611590325832, + 0.00906315352767706, + 0.012322651222348213, + 0.0077303024008870125, + 0.011846845969557762, + 0.005257714539766312, + 0.00490514375269413, + 0.013974757865071297, + 0.004605347756296396, + 0.01389715913683176, + 0.010756200179457664, + 0.007120655383914709, + 0.01236707717180252, + 0.008984622545540333, + 0.00677085854113102, + 0.01349714957177639, + 0.010856046341359615, + 0.01350963395088911, + 0.014669929631054401, + 0.017856866121292114, + 0.03611772879958153, + 0.023233652114868164, + 0.023488003760576248, + 0.02785961516201496, + 0.01769077219069004, + 0.028963977470993996, + 0.01949124038219452, + 0.011987577192485332, + 0.032710980623960495, + 0.008391928859055042, + 0.010135355405509472, + 0.020825829356908798, + 0.014743129722774029, + 0.025707701221108437, + 0.012614566832780838, + 0.009758309461176395, + 0.016895929351449013, + 0.0123196542263031, + 0.01255708932876587, + 0.01625032350420952, + 0.021953780204057693, + 0.021985884755849838, + 0.007397704757750034, + 0.007542929146438837, + 0.008012649603188038, + 0.005442112218588591, + 0.006614411249756813, + 0.009157828986644745, + 0.005173712968826294, + 0.007605508901178837, + 0.0074821044690907, + 0.005076158791780472, + 0.00786157138645649, + 0.011833792552351952, + 0.005030746106058359, + 0.004558518063277006, + 0.0071337223052978516, + 0.004394858609884977, + 0.0029285235796123743, + 0.002427345374599099, + 0.0044762613251805305, + 0.004557221196591854, + 0.001707764808088541, + 0.002706330269575119, + 0.003040623152628541, + 0.0065462058410048485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.005259012337774038 + } + ], + "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/0000125/group_9/index_tensor([3804, 3804])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_9/index_tensor([3804, 3804])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..adaeece99bec813272b631d83e1eb2eb9a19c400 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_9/index_tensor([3804, 3804])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle close-up slowly pans right, following the majestic drift of a massive jellyfish filling the foreground, its form evoked by the graceful, sinuous lines of Art Nouveau. The creature's massive, translucent bell is a deep, luminous crimson, adorned with intricate, swirling gold patterns that capture and refract the light within. From beneath the bell, a mane of countless hair-like tentacles emerges, each radiating a vibrant yellow bioluminescence, creating a mesmerizing, flowing mass. The creature traverses the indigo-blue depths, where currents are imagined as stylized, sweeping curves of water. Its body pulsates gently and rhythmically, a slow, deliberate contraction and expansion that propels it with serene grace as its finest tentacles delicately twitch, rubbing against invisible, shimmering particles in the water. The scene is illuminated almost entirely by the creature itself, creating a dramatic interplay of warm light and deep shadows that immerses the image in an atmosphere of hypnotic, otherworldly beauty and profound tranquility.", + "seed": 20121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_9/index_3804_seed_20121_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": -0.024293097527697684, + "vlm_logprob": -0.024293097527697684, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.024293097527697684, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049191150814294815 + ], + "opposite_logprob_tokens": [ + -3.2991912364959717 + ], + "logprob_reward": -0.049191150814294815, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.049191150814294815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047871820628643036, + 0.04925555735826492, + 0.058935534209012985, + 0.05339434742927551, + 0.04074874147772789, + 0.4471561908721924, + 0.052777186036109924, + 0.03725556284189224, + 0.04606105759739876, + 0.16654393076896667 + ], + "credit_map": [ + 0.005828897003084421, + 0.005630582105368376, + 0.007139099296182394, + 0.004090068396180868, + 0.006818248424679041, + 0.006054259371012449, + 0.006189087405800819, + 0.006121575366705656, + 0.005211119074374437, + 0.007640541065484285, + 0.00686784228309989, + 0.007658299524337053, + 0.007513280026614666, + 0.0038163368590176105, + 0.00591938616707921, + 0.004628753289580345, + 0.006172884255647659, + 0.005584228318184614, + 0.007325868122279644, + 0.010123377665877342, + 0.008659900166094303, + 0.006230264902114868, + 0.00963500328361988, + 0.005204006098210812, + 0.003925594035536051, + 0.005852357484400272, + 0.0059848870150744915, + 0.0077323163859546185, + 0.0051612453535199165, + 0.011388090439140797, + 0.0067473421804606915, + 0.0066025154665112495, + 0.005724909249693155, + 0.0035746037028729916, + 0.005919851362705231, + 0.007717191707342863, + 0.005210947711020708, + 0.004424263257533312, + 0.004584518726915121, + 0.003592453896999359, + 0.018947789445519447, + 0.058503977954387665, + 0.01116254087537527, + 0.1416887789964676, + 0.04176291078329086, + 0.02694852091372013, + 0.03946209326386452, + 0.10867960751056671, + 0.004920638632029295, + 0.00667935935780406, + 0.0036363324616104364, + 0.014304360374808311, + 0.0040599629282951355, + 0.0034225599374622107, + 0.012091200798749924, + 0.0036627703811973333, + 0.0038862074725329876, + 0.00374488509260118, + 0.00695269787684083, + 0.0063603282906115055, + 0.003547609318047762, + 0.003663780866190791, + 0.003398723667487502, + 0.005701330490410328, + 0.0042710090056061745, + 0.004801608622074127, + 0.003309372579678893, + 0.005259939935058355, + 0.0033298959024250507, + 0.007236042525619268, + 0.007537404075264931, + 0.010315782390534878, + 0.016103176400065422, + 0.01577860862016678, + 0.012536212801933289, + 0.023405766114592552, + 0.016133880242705345, + 0.02530842088162899, + 0.030017169192433357, + 0.02726069651544094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.049191150814294815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009283285471610725 + ], + "opposite_logprob_tokens": [ + -7.0009284019470215 + ], + "logprob_reward": -0.0009283285471610725, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009283285471610725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08082526177167892, + 0.07302147150039673, + 0.0498884953558445, + 0.034436069428920746, + 0.05159740149974823, + 0.32648780941963196, + 0.057725366204977036, + 0.04095413535833359, + 0.053005050867795944, + 0.23205892741680145 + ], + "credit_map": [ + 0.009139541536569595, + 0.010475022718310356, + 0.008122829720377922, + 0.00916216243058443, + 0.011553340591490269, + 0.009575366973876953, + 0.013276887126266956, + 0.009520108811557293, + 0.008109355345368385, + 0.012791704386472702, + 0.008683551102876663, + 0.007771327160298824, + 0.009590884670615196, + 0.00809730589389801, + 0.008838263340294361, + 0.009139081463217735, + 0.006220006849616766, + 0.006477884016931057, + 0.0088052311912179, + 0.005479338113218546, + 0.006244319956749678, + 0.004748809617012739, + 0.007336630020290613, + 0.004576276522129774, + 0.0032452517189085484, + 0.0035405305679887533, + 0.002940686419606209, + 0.0024488596245646477, + 0.003944079391658306, + 0.008108546957373619, + 0.0058538345620036125, + 0.004354280885308981, + 0.004681589547544718, + 0.005927043966948986, + 0.004587553441524506, + 0.0038355274591594934, + 0.011829436756670475, + 0.008373158983886242, + 0.005916974972933531, + 0.006446114741265774, + 0.013379302807152271, + 0.045507628470659256, + 0.006359214894473553, + 0.10673876851797104, + 0.029701964929699898, + 0.019262826070189476, + 0.027413496747612953, + 0.07812460511922836, + 0.0023951008915901184, + 0.006217018235474825, + 0.0024066967889666557, + 0.02077433280646801, + 0.004386477172374725, + 0.0036363499239087105, + 0.014277084730565548, + 0.0036323019303381443, + 0.004267971962690353, + 0.005300906952470541, + 0.00375927472487092, + 0.00568412384018302, + 0.005557653959840536, + 0.004312379751354456, + 0.005475570913404226, + 0.006596254650503397, + 0.004517256282269955, + 0.006045141723006964, + 0.0052452352829277515, + 0.009088114835321903, + 0.004260603338479996, + 0.006736015900969505, + 0.007037736009806395, + 0.010074945166707039, + 0.02455487847328186, + 0.021933982148766518, + 0.017057577148079872, + 0.0365249440073967, + 0.022290457040071487, + 0.033383190631866455, + 0.045949384570121765, + 0.030364491045475006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.0009283285471610725 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006379238329827785 + ], + "opposite_logprob_tokens": [ + -7.375638008117676 + ], + "logprob_reward": -0.0006379238329827785, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006379238329827785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15300823748111725, + 0.20791181921958923, + 0.04728922247886658, + 0.10225366055965424, + 0.05980169400572777, + 0.049500443041324615, + 0.09845064580440521, + 0.05898245424032211, + 0.08268745243549347, + 0.14011436700820923 + ], + "credit_map": [ + 0.01701365038752556, + 0.01961386203765869, + 0.018318064510822296, + 0.02476947009563446, + 0.01663530059158802, + 0.01729585975408554, + 0.022500446066260338, + 0.016861578449606895, + 0.02447589486837387, + 0.03132366016507149, + 0.03056316450238228, + 0.033977851271629333, + 0.026111464947462082, + 0.016262879595160484, + 0.02415008842945099, + 0.021046804264187813, + 0.00492115831002593, + 0.00514717772603035, + 0.006989101879298687, + 0.004106197040528059, + 0.006160292308777571, + 0.006570326630026102, + 0.007475199643522501, + 0.0059197647497057915, + 0.010246175341308117, + 0.00791826844215393, + 0.009620518423616886, + 0.005820890888571739, + 0.019496874883770943, + 0.023150457069277763, + 0.015749113634228706, + 0.010251365602016449, + 0.008086957968771458, + 0.005851010326296091, + 0.007714718580245972, + 0.009752380661666393, + 0.008021289482712746, + 0.008515508845448494, + 0.00593408290296793, + 0.005925741977989674, + 0.004339784849435091, + 0.006635566242039204, + 0.00396127812564373, + 0.006290184799581766, + 0.006619137711822987, + 0.00431307777762413, + 0.008258060552179813, + 0.009083352982997894, + 0.0035873143933713436, + 0.011149543337523937, + 0.0028522938955575228, + 0.036914337426424026, + 0.006424313876777887, + 0.0060519324615597725, + 0.025718146935105324, + 0.005752761848270893, + 0.005896293092519045, + 0.006077544763684273, + 0.007186777889728546, + 0.0077453418634831905, + 0.005102456081658602, + 0.007901116274297237, + 0.007835173048079014, + 0.011237746104598045, + 0.007185411173850298, + 0.008957182057201862, + 0.008984587155282497, + 0.012006527744233608, + 0.008868885226547718, + 0.011261727660894394, + 0.008584832772612572, + 0.016838299110531807, + 0.014919478446245193, + 0.015165970660746098, + 0.011421012692153454, + 0.017991812899708748, + 0.013838628306984901, + 0.018969789147377014, + 0.021060431376099586, + 0.026747260242700577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0006379238329827785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007175016799010336 + ], + "opposite_logprob_tokens": [ + -7.250717639923096 + ], + "logprob_reward": -0.0007175016799010336, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007175016799010336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05414123833179474, + 0.04677090421319008, + 0.04399297758936882, + 0.04885278269648552, + 0.04616497457027435, + 0.2636679410934448, + 0.030973393470048904, + 0.04130284860730171, + 0.08532039821147919, + 0.33881258964538574 + ], + "credit_map": [ + 0.006614469923079014, + 0.007844004780054092, + 0.009135629050433636, + 0.006307055242359638, + 0.007202274166047573, + 0.0049467491917312145, + 0.00637746462598443, + 0.00571358622983098, + 0.006254373583942652, + 0.00685527827590704, + 0.008477668277919292, + 0.005111358128488064, + 0.0069783651269972324, + 0.0035002671647816896, + 0.005322335287928581, + 0.0042712572030723095, + 0.00526286568492651, + 0.0050825439393520355, + 0.007967638783156872, + 0.005540193058550358, + 0.00596381351351738, + 0.004363551270216703, + 0.0052461177110672, + 0.004566254559904337, + 0.004063611850142479, + 0.003373592160642147, + 0.005115939769893885, + 0.004770144820213318, + 0.0066014728508889675, + 0.012852886691689491, + 0.005471271928399801, + 0.006603860761970282, + 0.009429693222045898, + 0.004369611851871014, + 0.0066922674886882305, + 0.00599337462335825, + 0.0067537520080804825, + 0.0036028302274644375, + 0.004765513818711042, + 0.004557930864393711, + 0.011894134804606438, + 0.03512588143348694, + 0.00602132361382246, + 0.0831008031964302, + 0.02497989498078823, + 0.01622667722404003, + 0.023237669840455055, + 0.06308155506849289, + 0.0025416866410523653, + 0.004483665805310011, + 0.001965878764167428, + 0.007660045754164457, + 0.0037980997003614902, + 0.0033289494458585978, + 0.004118948243558407, + 0.0030761188827455044, + 0.0035144411958754063, + 0.004978504031896591, + 0.004430845845490694, + 0.004943036939948797, + 0.004414708353579044, + 0.005993609316647053, + 0.004219096153974533, + 0.00880860723555088, + 0.007588672451674938, + 0.007867489010095596, + 0.008200244978070259, + 0.013339008204638958, + 0.007224549539387226, + 0.012362205423414707, + 0.01168874278664589, + 0.017049480229616165, + 0.027842527255415916, + 0.028829272836446762, + 0.02540609799325466, + 0.04864763095974922, + 0.030782775953412056, + 0.05355142429471016, + 0.06531613320112228, + 0.058436717838048935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.0007175016799010336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999058276414871 + ], + "opposite_logprob_tokens": [ + -2.694990634918213 + ], + "logprob_reward": -0.06999058276414871, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999058276414871, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054567139595746994, + 0.06492529064416885, + 0.019371848553419113, + 0.062198977917432785, + 0.09010376036167145, + 0.29553672671318054, + 0.059890955686569214, + 0.17255298793315887, + 0.025738632306456566, + 0.15511362254619598 + ], + "credit_map": [ + 0.006927184294909239, + 0.006916698534041643, + 0.008293847553431988, + 0.00494942395016551, + 0.0062422300688922405, + 0.007655134424567223, + 0.007161120884120464, + 0.00642150454223156, + 0.007406098302453756, + 0.0069606611505150795, + 0.007659225724637508, + 0.004961247555911541, + 0.007751057390123606, + 0.00660167820751667, + 0.009848249144852161, + 0.013737081550061703, + 0.0018196535529568791, + 0.0015945498598739505, + 0.002103549661114812, + 0.0030593015253543854, + 0.002962564118206501, + 0.0032446125987917185, + 0.0026543664280325174, + 0.0019332518568262458, + 0.006144488696008921, + 0.005675409454852343, + 0.006195229012519121, + 0.004108656197786331, + 0.010632702149450779, + 0.014188253320753574, + 0.008048716001212597, + 0.007205531001091003, + 0.012309656478464603, + 0.008256308734416962, + 0.012906200252473354, + 0.011226640082895756, + 0.015829896554350853, + 0.009063675999641418, + 0.011317305266857147, + 0.009194078855216503, + 0.011461695656180382, + 0.04288367182016373, + 0.006215957459062338, + 0.09781460464000702, + 0.027097702026367188, + 0.01719294674694538, + 0.02339940331876278, + 0.06947074830532074, + 0.0056701903231441975, + 0.007728285621851683, + 0.004925196524709463, + 0.011849180795252323, + 0.005669933278113604, + 0.009182367473840714, + 0.008027707226574421, + 0.006838098168373108, + 0.01956707425415516, + 0.019567379727959633, + 0.022184746339917183, + 0.023710504174232483, + 0.01950065977871418, + 0.020898109301924706, + 0.01880018226802349, + 0.02832433581352234, + 0.0019100072095170617, + 0.002518974244594574, + 0.002130651380866766, + 0.004484462086111307, + 0.001780515885911882, + 0.0036944299936294556, + 0.0039346893317997456, + 0.005284903571009636, + 0.015235453844070435, + 0.012900127097964287, + 0.012583885341882706, + 0.02720453403890133, + 0.013202553614974022, + 0.020113540813326836, + 0.029294949024915695, + 0.024578595533967018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.06999058276414871 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up slowly pans right, following the majestic drift of a massive jellyfish filling the foreground, its form evoked by the graceful, sinuous lines of Art Nouveau. The creature's massive, translucent bell is a deep, luminous crimson, adorned with intricate, swirling gold patterns that capture and refract the light within. From beneath the bell, a mane of countless hair-like tentacles emerges, each radiating a vibrant yellow bioluminescence, creating a mesmerizing, flowing mass. The creature traverses the indigo-blue depths, where currents are imagined as stylized, sweeping curves of water. Its body pulsates gently and rhythmically, a slow, deliberate contraction and expansion that propels it with serene grace as its finest tentacles delicately twitch, rubbing against invisible, shimmering particles in the water. The scene is illuminated almost entirely by the creature itself, creating a dramatic interplay of warm light and deep shadows that immerses the image in an atmosphere of hypnotic, otherworldly beauty and profound tranquility.", + "seed": 20122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_9/index_3804_seed_20122_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": -0.039349909324664624, + "vlm_logprob": -0.039349909324664624, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.039349909324664624, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12264542281627655 + ], + "opposite_logprob_tokens": [ + -2.247645378112793 + ], + "logprob_reward": -0.12264542281627655, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12264542281627655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036684758961200714, + 0.05258945748209953, + 0.03250177204608917, + 0.10607625544071198, + 0.015152769163250923, + 0.029330648481845856, + 0.016677388921380043, + 0.5947833061218262, + 0.10341594368219376, + 0.01278770063072443 + ], + "credit_map": [ + 0.004516515415161848, + 0.00782005675137043, + 0.0020335728768259287, + 0.004889516159892082, + 0.005033210851252079, + 0.0029030090663582087, + 0.004787653684616089, + 0.004701224621385336, + 0.0042521231807768345, + 0.009592496789991856, + 0.00372942048124969, + 0.006146154832094908, + 0.009914956986904144, + 0.004129326436668634, + 0.01010227482765913, + 0.004722704645246267, + 0.004127838183194399, + 0.004535371903330088, + 0.002296966966241598, + 0.005376903805881739, + 0.0038964664563536644, + 0.0042696925811469555, + 0.003433776553720236, + 0.004564756993204355, + 0.011247307062149048, + 0.020971374586224556, + 0.012051761150360107, + 0.014199095778167248, + 0.016265928745269775, + 0.009002171456813812, + 0.013247121125459671, + 0.009091491810977459, + 0.0014859194634482265, + 0.0020979621913284063, + 0.0018099832814186811, + 0.0020881120581179857, + 0.0030056550167500973, + 0.0014122016727924347, + 0.0017624517204239964, + 0.001490483758971095, + 0.0028912166599184275, + 0.004114797804504633, + 0.0024918271228671074, + 0.0030000335536897182, + 0.0035598224494606256, + 0.005623024422675371, + 0.003315581940114498, + 0.004334342200309038, + 0.0016339024296030402, + 0.003135772654786706, + 0.0011770332930609584, + 0.0022366323973983526, + 0.0019509298726916313, + 0.0025905740913003683, + 0.0021224808879196644, + 0.0018300639931112528, + 0.1644441783428192, + 0.06421352922916412, + 0.028179556131362915, + 0.020308278501033783, + 0.16972358524799347, + 0.037278082221746445, + 0.05822080373764038, + 0.05241529643535614, + 0.03261378034949303, + 0.008383027277886868, + 0.005893413443118334, + 0.005118323024362326, + 0.007881475612521172, + 0.005422156769782305, + 0.021183589473366737, + 0.016920180991292, + 0.001838108990341425, + 0.0011247402289882302, + 0.0007947373669594526, + 0.0011575475800782442, + 0.0011139159323647618, + 0.0011415805201977491, + 0.004481000360101461, + 0.0011360698845237494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.12264542281627655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024963426403701305 + ], + "opposite_logprob_tokens": [ + -6.002496242523193 + ], + "logprob_reward": -0.0024963426403701305, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024963426403701305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06111656501889229, + 0.28682741522789, + 0.046596601605415344, + 0.03001941367983818, + 0.06124715879559517, + 0.03972223028540611, + 0.023458346724510193, + 0.37245362997055054, + 0.035743094980716705, + 0.04281552508473396 + ], + "credit_map": [ + 0.006845836993306875, + 0.009049452841281891, + 0.004150673281401396, + 0.009184177964925766, + 0.007956510409712791, + 0.005481670144945383, + 0.009609086439013481, + 0.008839157409965992, + 0.023287365213036537, + 0.06002321466803551, + 0.017344767227768898, + 0.03786639869213104, + 0.05305956304073334, + 0.02128603495657444, + 0.055565815418958664, + 0.018394289538264275, + 0.005532792303711176, + 0.005684537813067436, + 0.005549195688217878, + 0.006375283934175968, + 0.00480542192235589, + 0.006265327334403992, + 0.005651059560477734, + 0.006732986308634281, + 0.003643547184765339, + 0.003966592252254486, + 0.0029907410498708487, + 0.0037313937209546566, + 0.0037608693819493055, + 0.0029289203230291605, + 0.005119798704981804, + 0.003877555252984166, + 0.004705395083874464, + 0.00924170296639204, + 0.005016091279685497, + 0.007507867179811001, + 0.009676392190158367, + 0.006112736649811268, + 0.012005915865302086, + 0.006981066428124905, + 0.005137985106557608, + 0.007911975495517254, + 0.003599715419113636, + 0.004140438046306372, + 0.0059578996151685715, + 0.004334727767854929, + 0.0040983580984175205, + 0.004541132599115372, + 0.0020399163477122784, + 0.003865770995616913, + 0.0025139469653367996, + 0.002716962480917573, + 0.0030905120074748993, + 0.003612304339185357, + 0.003309079445898533, + 0.002309854608029127, + 0.10345375537872314, + 0.040023963898420334, + 0.017173385247588158, + 0.012530586682260036, + 0.106007881462574, + 0.023770106956362724, + 0.03678715229034424, + 0.03270679712295532, + 0.011721485294401646, + 0.0027445382438600063, + 0.00251045823097229, + 0.0019237976521253586, + 0.003130261553451419, + 0.002111945068463683, + 0.002548019867390394, + 0.00905259046703577, + 0.005246676038950682, + 0.007249015849083662, + 0.001839235657826066, + 0.003880714764818549, + 0.005455328617244959, + 0.003860101103782654, + 0.007230491377413273, + 0.008053965866565704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.0024963426403701305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005654881824739277 + ], + "opposite_logprob_tokens": [ + -7.500565528869629 + ], + "logprob_reward": -0.0005654881824739277, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005654881824739277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06453536450862885, + 0.09665913134813309, + 0.07162782549858093, + 0.14057089388370514, + 0.039974749088287354, + 0.21881745755672455, + 0.07266660779714584, + 0.0874389261007309, + 0.11032222956418991, + 0.09738675504922867 + ], + "credit_map": [ + 0.008169733919203281, + 0.007868134416639805, + 0.0073344106785953045, + 0.011291923001408577, + 0.006495386827737093, + 0.005934490356594324, + 0.010072678327560425, + 0.007368606980890036, + 0.009850804693996906, + 0.019897736608982086, + 0.00446596834808588, + 0.0113753080368042, + 0.017103927209973335, + 0.00894222967326641, + 0.016638237982988358, + 0.008384916931390762, + 0.00828962679952383, + 0.011410142295062542, + 0.005219189915806055, + 0.01057465560734272, + 0.009104610420763493, + 0.010465441271662712, + 0.007639078423380852, + 0.008925080299377441, + 0.015542158856987953, + 0.020315635949373245, + 0.014189517125487328, + 0.018093816936016083, + 0.02319035306572914, + 0.012705771252512932, + 0.017456429079174995, + 0.01907721534371376, + 0.004523819778114557, + 0.006512750405818224, + 0.003971146419644356, + 0.005869160871952772, + 0.005818106234073639, + 0.004756696987897158, + 0.004861472174525261, + 0.0036616013385355473, + 0.024784762412309647, + 0.050955306738615036, + 0.021054837852716446, + 0.02239721454679966, + 0.03138429671525955, + 0.038349736481904984, + 0.01555162388831377, + 0.014339683577418327, + 0.0062772007659077644, + 0.01221546158194542, + 0.006411066744476557, + 0.009859508834779263, + 0.009223180823028088, + 0.009751709178090096, + 0.011991755105555058, + 0.006936726160347462, + 0.017698148265480995, + 0.011623620055615902, + 0.00555279478430748, + 0.012241964228451252, + 0.016850020736455917, + 0.008416240103542805, + 0.007917488925158978, + 0.007138650864362717, + 0.015723854303359985, + 0.007526575122028589, + 0.005359533708542585, + 0.008111348375678062, + 0.012901350855827332, + 0.008311710320413113, + 0.04214831069111824, + 0.010239542461931705, + 0.014981329441070557, + 0.010456448420882225, + 0.007325889077037573, + 0.012382810935378075, + 0.009145611897110939, + 0.008919650688767433, + 0.026888472959399223, + 0.007286543026566505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.0005654881824739277 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010417040903121233 + ], + "opposite_logprob_tokens": [ + -6.876041889190674 + ], + "logprob_reward": -0.0010417040903121233, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010417040903121233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24512879550457, + 0.01726469025015831, + 0.018130188807845116, + 0.029301365837454796, + 0.01385531760752201, + 0.03454775735735893, + 0.07959800213575363, + 0.4824742376804352, + 0.04679306969046593, + 0.03290654718875885 + ], + "credit_map": [ + 0.02500223182141781, + 0.027315057814121246, + 0.022713245823979378, + 0.040896013379096985, + 0.025010885670781136, + 0.01913774572312832, + 0.05288323387503624, + 0.03217039629817009, + 0.0018047738121822476, + 0.0023392231669276953, + 0.001776886056177318, + 0.0020813243463635445, + 0.002822545822709799, + 0.0017411543522030115, + 0.0032109939493238926, + 0.0014877881621941924, + 0.0020990711636841297, + 0.002207817742601037, + 0.0015901584411039948, + 0.003890945576131344, + 0.0018372302874922752, + 0.0022702461574226618, + 0.002284587826579809, + 0.0019501320784911513, + 0.003725481452420354, + 0.006498855538666248, + 0.0025044006761163473, + 0.003877058392390609, + 0.0056341043673455715, + 0.0024519385769963264, + 0.0026613168884068727, + 0.001948212506249547, + 0.0016003464115783572, + 0.0024483974557369947, + 0.001976622035726905, + 0.0017485481221228838, + 0.0022153782192617655, + 0.0012064784532412887, + 0.0012882662704214454, + 0.0013712812215089798, + 0.004137292504310608, + 0.0062931003049016, + 0.0025511186104267836, + 0.0039815595373511314, + 0.004760646726936102, + 0.0066844262182712555, + 0.003254397539421916, + 0.0028852135874330997, + 0.01100220438092947, + 0.010198057629168034, + 0.005690551828593016, + 0.007924280129373074, + 0.008031865581870079, + 0.011640235781669617, + 0.010811435990035534, + 0.014299373142421246, + 0.13248059153556824, + 0.05190102010965347, + 0.022434616461396217, + 0.016946297138929367, + 0.1383761167526245, + 0.030311426147818565, + 0.047716137021780014, + 0.04230804368853569, + 0.010757901705801487, + 0.0031852228567004204, + 0.0024081352166831493, + 0.0031508805695921183, + 0.004680685233324766, + 0.0032584320288151503, + 0.016455145552754402, + 0.0028966625686734915, + 0.0036621352192014456, + 0.0025632234755903482, + 0.0013846565270796418, + 0.0024596008006483316, + 0.0027734176255762577, + 0.0020908869337290525, + 0.010413389652967453, + 0.007559236604720354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0010417040903121233 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000058889389038 + ], + "opposite_logprob_tokens": [ + -2.695000648498535 + ], + "logprob_reward": -0.07000058889389038, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000058889389038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02685103937983513, + 0.042056769132614136, + 0.013573450036346912, + 0.03994359076023102, + 0.009414147585630417, + 0.026416786015033722, + 0.027673844248056412, + 0.7245360016822815, + 0.03395172581076622, + 0.05558261275291443 + ], + "credit_map": [ + 0.0032302674371749163, + 0.00430545536801219, + 0.0022038740571588278, + 0.003692161524668336, + 0.004359123297035694, + 0.0019316787365823984, + 0.0038997766096144915, + 0.003228703746572137, + 0.0035220321733504534, + 0.007644800003618002, + 0.003385294694453478, + 0.004797142464667559, + 0.007250582333654165, + 0.0034758250694721937, + 0.007852725684642792, + 0.004128369502723217, + 0.0017410871805623174, + 0.0015752374893054366, + 0.0013911692658439279, + 0.0018629737896844745, + 0.0016971685690805316, + 0.0017397667979821563, + 0.0017022137762978673, + 0.0018638320034369826, + 0.0037116368766874075, + 0.008303796872496605, + 0.0029615922830998898, + 0.005072678439319134, + 0.0072416518814861774, + 0.0035772754345089197, + 0.004958652891218662, + 0.004116306081414223, + 0.0010005353251472116, + 0.0011464864946901798, + 0.0012893312377855182, + 0.0012245397083461285, + 0.0014245858183130622, + 0.001480646664276719, + 0.0008688081870786846, + 0.0009792145574465394, + 0.002818081993609667, + 0.0037843878380954266, + 0.0029094626661390066, + 0.0026800069026649, + 0.0033687197137624025, + 0.004914749413728714, + 0.0027880175039172173, + 0.0031533576548099518, + 0.0038555923383682966, + 0.0047760517336428165, + 0.002456626622006297, + 0.003603516612201929, + 0.0030739805661141872, + 0.0032981117255985737, + 0.0026008272543549538, + 0.004009134601801634, + 0.19983650743961334, + 0.07716318964958191, + 0.033882852643728256, + 0.02469693496823311, + 0.20728890597820282, + 0.04587164148688316, + 0.07233728468418121, + 0.06345877796411514, + 0.0074518341571092606, + 0.00236079515889287, + 0.0020410758443176746, + 0.0026823573280125856, + 0.002752576256170869, + 0.0025214666966348886, + 0.011290444061160088, + 0.0028511786367744207, + 0.007793190889060497, + 0.0036009184550493956, + 0.0030823159031569958, + 0.004676579497754574, + 0.005070296116173267, + 0.004615205340087414, + 0.020269427448511124, + 0.006474677473306656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -0.07000058889389038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up slowly pans right, following the majestic drift of a massive jellyfish filling the foreground, its form evoked by the graceful, sinuous lines of Art Nouveau. The creature's massive, translucent bell is a deep, luminous crimson, adorned with intricate, swirling gold patterns that capture and refract the light within. From beneath the bell, a mane of countless hair-like tentacles emerges, each radiating a vibrant yellow bioluminescence, creating a mesmerizing, flowing mass. The creature traverses the indigo-blue depths, where currents are imagined as stylized, sweeping curves of water. Its body pulsates gently and rhythmically, a slow, deliberate contraction and expansion that propels it with serene grace as its finest tentacles delicately twitch, rubbing against invisible, shimmering particles in the water. The scene is illuminated almost entirely by the creature itself, creating a dramatic interplay of warm light and deep shadows that immerses the image in an atmosphere of hypnotic, otherworldly beauty and profound tranquility.", + "seed": 20123, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_9/index_3804_seed_20123_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": -0.02890470078564249, + "vlm_logprob": -0.02890470078564249, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.02890470078564249, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0533316545188427 + ], + "opposite_logprob_tokens": [ + -3.0533316135406494 + ], + "logprob_reward": -0.0533316545188427, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0533316545188427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024421388283371925, + 0.017460161820054054, + 0.08811087906360626, + 0.02555990219116211, + 0.024850759655237198, + 0.023939575999975204, + 0.7183776497840881, + 0.03669792786240578, + 0.009149577468633652, + 0.03143223747611046 + ], + "credit_map": [ + 0.0018487237393856049, + 0.002599089639261365, + 0.0020482155960053205, + 0.00529147544875741, + 0.00310489721596241, + 0.0025199358351528645, + 0.0033823237754404545, + 0.003626723773777485, + 0.0015488985227420926, + 0.002603048225864768, + 0.0015548106748610735, + 0.002358496654778719, + 0.0022124196402728558, + 0.002190805971622467, + 0.002194203669205308, + 0.0027974757831543684, + 0.007016638293862343, + 0.003596986411139369, + 0.0040488033555448055, + 0.037139806896448135, + 0.005039367824792862, + 0.003973361570388079, + 0.022551335394382477, + 0.004744562320411205, + 0.001818341203033924, + 0.0026765624061226845, + 0.0023190558422356844, + 0.0035816265735775232, + 0.005776524543762207, + 0.002722240285947919, + 0.0040680947713553905, + 0.00259745423682034, + 0.003468260867521167, + 0.0040826047770679, + 0.002273352351039648, + 0.00301340501755476, + 0.0035684884060174227, + 0.002732135122641921, + 0.003661434631794691, + 0.0020510756876319647, + 0.0025759737472981215, + 0.004299481399357319, + 0.0025095290038734674, + 0.003184392349794507, + 0.003378171008080244, + 0.0022783020976930857, + 0.0025969119742512703, + 0.003116811625659466, + 0.022357195615768433, + 0.1045285314321518, + 0.03036319836974144, + 0.03901620954275131, + 0.2189463973045349, + 0.03810787945985794, + 0.22409182786941528, + 0.04096636921167374, + 0.0038259103894233704, + 0.0028962413780391216, + 0.0035276017151772976, + 0.006396255921572447, + 0.002923363819718361, + 0.0027870475314557552, + 0.00966779887676239, + 0.0046737040393054485, + 0.0012408833717927337, + 0.0012100409949198365, + 0.0010077835759148002, + 0.0013138025533407927, + 0.0010089268907904625, + 0.0010430325055494905, + 0.0013116337358951569, + 0.0010134720942005515, + 0.002557532861828804, + 0.010112074203789234, + 0.004023526329547167, + 0.0030209047254174948, + 0.006018882617354393, + 0.001562069053761661, + 0.0028140805661678314, + 0.0013231666525825858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.0533316545188427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011841197265312076 + ], + "opposite_logprob_tokens": [ + -6.751183986663818 + ], + "logprob_reward": -0.0011841197265312076, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011841197265312076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0076637864112854, + 0.004441330675035715, + 0.012451479211449623, + 0.003684160765260458, + 0.0034032573457807302, + 0.003384519601240754, + 0.918304979801178, + 0.040631700307130814, + 0.00448644021525979, + 0.0015483289025723934 + ], + "credit_map": [ + 0.0007535043987445533, + 0.0007371809333562851, + 0.0004620141990017146, + 0.0010717858094722033, + 0.0016556112095713615, + 0.0008840410737320781, + 0.0011497455416247249, + 0.0009499029256403446, + 0.000528458971530199, + 0.0003632641746662557, + 0.00021512442617677152, + 0.0005966854514554143, + 0.00041512364987283945, + 0.0004938288475386798, + 0.0012821038253605366, + 0.0005467412993311882, + 0.001084318500943482, + 0.0007159821107052267, + 0.0008385560940951109, + 0.004628239665180445, + 0.0010035204468294978, + 0.0008349384879693389, + 0.0025852660182863474, + 0.0007606578292325139, + 0.000251934165135026, + 0.0004246545722708106, + 0.0003626600082498044, + 0.0005632407264783978, + 0.0005654800916090608, + 0.00040863739559426904, + 0.000699374417308718, + 0.0004081794759258628, + 0.0002660898317117244, + 0.0005567158223129809, + 0.0003037681744899601, + 0.0004489783605094999, + 0.0006251474260352552, + 0.0003492858959361911, + 0.0005706899100914598, + 0.0002825816045515239, + 0.0003139020991511643, + 0.0005094187799841166, + 0.0003224831889383495, + 0.00032637984259054065, + 0.0007669816259294748, + 0.0004058009944856167, + 0.0004374150012154132, + 0.0003021378943230957, + 0.02834480069577694, + 0.13334999978542328, + 0.03937792778015137, + 0.050035685300827026, + 0.27990978956222534, + 0.04857010394334793, + 0.28596562147140503, + 0.052751049399375916, + 0.0038409035187214613, + 0.003215018194168806, + 0.0037132028955966234, + 0.007072728127241135, + 0.0031652268953621387, + 0.0031447976361960173, + 0.011331441812217236, + 0.005148380529135466, + 0.0004995083436369896, + 0.00055589439580217, + 0.0005417750217020512, + 0.000767274759709835, + 0.0005561645957641304, + 0.000406258855946362, + 0.0006150391418486834, + 0.0005445248680189252, + 0.00021055979595985264, + 0.00016355188563466072, + 0.0002702547935768962, + 0.0002112757647410035, + 0.00017943058628588915, + 0.0001450796116841957, + 0.0002264046634081751, + 0.00014177167031448334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.0011841197265312076 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004400238685775548 + ], + "opposite_logprob_tokens": [ + -7.7504401206970215 + ], + "logprob_reward": -0.0004400238685775548, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004400238685775548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010306041687726974, + 0.00700409198179841, + 0.027405356988310814, + 0.01434608269482851, + 0.005485893227159977, + 0.01156004797667265, + 0.8864191174507141, + 0.00467862980440259, + 0.0080875214189291, + 0.024707220494747162 + ], + "credit_map": [ + 0.001261911471374333, + 0.0009737422806210816, + 0.0009474862017668784, + 0.001432010787539184, + 0.001656023901887238, + 0.0010061453795060515, + 0.0012142810737714171, + 0.0018144407076761127, + 0.0007842068444006145, + 0.0006357615347951651, + 0.0008853111648932099, + 0.001234612544067204, + 0.0006503271288238466, + 0.000857509090565145, + 0.001107294810935855, + 0.0008490686886943877, + 0.002091373084113002, + 0.0011124394368380308, + 0.0013658879324793816, + 0.011216682381927967, + 0.0018114325357601047, + 0.0015327901346608996, + 0.006636112928390503, + 0.0016386376228183508, + 0.0009925210615620017, + 0.0012922900496050715, + 0.001949281431734562, + 0.001879386487416923, + 0.0027351537719368935, + 0.0013761252630501986, + 0.0024975903797894716, + 0.0016237335512414575, + 0.0007248737965710461, + 0.000737883907277137, + 0.000654767092783004, + 0.0007093093008734286, + 0.0007846357184462249, + 0.0006449724896810949, + 0.0006846271571703255, + 0.0005448234733194113, + 0.0011625925544649363, + 0.001977750798687339, + 0.0013644096907228231, + 0.0012356283841654658, + 0.00206812284886837, + 0.0010247351601719856, + 0.0011889527086168528, + 0.0015378554817289114, + 0.027148881927132607, + 0.12851345539093018, + 0.03657012805342674, + 0.0476946197450161, + 0.2706945240497589, + 0.04611536115407944, + 0.27894115447998047, + 0.050740983337163925, + 0.0005299526965245605, + 0.0007100108196027577, + 0.00048394716577604413, + 0.000619606114923954, + 0.00066179916029796, + 0.0005016743671149015, + 0.0005909489118494093, + 0.0005806906265206635, + 0.0008981294231489301, + 0.0008684393251314759, + 0.0011394880712032318, + 0.0014087276067584753, + 0.0007545157568529248, + 0.0007729397621005774, + 0.0011222213506698608, + 0.0011230604723095894, + 0.0016821817262098193, + 0.008282788097858429, + 0.0033163258340209723, + 0.001956768101081252, + 0.005221622530370951, + 0.0012190729612484574, + 0.0019946438260376453, + 0.0010338175343349576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.0004400238685775548 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005596501869149506 + ], + "opposite_logprob_tokens": [ + -7.5005598068237305 + ], + "logprob_reward": -0.0005596501869149506, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005596501869149506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01779906265437603, + 0.036612994968891144, + 0.020549502223730087, + 0.0070130187086761, + 0.005959841422736645, + 0.0036850194446742535, + 0.8797785639762878, + 0.01084528025239706, + 0.007985834963619709, + 0.009770845994353294 + ], + "credit_map": [ + 0.0030708538834005594, + 0.0020698460284620523, + 0.0014665117487311363, + 0.002601421670988202, + 0.0021000767592340708, + 0.0029995162039995193, + 0.0017475273925811052, + 0.0017433111788704991, + 0.00424998952075839, + 0.0036074421368539333, + 0.002033836441114545, + 0.0039914436638355255, + 0.003242124104872346, + 0.003905058139935136, + 0.011372673325240612, + 0.004210432525724173, + 0.0017376726027578115, + 0.0009010524954646826, + 0.0008434736519120634, + 0.008497083559632301, + 0.0014946026494726539, + 0.0011298558674752712, + 0.004718019161373377, + 0.001227742526680231, + 0.0004105320549570024, + 0.0007312989328056574, + 0.0007457992760464549, + 0.0010007277596741915, + 0.0011589368805289268, + 0.0007932072039693594, + 0.001434655743651092, + 0.0007378618465736508, + 0.0005150313372723758, + 0.0013727799523621798, + 0.000421675737015903, + 0.0006260701338760555, + 0.0010777295101433992, + 0.0006973501876927912, + 0.0008553075022064149, + 0.0003938973823096603, + 0.00032021949300542474, + 0.0005090996855869889, + 0.0003657635534182191, + 0.0004746493068523705, + 0.0007884891238063574, + 0.0003667757846415043, + 0.00046043310430832207, + 0.0003995898296125233, + 0.027062298730015755, + 0.12798233330249786, + 0.036994680762290955, + 0.047369327396154404, + 0.26874420046806335, + 0.046090271323919296, + 0.2753720283508301, + 0.050163526087999344, + 0.0009988130768761039, + 0.0011065442813560367, + 0.0009060252923518419, + 0.0020419289357960224, + 0.0010101940715685487, + 0.0010315560502931476, + 0.0023832402657717466, + 0.0013669787440449, + 0.000902866420801729, + 0.001041313516907394, + 0.0009627625695429742, + 0.0013375615235418081, + 0.0008186097256839275, + 0.0007914498564787209, + 0.0011466183932498097, + 0.0009846537141129375, + 0.0008670884999446571, + 0.002092367969453335, + 0.0011427052086219192, + 0.0009628594270907342, + 0.0016568974824622273, + 0.0007808844675309956, + 0.0015763677656650543, + 0.0006916762213222682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.0005596501869149506 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900805562734604 + ], + "opposite_logprob_tokens": [ + -2.464008092880249 + ], + "logprob_reward": -0.08900805562734604, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900805562734604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010011599399149418, + 0.011531798169016838, + 0.0028288676403462887, + 0.006832712795585394, + 0.0026165200397372246, + 0.00447078188881278, + 0.9456363916397095, + 0.004493876360356808, + 0.00394790293648839, + 0.00762952258810401 + ], + "credit_map": [ + 0.0010271824430674314, + 0.001545304199680686, + 0.0009309331653639674, + 0.001943202456459403, + 0.0014446560526266694, + 0.001009989995509386, + 0.0009933297988027334, + 0.0011170010548084974, + 0.0014041629619896412, + 0.0012720099184662104, + 0.00047321515739895403, + 0.0011788165429607034, + 0.0010354416444897652, + 0.0011389185674488544, + 0.0036303207743912935, + 0.0013989126309752464, + 0.00030366124701686203, + 0.00026934529887512326, + 0.0002633718540892005, + 0.0005652737454511225, + 0.0003015057009179145, + 0.0003699205117300153, + 0.00041412649443373084, + 0.00034166278783231974, + 0.00045728246914222836, + 0.0005013738991692662, + 0.0011458146618679166, + 0.0008196730632334948, + 0.0015299094375222921, + 0.0005564858438447118, + 0.0011072049383074045, + 0.0007149688899517059, + 0.0002461835974827409, + 0.0006362955900840461, + 0.00032391370041295886, + 0.0003270594752393663, + 0.0003902291355188936, + 0.00021809505415149033, + 0.0002906927256844938, + 0.0001840507611632347, + 0.0004645663721021265, + 0.0008318913751281798, + 0.0004961329977959394, + 0.0004931282601319253, + 0.0007338526775129139, + 0.0003270854940637946, + 0.00043163125519640744, + 0.0006924932822585106, + 0.029273366555571556, + 0.1361447125673294, + 0.03897635266184807, + 0.05159248411655426, + 0.28759288787841797, + 0.049526114016771317, + 0.29824721813201904, + 0.054283298552036285, + 0.00046872979146428406, + 0.00044329409138299525, + 0.00040976813761517406, + 0.0008124162559397519, + 0.00040373794035986066, + 0.00036452393396757543, + 0.0010733120143413544, + 0.0005180941079743207, + 0.0007855218136683106, + 0.000561641703825444, + 0.00042689256952144206, + 0.0003788158646784723, + 0.0004288078343961388, + 0.0003500002494547516, + 0.0005008109728805721, + 0.0005154121317900717, + 0.00048144353786483407, + 0.0026351362466812134, + 0.001185717643238604, + 0.0006068124203011394, + 0.0015384008875116706, + 0.0003414735838305205, + 0.0004928644630126655, + 0.0003476732817944139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.08900805562734604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up slowly pans right, following the majestic drift of a massive jellyfish filling the foreground, its form evoked by the graceful, sinuous lines of Art Nouveau. The creature's massive, translucent bell is a deep, luminous crimson, adorned with intricate, swirling gold patterns that capture and refract the light within. From beneath the bell, a mane of countless hair-like tentacles emerges, each radiating a vibrant yellow bioluminescence, creating a mesmerizing, flowing mass. The creature traverses the indigo-blue depths, where currents are imagined as stylized, sweeping curves of water. Its body pulsates gently and rhythmically, a slow, deliberate contraction and expansion that propels it with serene grace as its finest tentacles delicately twitch, rubbing against invisible, shimmering particles in the water. The scene is illuminated almost entirely by the creature itself, creating a dramatic interplay of warm light and deep shadows that immerses the image in an atmosphere of hypnotic, otherworldly beauty and profound tranquility.", + "seed": 20124, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_9/index_3804_seed_20124_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": -0.028182701126206668, + "vlm_logprob": -0.028182701126206668, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.028182701126206668, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.098756343126297 + ], + "opposite_logprob_tokens": [ + -2.4737563133239746 + ], + "logprob_reward": -0.098756343126297, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.098756343126297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07924064993858337, + 0.06660179048776627, + 0.04256949573755264, + 0.02401742897927761, + 0.08741141855716705, + 0.09385815262794495, + 0.411249577999115, + 0.06822790205478668, + 0.04155564680695534, + 0.08526791632175446 + ], + "credit_map": [ + 0.00846609566360712, + 0.008500363677740097, + 0.012170104309916496, + 0.010734178125858307, + 0.008167438209056854, + 0.008189433254301548, + 0.011099547147750854, + 0.01191349234431982, + 0.008332377299666405, + 0.009865113534033298, + 0.008290363475680351, + 0.008196771144866943, + 0.005110498517751694, + 0.009519880637526512, + 0.008190702646970749, + 0.009096082299947739, + 0.006790514569729567, + 0.003281377488747239, + 0.0052930680103600025, + 0.004490479826927185, + 0.0033830103930085897, + 0.006093928124755621, + 0.004905304871499538, + 0.008331816643476486, + 0.002920323284342885, + 0.002929758047685027, + 0.0031133994925767183, + 0.004061670042574406, + 0.0018579037860035896, + 0.002460119314491749, + 0.0032059724908322096, + 0.003468282986432314, + 0.012498517520725727, + 0.007440425921231508, + 0.00745545094832778, + 0.009795756079256535, + 0.010108188726007938, + 0.01126693282276392, + 0.01744046062231064, + 0.01140568871051073, + 0.011379694566130638, + 0.011649532243609428, + 0.02016703598201275, + 0.011171591468155384, + 0.008301221765577793, + 0.009096172638237476, + 0.01061511691659689, + 0.01147778332233429, + 0.025681467726826668, + 0.021536752581596375, + 0.05071355774998665, + 0.1319422721862793, + 0.013036902993917465, + 0.038184165954589844, + 0.02195233665406704, + 0.10820212215185165, + 0.006060936488211155, + 0.007040240801870823, + 0.006135309115052223, + 0.014221454970538616, + 0.006479822099208832, + 0.0076866899617016315, + 0.008374334312975407, + 0.012229111045598984, + 0.0053158835507929325, + 0.005041119176894426, + 0.005321042612195015, + 0.004403292201459408, + 0.004519916605204344, + 0.0050320313312113285, + 0.006032832898199558, + 0.005889528896659613, + 0.006327611859887838, + 0.009792931377887726, + 0.007670288905501366, + 0.021888243034482002, + 0.006488401908427477, + 0.009966710582375526, + 0.011245026253163815, + 0.011888701468706131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.098756343126297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017212113598361611 + ], + "opposite_logprob_tokens": [ + -6.376721382141113 + ], + "logprob_reward": -0.0017212113598361611, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017212113598361611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10279291868209839, + 0.05714604631066322, + 0.03611891716718674, + 0.03447987511754036, + 0.1619209200143814, + 0.05634789168834686, + 0.1105722188949585, + 0.22908805310726166, + 0.054726116359233856, + 0.15680699050426483 + ], + "credit_map": [ + 0.011109261773526669, + 0.008912965655326843, + 0.014356506057083607, + 0.01312547829002142, + 0.010065208189189434, + 0.01094811037182808, + 0.018177716061472893, + 0.016097672283649445, + 0.006673041731119156, + 0.004964154213666916, + 0.009752998128533363, + 0.006877724081277847, + 0.005998446140438318, + 0.008040194399654865, + 0.006957828067243099, + 0.007881656289100647, + 0.005716173909604549, + 0.004093598108738661, + 0.003774196607992053, + 0.004484891891479492, + 0.003602738492190838, + 0.003960488364100456, + 0.005381430499255657, + 0.005105395335704088, + 0.004007062409073114, + 0.006106697488576174, + 0.003943181596696377, + 0.004492681007832289, + 0.0036908271722495556, + 0.0035207169130444527, + 0.004226148594170809, + 0.004492559935897589, + 0.026405436918139458, + 0.013831703923642635, + 0.018870418891310692, + 0.018512796610593796, + 0.014281345531344414, + 0.02147267758846283, + 0.024735940620303154, + 0.023810608312487602, + 0.008565573953092098, + 0.00780511274933815, + 0.007284481078386307, + 0.0061265770345926285, + 0.006240738555788994, + 0.006065043155103922, + 0.007726565003395081, + 0.006533802021294832, + 0.007074369583278894, + 0.006894747260957956, + 0.013893242925405502, + 0.03165721520781517, + 0.00898378249257803, + 0.009541455656290054, + 0.008370772935450077, + 0.02415662817656994, + 0.019018765538930893, + 0.02956329844892025, + 0.018805984407663345, + 0.06828542798757553, + 0.009686822071671486, + 0.015258077532052994, + 0.021236272528767586, + 0.04723340645432472, + 0.004979780409485102, + 0.006553850136697292, + 0.006252674385905266, + 0.007111836224794388, + 0.006657250691205263, + 0.006115682888776064, + 0.008803844451904297, + 0.008251198567450047, + 0.012161573395133018, + 0.021807443350553513, + 0.012966233305633068, + 0.045824624598026276, + 0.007673174608498812, + 0.014557676389813423, + 0.014307359233498573, + 0.02750890888273716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.0017212113598361611 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017174031818285584 + ], + "opposite_logprob_tokens": [ + -6.376717567443848 + ], + "logprob_reward": -0.0017174031818285584, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017174031818285584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02668486163020134, + 0.02281632274389267, + 0.022233057767152786, + 0.020670656114816666, + 0.2019639015197754, + 0.09093184024095535, + 0.13692674040794373, + 0.35945454239845276, + 0.015007292851805687, + 0.10331085324287415 + ], + "credit_map": [ + 0.0027827986050397158, + 0.0028375042602419853, + 0.004480724688619375, + 0.0033893242944031954, + 0.0018885624594986439, + 0.0032199211418628693, + 0.003239415353164077, + 0.004846605472266674, + 0.0026175447273999453, + 0.002818909240886569, + 0.0025415215641260147, + 0.0027963067404925823, + 0.0023643793538212776, + 0.0037617969792336226, + 0.0026195701211690903, + 0.003296288661658764, + 0.0024194729048758745, + 0.002471483778208494, + 0.0032244815956801176, + 0.0036680635530501604, + 0.0018739810911938548, + 0.0022174431942403316, + 0.003276517614722252, + 0.00308160949498415, + 0.0030990722589194775, + 0.0018152330303564668, + 0.0018020856659859419, + 0.0023660333827137947, + 0.001600481802597642, + 0.0028777578845620155, + 0.0029193300288170576, + 0.004190658684819937, + 0.031565722078084946, + 0.016902200877666473, + 0.013309158384799957, + 0.01841789111495018, + 0.018215538933873177, + 0.03398766741156578, + 0.02936496213078499, + 0.04020073637366295, + 0.010960069485008717, + 0.010578849352896214, + 0.011897070333361626, + 0.009251962415874004, + 0.01125164981931448, + 0.011424258351325989, + 0.013369596563279629, + 0.012198368087410927, + 0.01020018570125103, + 0.009373443201184273, + 0.015884391963481903, + 0.03826242312788963, + 0.010424351319670677, + 0.011591324582695961, + 0.012397675774991512, + 0.02879292704164982, + 0.029758943244814873, + 0.04261985421180725, + 0.030294742435216904, + 0.1045137420296669, + 0.01572839543223381, + 0.024779653176665306, + 0.035598721355199814, + 0.07616045325994492, + 0.001792863360606134, + 0.001995150465518236, + 0.0020716143772006035, + 0.0016870062099769711, + 0.002004646696150303, + 0.0013053640723228455, + 0.002636733464896679, + 0.0015139123424887657, + 0.0111298942938447, + 0.014069312252104282, + 0.015965821221470833, + 0.021501433104276657, + 0.010109633207321167, + 0.007074829190969467, + 0.012055963277816772, + 0.01140394527465105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.0017174031818285584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006337541271932423 + ], + "opposite_logprob_tokens": [ + -7.375633716583252 + ], + "logprob_reward": -0.0006337541271932423, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006337541271932423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0875631719827652, + 0.06820279359817505, + 0.07786450535058975, + 0.045197490602731705, + 0.05061903968453407, + 0.05043262988328934, + 0.159242644906044, + 0.2229447364807129, + 0.02347829006612301, + 0.21445471048355103 + ], + "credit_map": [ + 0.006299215368926525, + 0.006371479015797377, + 0.008699458092451096, + 0.008324774913489819, + 0.00529630109667778, + 0.014741769060492516, + 0.01183333806693554, + 0.025996824726462364, + 0.009581132791936398, + 0.010118464007973671, + 0.007661975454539061, + 0.007910697720944881, + 0.005314161069691181, + 0.009912431240081787, + 0.007568263448774815, + 0.010135666467249393, + 0.011136640794575214, + 0.004021591041237116, + 0.007024757564067841, + 0.007062257267534733, + 0.004800337832421064, + 0.014576289802789688, + 0.007948163896799088, + 0.02129446156322956, + 0.005896695889532566, + 0.00440134247764945, + 0.0046917106956243515, + 0.00411874009296298, + 0.002939838683232665, + 0.007371230982244015, + 0.00426048319786787, + 0.011517446488142014, + 0.00797947682440281, + 0.004814295098185539, + 0.0043605174869298935, + 0.006377820391207933, + 0.005387074314057827, + 0.00578655581921339, + 0.010736562311649323, + 0.005176729988306761, + 0.0042898752726614475, + 0.004832375328987837, + 0.009501890279352665, + 0.005225289613008499, + 0.0059954021126031876, + 0.005761669483035803, + 0.006832945626229048, + 0.007993178442120552, + 0.010137438774108887, + 0.007826076820492744, + 0.018703395500779152, + 0.04994899034500122, + 0.0046014669351279736, + 0.015551241114735603, + 0.008962532505393028, + 0.04351150617003441, + 0.019147811457514763, + 0.027114344760775566, + 0.016300208866596222, + 0.06554210931062698, + 0.010370900854468346, + 0.014946872368454933, + 0.02238215133547783, + 0.047140322625637054, + 0.0024710758589208126, + 0.0029890926089137793, + 0.0023250326048582792, + 0.002959195524454117, + 0.002122309524565935, + 0.002192261628806591, + 0.004432929679751396, + 0.003986391704529524, + 0.012293624691665173, + 0.029707033187150955, + 0.0148324528709054, + 0.07023418694734573, + 0.010239338502287865, + 0.02290465496480465, + 0.02069377712905407, + 0.03354963660240173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0006337541271932423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive jellyfish with a deep, luminous crimson bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish feature intricate, swirling gold patterns on its bell and a mane of yellow bioluminescent tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the indigo-blue depths of water with stylized, sweeping curves representing currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the jellyfish exhibit a slow, rhythmic pulsating motion where its body contracts and expands?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808479383587837 + ], + "opposite_logprob_tokens": [ + -3.2880847454071045 + ], + "logprob_reward": -0.03808479383587837, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808479383587837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06914368271827698, + 0.10554824769496918, + 0.04783695936203003, + 0.01601172797381878, + 0.11956461519002914, + 0.12133792042732239, + 0.04812498018145561, + 0.3211340606212616, + 0.06580229103565216, + 0.08549544215202332 + ], + "credit_map": [ + 0.008474561385810375, + 0.006503287702798843, + 0.009688536636531353, + 0.007954925298690796, + 0.007907360792160034, + 0.0071165659464895725, + 0.011037728749215603, + 0.01046072319149971, + 0.013404248282313347, + 0.017213737592101097, + 0.011573893949389458, + 0.013068846426904202, + 0.008336749859154224, + 0.01456883829087019, + 0.011734725907444954, + 0.015647226944565773, + 0.00833814311772585, + 0.0038460169453173876, + 0.006180272437632084, + 0.004687872249633074, + 0.0034989973064512014, + 0.006797472946345806, + 0.006057521793991327, + 0.008430670946836472, + 0.0017435700865462422, + 0.0015447940677404404, + 0.0021424675360322, + 0.0021187688689678907, + 0.0017452673055231571, + 0.001613709726370871, + 0.0029981452971696854, + 0.002105007879436016, + 0.02349480614066124, + 0.007547200191766024, + 0.010606026276946068, + 0.012329579330980778, + 0.01034486573189497, + 0.01878872513771057, + 0.0171597171574831, + 0.019293716177344322, + 0.01676984317600727, + 0.014619916677474976, + 0.01650828681886196, + 0.015202272683382034, + 0.012188316322863102, + 0.0152596989646554, + 0.01556907594203949, + 0.015220526605844498, + 0.003653038991615176, + 0.006046151276677847, + 0.005696685053408146, + 0.008507725782692432, + 0.006288668606430292, + 0.004616434220224619, + 0.004777821712195873, + 0.008538458496332169, + 0.02708819881081581, + 0.044250208884477615, + 0.021623751148581505, + 0.09571635723114014, + 0.01485158409923315, + 0.020563511177897453, + 0.033563047647476196, + 0.06347742676734924, + 0.006467287428677082, + 0.009676273912191391, + 0.009085046127438545, + 0.008621441200375557, + 0.006420212332159281, + 0.007551571819931269, + 0.009638365358114243, + 0.00834210216999054, + 0.006373555865138769, + 0.01008352916687727, + 0.007739747874438763, + 0.019454380497336388, + 0.005801346618682146, + 0.007272326387465, + 0.015107844024896622, + 0.01366272009909153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.03808479383587837 + } + ], + "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/0000128/group_0/index_tensor([4166, 4166])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_0/index_tensor([4166, 4166])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bbdad1e14df72733218d9d6daca0bedb0055cdf4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_0/index_tensor([4166, 4166])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, horizontal, static shot focuses squarely on an absolutely pure, unadulterated white background. In the center, a child-like bullfrog, drawn in simple, bold black lines, stands at the apex of its roar. Its outline is rendered in minimal, bold black lines, devoid of color fill or shading. Its body leans back slightly from exertion, its limbs splayed outward in a tense pose. Its signature cheeks bulge dramatically to the sides, forming two perfectly round air pockets even larger than its head, dominating the frame. Its mouth gapes into a black O, as if unleashing a silent yet earth-shaking roar. The composition is simple, with stark black and white. Through the extreme exaggeration of the dynamic moment, the simplest lines successfully convey a comical yet explosive visual impact, creating a unique and intense atmosphere.", + "seed": 4476, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_0/index_4166_seed_4476_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": -0.20328593030571937, + "vlm_logprob": -0.20328593030571937, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.20328593030571937, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.758463442325592 + ], + "opposite_logprob_tokens": [ + -0.633463442325592 + ], + "logprob_reward": -0.758463442325592, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.758463442325592, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21495234966278076, + 0.13827824592590332, + 0.16006340086460114, + 0.03859694302082062, + 0.06611078232526779, + 0.1969934105873108, + 0.023426279425621033, + 0.13157056272029877, + 0.010476991534233093, + 0.019531024619936943 + ], + "credit_map": [ + 0.06937676668167114, + 0.012853636406362057, + 0.028751637786626816, + 0.015681905671954155, + 0.009370566345751286, + 0.04936716705560684, + 0.014344832859933376, + 0.015205858275294304, + 0.04260997101664543, + 0.007717607077211142, + 0.020063010975718498, + 0.009237002581357956, + 0.006535121239721775, + 0.02983287163078785, + 0.009604676626622677, + 0.012677998282015324, + 0.048871930688619614, + 0.009855562821030617, + 0.02307453379034996, + 0.011470837518572807, + 0.006867284886538982, + 0.03438449278473854, + 0.010668754577636719, + 0.014870021492242813, + 0.013165184296667576, + 0.002099611097946763, + 0.0045737773180007935, + 0.0025696312077343464, + 0.0017403895035386086, + 0.008012903854250908, + 0.003427156014367938, + 0.0030082883313298225, + 0.022482924163341522, + 0.004404714331030846, + 0.009324446320533752, + 0.003251443151384592, + 0.0029556730296462774, + 0.014789052307605743, + 0.0032026616390794516, + 0.0056998757645487785, + 0.0749419629573822, + 0.00926293432712555, + 0.02638307958841324, + 0.00832284428179264, + 0.006645855959504843, + 0.05049603804945946, + 0.007425601594150066, + 0.01351509615778923, + 0.007730301469564438, + 0.0018805696163326502, + 0.0031268203165382147, + 0.0012737024808302522, + 0.001087696640752256, + 0.004869458731263876, + 0.0015869929920881987, + 0.0018707404378801584, + 0.046936120837926865, + 0.006755958776921034, + 0.01757042668759823, + 0.007130815647542477, + 0.0049061644822359085, + 0.03247465193271637, + 0.005945997778326273, + 0.009850453585386276, + 0.0018091718666255474, + 0.0010763510363176465, + 0.0017453166656196117, + 0.0010044628288596869, + 0.0008454956696368754, + 0.0016747914487496018, + 0.0013202293775975704, + 0.0010011737467721105, + 0.0069059282541275024, + 0.0013640185352414846, + 0.0024297283962368965, + 0.001013552537187934, + 0.0008824272081255913, + 0.004229689482599497, + 0.0011556897079572082, + 0.0015499917790293694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.758463442325592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033750079572200775 + ], + "opposite_logprob_tokens": [ + -3.408750057220459 + ], + "logprob_reward": -0.033750079572200775, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033750079572200775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24007640779018402, + 0.15118050575256348, + 0.13226181268692017, + 0.057392850518226624, + 0.038176462054252625, + 0.17165113985538483, + 0.02824004925787449, + 0.08909928798675537, + 0.06796439737081528, + 0.023957133293151855 + ], + "credit_map": [ + 0.07614972442388535, + 0.018345756456255913, + 0.03013679012656212, + 0.0162709578871727, + 0.014155550859868526, + 0.05363800376653671, + 0.014030196703970432, + 0.017349425703287125, + 0.04817471280694008, + 0.010927602648735046, + 0.020379584282636642, + 0.007942534051835537, + 0.00884520448744297, + 0.033127762377262115, + 0.008322780020534992, + 0.013460327871143818, + 0.04043852910399437, + 0.008992019109427929, + 0.018842708319425583, + 0.008900023996829987, + 0.005666733719408512, + 0.028014792129397392, + 0.008561917580664158, + 0.012845083139836788, + 0.018933719024062157, + 0.004106324631720781, + 0.006884293165057898, + 0.0035415138117969036, + 0.003024105681106448, + 0.011570392176508904, + 0.00447517354041338, + 0.004857327323406935, + 0.0067294505424797535, + 0.002903463551774621, + 0.005776089150458574, + 0.0036940178833901882, + 0.003317971248179674, + 0.005995040759444237, + 0.004993996117264032, + 0.004766430240124464, + 0.05879908427596092, + 0.009799693711102009, + 0.02300766296684742, + 0.00942584965378046, + 0.008461673744022846, + 0.04080621525645256, + 0.00870436616241932, + 0.01264659408479929, + 0.007700629532337189, + 0.0023473226465284824, + 0.0042026652954518795, + 0.002046797890216112, + 0.0017672254471108317, + 0.005248336587101221, + 0.002589672803878784, + 0.0023373982403427362, + 0.024457694962620735, + 0.008351792581379414, + 0.01138724759221077, + 0.006596812047064304, + 0.005752179771661758, + 0.01855182833969593, + 0.006042917724698782, + 0.007958811707794666, + 0.013959064148366451, + 0.006924582179635763, + 0.011354724876582623, + 0.005824458785355091, + 0.00495865847915411, + 0.011288539506494999, + 0.0075384462252259254, + 0.006115917582064867, + 0.005363401025533676, + 0.0019457921152934432, + 0.0035473965108394623, + 0.0018710116855800152, + 0.0020528691820800304, + 0.004398007411509752, + 0.0027116837445646524, + 0.0020669696386903524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.033750079572200775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789785385131836 + ], + "opposite_logprob_tokens": [ + -2.5789785385131836 + ], + "logprob_reward": -0.0789785385131836, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789785385131836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14059928059577942, + 0.17243513464927673, + 0.15384580194950104, + 0.04272063449025154, + 0.06372223049402237, + 0.06619035452604294, + 0.09120020270347595, + 0.12886033952236176, + 0.09979713708162308, + 0.040628883987665176 + ], + "credit_map": [ + 0.03129974752664566, + 0.009128117002546787, + 0.01985703967511654, + 0.011351082473993301, + 0.01041773147881031, + 0.02709261141717434, + 0.0171823687851429, + 0.014270573854446411, + 0.03882937878370285, + 0.010934271849691868, + 0.025978999212384224, + 0.013841291889548302, + 0.011643721722066402, + 0.033599551767110825, + 0.017910338938236237, + 0.01969757117331028, + 0.03461242467164993, + 0.014389757998287678, + 0.022493669763207436, + 0.012229712679982185, + 0.011237741447985172, + 0.026608919724822044, + 0.01596662402153015, + 0.01630694605410099, + 0.007865266874432564, + 0.005837730597704649, + 0.00517412181943655, + 0.0040434496477246284, + 0.00393605837598443, + 0.006062791217118502, + 0.0058282529935240746, + 0.003972959239035845, + 0.018079450353980064, + 0.004579359665513039, + 0.007984060794115067, + 0.0046473173424601555, + 0.005381305702030659, + 0.011356298811733723, + 0.006412712391465902, + 0.005281721707433462, + 0.01921217329800129, + 0.005312739871442318, + 0.008011306636035442, + 0.004111544694751501, + 0.004692521877586842, + 0.01439366303384304, + 0.005251503083854914, + 0.005204898305237293, + 0.021870432421565056, + 0.009928185492753983, + 0.01329096406698227, + 0.006687641143798828, + 0.006344600114971399, + 0.015279365703463554, + 0.009064191952347755, + 0.008734830655157566, + 0.032399773597717285, + 0.015292037278413773, + 0.018321754410862923, + 0.009127245284616947, + 0.00852931197732687, + 0.024108875542879105, + 0.009918446652591228, + 0.011162906885147095, + 0.019586531445384026, + 0.00930430181324482, + 0.016420559957623482, + 0.009240631014108658, + 0.007839213125407696, + 0.01611892133951187, + 0.012448005378246307, + 0.008838974870741367, + 0.010686379857361317, + 0.003100386820733547, + 0.005402884446084499, + 0.0033398624509572983, + 0.002761614741757512, + 0.007320657838135958, + 0.004193707834929228, + 0.003823390696197748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0789785385131836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704899907112122 + ], + "opposite_logprob_tokens": [ + -2.127048969268799 + ], + "logprob_reward": -0.12704899907112122, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704899907112122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13653914630413055, + 0.16581398248672485, + 0.05340619385242462, + 0.08308128267526627, + 0.07743065804243088, + 0.1578415483236313, + 0.06647331267595291, + 0.14924195408821106, + 0.07840920984745026, + 0.031762681901454926 + ], + "credit_map": [ + 0.026297589763998985, + 0.012051899917423725, + 0.019477127119898796, + 0.012894032523036003, + 0.011498411186039448, + 0.02400108613073826, + 0.016929572448134422, + 0.01338944025337696, + 0.044131409376859665, + 0.011131463572382927, + 0.022172432392835617, + 0.014275545254349709, + 0.011779877357184887, + 0.03003082424402237, + 0.018658962100744247, + 0.01363347377628088, + 0.011686502955853939, + 0.0045082299038767815, + 0.006014647427946329, + 0.005088421516120434, + 0.005232383031398058, + 0.009781449101865292, + 0.006921548396348953, + 0.004173021297901869, + 0.02895277366042137, + 0.004050308372825384, + 0.010476668365299702, + 0.004424371290951967, + 0.004014125093817711, + 0.018768904730677605, + 0.004674817901104689, + 0.007719326298683882, + 0.01960829645395279, + 0.006536142900586128, + 0.010302472859621048, + 0.00686675263568759, + 0.00490057235583663, + 0.01415509544312954, + 0.007728623226284981, + 0.007332708686590195, + 0.04939908906817436, + 0.009513973258435726, + 0.021801777184009552, + 0.009450790472328663, + 0.009221973828971386, + 0.036544062197208405, + 0.008907172828912735, + 0.013002725318074226, + 0.015575156547129154, + 0.005916160997003317, + 0.01062886230647564, + 0.0043056127615273, + 0.005616514943540096, + 0.012163943611085415, + 0.005279533565044403, + 0.006987533997744322, + 0.0490562841296196, + 0.008185414597392082, + 0.019036998972296715, + 0.009998905472457409, + 0.007996615022420883, + 0.03429676592350006, + 0.009043483063578606, + 0.011627499014139175, + 0.017566034570336342, + 0.006275196094065905, + 0.01348511315882206, + 0.006549614481627941, + 0.0051243966445326805, + 0.013845925219357014, + 0.008143496699631214, + 0.007419439498335123, + 0.006985136307775974, + 0.003179319202899933, + 0.004222852177917957, + 0.002950407098978758, + 0.002690940396860242, + 0.004855982959270477, + 0.003612709231674671, + 0.003265337087213993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.12704899907112122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188592046499252 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.018188592046499252, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188592046499252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17950952053070068, + 0.08404229581356049, + 0.09309780597686768, + 0.11168662458658218, + 0.11736855655908585, + 0.1536228060722351, + 0.05369333177804947, + 0.0634860172867775, + 0.05951301380991936, + 0.08398007601499557 + ], + "credit_map": [ + 0.03764226660132408, + 0.014023294672369957, + 0.026140907779335976, + 0.014225222170352936, + 0.017524197697639465, + 0.02889746055006981, + 0.020920593291521072, + 0.020135561004281044, + 0.015315729193389416, + 0.007515657227486372, + 0.01252285111695528, + 0.009061582386493683, + 0.008919714018702507, + 0.011608920060098171, + 0.010623816400766373, + 0.008474013768136501, + 0.022160658612847328, + 0.006770827807486057, + 0.01371129509061575, + 0.009078266099095345, + 0.005470088683068752, + 0.017834670841693878, + 0.009517723694443703, + 0.008554267697036266, + 0.027963025495409966, + 0.011803103610873222, + 0.016185011714696884, + 0.009277351200580597, + 0.007529588881880045, + 0.01995173841714859, + 0.010323813185095787, + 0.008652975782752037, + 0.02701699361205101, + 0.009330757893621922, + 0.014197265729308128, + 0.008486175909638405, + 0.01251025591045618, + 0.019575899466872215, + 0.013729795813560486, + 0.012521399185061455, + 0.0560251921415329, + 0.007616021670401096, + 0.01917135901749134, + 0.0081026591360569, + 0.007183856796473265, + 0.0361957810819149, + 0.008887519128620625, + 0.010440408252179623, + 0.01283019594848156, + 0.005813031457364559, + 0.007671690545976162, + 0.004520541522651911, + 0.003752374090254307, + 0.009114555083215237, + 0.005066789221018553, + 0.004924145061522722, + 0.011984934099018574, + 0.006160371471196413, + 0.009384549222886562, + 0.005279003642499447, + 0.0052091386169195175, + 0.0106471823528409, + 0.007769712246954441, + 0.007051120512187481, + 0.015291307121515274, + 0.005680397152900696, + 0.00877896137535572, + 0.004722596146166325, + 0.0036935533862560987, + 0.010770573280751705, + 0.005661775358021259, + 0.004913845099508762, + 0.013076432049274445, + 0.008974047377705574, + 0.01303280983120203, + 0.008103410713374615, + 0.008786066435277462, + 0.011937885545194149, + 0.011188753880560398, + 0.008880666457116604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.018188592046499252 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, horizontal, static shot focuses squarely on an absolutely pure, unadulterated white background. In the center, a child-like bullfrog, drawn in simple, bold black lines, stands at the apex of its roar. Its outline is rendered in minimal, bold black lines, devoid of color fill or shading. Its body leans back slightly from exertion, its limbs splayed outward in a tense pose. Its signature cheeks bulge dramatically to the sides, forming two perfectly round air pockets even larger than its head, dominating the frame. Its mouth gapes into a black O, as if unleashing a silent yet earth-shaking roar. The composition is simple, with stark black and white. Through the extreme exaggeration of the dynamic moment, the simplest lines successfully convey a comical yet explosive visual impact, creating a unique and intense atmosphere.", + "seed": 4477, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_0/index_4166_seed_4477_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": -0.203681056573987, + "vlm_logprob": -0.203681056573987, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.203681056573987, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8269830346107483 + ], + "opposite_logprob_tokens": [ + -0.5769830346107483 + ], + "logprob_reward": -0.8269830346107483, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8269830346107483, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23672311007976532, + 0.08163785189390182, + 0.1041116788983345, + 0.2502673864364624, + 0.23978109657764435, + 0.04334733262658119, + 0.006319413892924786, + 0.00779403280466795, + 0.00948522798717022, + 0.020532798022031784 + ], + "credit_map": [ + 0.0807599350810051, + 0.013154120184481144, + 0.03738482668995857, + 0.012049923650920391, + 0.005927100777626038, + 0.039602331817150116, + 0.011047997511923313, + 0.03679688647389412, + 0.02268318459391594, + 0.007209615781903267, + 0.015015212818980217, + 0.005990177392959595, + 0.0038635022938251495, + 0.01425238884985447, + 0.003989233169704676, + 0.008634543977677822, + 0.020905086770653725, + 0.014811079949140549, + 0.0173141248524189, + 0.00725126639008522, + 0.009386287070810795, + 0.015159661881625652, + 0.005501304753124714, + 0.013782868161797523, + 0.08596348762512207, + 0.012027189135551453, + 0.04342170059680939, + 0.009756121784448624, + 0.006548744160681963, + 0.04292486980557442, + 0.010195127688348293, + 0.0394301638007164, + 0.07287070900201797, + 0.020424047484993935, + 0.0404709093272686, + 0.010805227793753147, + 0.011852017603814602, + 0.03854295611381531, + 0.009932279586791992, + 0.034882962703704834, + 0.012685242109000683, + 0.004505868069827557, + 0.006984897889196873, + 0.00261882646009326, + 0.00252591073513031, + 0.0066498275846242905, + 0.0020186237525194883, + 0.005358139052987099, + 0.001293450826779008, + 0.0006632677977904677, + 0.0008648847579024732, + 0.0006289499578997493, + 0.000527238822542131, + 0.0009398228721693158, + 0.0005765886744484305, + 0.000825210299808532, + 0.00254653487354517, + 0.0003824148734565824, + 0.0013207976007834077, + 0.0004247799515724182, + 0.0003284606209490448, + 0.0012070112861692905, + 0.0004571556346490979, + 0.0011268788948655128, + 0.0017485079588368535, + 0.0014858231879770756, + 0.0014674130361527205, + 0.0007979395450092852, + 0.0008392444578930736, + 0.0013548837741836905, + 0.0006875371909700334, + 0.0011038795346394181, + 0.0045870584435760975, + 0.0021740871015936136, + 0.0034033851698040962, + 0.000985450460575521, + 0.0011963995639234781, + 0.0034699412062764168, + 0.0008633061079308391, + 0.003853170434013009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.8269830346107483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638625167310238 + ], + "opposite_logprob_tokens": [ + -3.651386260986328 + ], + "logprob_reward": -0.02638625167310238, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638625167310238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18159566819667816, + 0.08948615938425064, + 0.12834806740283966, + 0.20159408450126648, + 0.15590061247348785, + 0.014579381793737411, + 0.021623235195875168, + 0.03204035758972168, + 0.05894021689891815, + 0.1158922016620636 + ], + "credit_map": [ + 0.06827357411384583, + 0.007547331042587757, + 0.033380988985300064, + 0.007791562471538782, + 0.006483383476734161, + 0.02990889549255371, + 0.006801884155720472, + 0.021408047527074814, + 0.02330237627029419, + 0.00808656681329012, + 0.0138696925714612, + 0.006819576025009155, + 0.004562658257782459, + 0.014101369306445122, + 0.005956584587693214, + 0.01278733927756548, + 0.025676880031824112, + 0.015207519754767418, + 0.016203969717025757, + 0.01284999679774046, + 0.013772142119705677, + 0.016859328374266624, + 0.012608388438820839, + 0.015169846825301647, + 0.06532230228185654, + 0.01013648696243763, + 0.03185936063528061, + 0.008137061260640621, + 0.005754266399890184, + 0.03684404864907265, + 0.008642436936497688, + 0.03489813581109047, + 0.021263059228658676, + 0.01954638585448265, + 0.016049616038799286, + 0.02137499488890171, + 0.01598522812128067, + 0.022189905866980553, + 0.01742405816912651, + 0.022067364305257797, + 0.0028412211686372757, + 0.001818216172978282, + 0.0018567930674180388, + 0.0011284417705610394, + 0.0009766275761649013, + 0.002857665531337261, + 0.000985671766102314, + 0.0021147446241229773, + 0.007028946653008461, + 0.001590566709637642, + 0.004626626614481211, + 0.0013090958818793297, + 0.0009741080575622618, + 0.002980069722980261, + 0.0010192678309977055, + 0.0020945537835359573, + 0.0068811532109975815, + 0.003948253579437733, + 0.0048621948808431625, + 0.0031017386354506016, + 0.003011178458109498, + 0.0043051172979176044, + 0.0028620429802685976, + 0.0030686764512211084, + 0.011532094329595566, + 0.00835818238556385, + 0.010021664202213287, + 0.0049987854436039925, + 0.004508909769356251, + 0.008967701345682144, + 0.0038947423454374075, + 0.006658134516328573, + 0.022806964814662933, + 0.016134267672896385, + 0.0179104246199131, + 0.011267583817243576, + 0.010416053235530853, + 0.01637456752359867, + 0.009505902417004108, + 0.01147643942385912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.02638625167310238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031053423881531 + ], + "opposite_logprob_tokens": [ + -2.3503105640411377 + ], + "logprob_reward": -0.10031053423881531, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031053423881531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24274885654449463, + 0.1397249847650528, + 0.06794743239879608, + 0.15634433925151825, + 0.09288240969181061, + 0.07295936346054077, + 0.05672154948115349, + 0.07966600358486176, + 0.040806036442518234, + 0.050199124962091446 + ], + "credit_map": [ + 0.04947064071893692, + 0.026530113071203232, + 0.030653616413474083, + 0.01669374480843544, + 0.01704326830804348, + 0.04804742708802223, + 0.017229583114385605, + 0.03708042576909065, + 0.018941860646009445, + 0.016112422570586205, + 0.01712786592543125, + 0.01837645284831524, + 0.010385964997112751, + 0.01679908111691475, + 0.017757311463356018, + 0.02422400563955307, + 0.013990231789648533, + 0.009699113667011261, + 0.010597706772387028, + 0.004985854495316744, + 0.007157627493143082, + 0.00928316917270422, + 0.0043005854822695255, + 0.007933131419122219, + 0.029649898409843445, + 0.014433433301746845, + 0.019254684448242188, + 0.010589122772216797, + 0.00932256132364273, + 0.029231291264295578, + 0.010272029787302017, + 0.03359129652380943, + 0.010072536766529083, + 0.013815652579069138, + 0.010728754103183746, + 0.009305036626756191, + 0.011408681981265545, + 0.013273471966385841, + 0.009842494502663612, + 0.014435764402151108, + 0.012468939647078514, + 0.009336009621620178, + 0.008061529137194157, + 0.006671569775789976, + 0.010042501613497734, + 0.010712540708482265, + 0.006766307167708874, + 0.00889995414763689, + 0.013221733272075653, + 0.005877892952412367, + 0.007950554601848125, + 0.0056197564117610455, + 0.004470307845622301, + 0.006849096622318029, + 0.005093585234135389, + 0.007638616021722555, + 0.017894228920340538, + 0.011427415534853935, + 0.010769190266728401, + 0.0062078614719212055, + 0.007145212031900883, + 0.011103719472885132, + 0.00582111394032836, + 0.009297255426645279, + 0.006093141622841358, + 0.0038017851766198874, + 0.0068565658293664455, + 0.0031707854941487312, + 0.0025752943474799395, + 0.005414812825620174, + 0.0027209059335291386, + 0.01017274335026741, + 0.01322994101792574, + 0.0045966096222400665, + 0.009112427942454815, + 0.0033059800043702126, + 0.0027947048656642437, + 0.007257477845996618, + 0.0028985904064029455, + 0.007003384176641703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.10031053423881531 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865920543670654 + ], + "opposite_logprob_tokens": [ + -3.048659324645996 + ], + "logprob_reward": -0.04865920543670654, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865920543670654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17228905856609344, + 0.20031394064426422, + 0.20596612989902496, + 0.03249090537428856, + 0.11953926831483841, + 0.027144411578774452, + 0.029992740601301193, + 0.11143947392702103, + 0.03268033638596535, + 0.0681437999010086 + ], + "credit_map": [ + 0.025290055200457573, + 0.023617396131157875, + 0.023346588015556335, + 0.020243847742676735, + 0.016405951231718063, + 0.02296702191233635, + 0.017383676022291183, + 0.023034503683447838, + 0.032917510718107224, + 0.030777249485254288, + 0.03232480213046074, + 0.016939062625169754, + 0.020081104710698128, + 0.0296494010835886, + 0.015118878334760666, + 0.02250591851770878, + 0.03950396180152893, + 0.02978917397558689, + 0.030806753784418106, + 0.016373690217733383, + 0.01876436360180378, + 0.029565414413809776, + 0.012844480574131012, + 0.028318263590335846, + 0.004419079516083002, + 0.00323230167850852, + 0.005630206316709518, + 0.003116399049758911, + 0.002782992320135236, + 0.004854285623878241, + 0.004104705527424812, + 0.004350931849330664, + 0.018086226657032967, + 0.01889622211456299, + 0.015176341868937016, + 0.01073353923857212, + 0.016405418515205383, + 0.015280654653906822, + 0.010828143917024136, + 0.014132702723145485, + 0.003732098266482353, + 0.0044076163321733475, + 0.002973003778606653, + 0.00198933738283813, + 0.0031049116514623165, + 0.004075841512531042, + 0.0018141333712264895, + 0.005047464743256569, + 0.006073500029742718, + 0.0037783896550536156, + 0.004255576990544796, + 0.002389528090134263, + 0.0020192221272736788, + 0.003950205165892839, + 0.0022411285899579525, + 0.005285187624394894, + 0.018850937485694885, + 0.016861531883478165, + 0.014051960781216621, + 0.010321833193302155, + 0.013189108110964298, + 0.015498783439397812, + 0.011918791569769382, + 0.010746522806584835, + 0.005921524949371815, + 0.00428295461460948, + 0.0054230717942118645, + 0.0018998264567926526, + 0.0021765660494565964, + 0.00477884104475379, + 0.0016501209465786815, + 0.006547426339238882, + 0.017797932028770447, + 0.007037663366645575, + 0.011656043119728565, + 0.005385033320635557, + 0.004344827961176634, + 0.009483162313699722, + 0.004455915652215481, + 0.007983209565281868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.04865920543670654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14632800221443176, + 0.15350830554962158, + 0.057219937443733215, + 0.26989778876304626, + 0.08080913126468658, + 0.06086676940321922, + 0.048716481775045395, + 0.04552745074033737, + 0.044746335595846176, + 0.09237982332706451 + ], + "credit_map": [ + 0.03047707863152027, + 0.012056352570652962, + 0.018442535772919655, + 0.012802528217434883, + 0.015149061568081379, + 0.0226313304156065, + 0.014215418137609959, + 0.020553702488541603, + 0.02618144452571869, + 0.019132470712065697, + 0.02882327139377594, + 0.011130171827971935, + 0.014135124161839485, + 0.022577278316020966, + 0.011185767129063606, + 0.020342763513326645, + 0.008350404910743237, + 0.008874919265508652, + 0.00781889446079731, + 0.006880967412143946, + 0.006164211314171553, + 0.007614411413669586, + 0.006499319337308407, + 0.0050168088637292385, + 0.0758214145898819, + 0.012708509340882301, + 0.03812166675925255, + 0.012103580869734287, + 0.007862952537834644, + 0.052013516426086426, + 0.014127635397017002, + 0.057138532400131226, + 0.01245171669870615, + 0.011673682369291782, + 0.011764737777411938, + 0.0065687913447618484, + 0.007318480871617794, + 0.012291358783841133, + 0.007268535438925028, + 0.011471833102405071, + 0.010415507480502129, + 0.010130058974027634, + 0.009467133320868015, + 0.0046640620566904545, + 0.004968412220478058, + 0.007031719665974379, + 0.003980860114097595, + 0.010209014639258385, + 0.00747174397110939, + 0.00874132476747036, + 0.00625592190772295, + 0.004802325740456581, + 0.005729760508984327, + 0.0065672388300299644, + 0.004293181002140045, + 0.004854980856180191, + 0.007129305507987738, + 0.004333552438765764, + 0.006333040539175272, + 0.0037798357661813498, + 0.005244588945060968, + 0.0068863993510603905, + 0.004685488995164633, + 0.007135238964110613, + 0.008786638267338276, + 0.005285740830004215, + 0.006733049638569355, + 0.004138936288654804, + 0.0037729484029114246, + 0.006584340240806341, + 0.0038004072848707438, + 0.005644272547215223, + 0.027447126805782318, + 0.007932462729513645, + 0.01681666634976864, + 0.005909594241529703, + 0.005156924948096275, + 0.013338630087673664, + 0.004633706528693438, + 0.01114471722394228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.016066256910562515 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, horizontal, static shot focuses squarely on an absolutely pure, unadulterated white background. In the center, a child-like bullfrog, drawn in simple, bold black lines, stands at the apex of its roar. Its outline is rendered in minimal, bold black lines, devoid of color fill or shading. Its body leans back slightly from exertion, its limbs splayed outward in a tense pose. Its signature cheeks bulge dramatically to the sides, forming two perfectly round air pockets even larger than its head, dominating the frame. Its mouth gapes into a black O, as if unleashing a silent yet earth-shaking roar. The composition is simple, with stark black and white. Through the extreme exaggeration of the dynamic moment, the simplest lines successfully convey a comical yet explosive visual impact, creating a unique and intense atmosphere.", + "seed": 4478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_0/index_4166_seed_4478_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": -0.12828131802380086, + "vlm_logprob": -0.12828131802380086, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.12828131802380086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.350376158952713 + ], + "opposite_logprob_tokens": [ + -1.2253761291503906 + ], + "logprob_reward": -0.350376158952713, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.350376158952713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29451659321784973, + 0.08305258303880692, + 0.07448744028806686, + 0.15395714342594147, + 0.16403010487556458, + 0.11361809819936752, + 0.04682059958577156, + 0.028691915795207024, + 0.018628232181072235, + 0.022197343409061432 + ], + "credit_map": [ + 0.02494928613305092, + 0.09009852260351181, + 0.04998144134879112, + 0.020044460892677307, + 0.03043563850224018, + 0.024326592683792114, + 0.026834186166524887, + 0.027846474200487137, + 0.008104912005364895, + 0.020323602482676506, + 0.014065388590097427, + 0.006445392034947872, + 0.008294645696878433, + 0.009341880679130554, + 0.007810063194483519, + 0.008666695095598698, + 0.010003759525716305, + 0.006686059758067131, + 0.012696650810539722, + 0.006438255775719881, + 0.006745085120201111, + 0.018967928364872932, + 0.005907818209379911, + 0.007041885983198881, + 0.01171139720827341, + 0.03740910068154335, + 0.02644386887550354, + 0.016512196511030197, + 0.014600003138184547, + 0.017509372904896736, + 0.015233339741826057, + 0.014537862502038479, + 0.010911060497164726, + 0.03592672944068909, + 0.025371752679347992, + 0.016470851376652718, + 0.023827947676181793, + 0.010703545063734055, + 0.0254744291305542, + 0.015343782491981983, + 0.009573831222951412, + 0.03502323850989342, + 0.01953105628490448, + 0.009706487879157066, + 0.012896856293082237, + 0.011409946717321873, + 0.007744934409856796, + 0.007731739431619644, + 0.004496491048485041, + 0.009232258424162865, + 0.0076944478787481785, + 0.0046631828881800175, + 0.005967768374830484, + 0.00554494746029377, + 0.004593523219227791, + 0.004627978894859552, + 0.003618724411353469, + 0.004197051282972097, + 0.005009396467357874, + 0.003622011048719287, + 0.0030301224905997515, + 0.0033467866014689207, + 0.002399437827989459, + 0.0034683875273913145, + 0.002439245581626892, + 0.0024558245204389095, + 0.003278150921687484, + 0.0016096018953248858, + 0.0021494359243661165, + 0.0028610716108232737, + 0.0016590239247307181, + 0.0021758778020739555, + 0.0017210125224664807, + 0.0038656957913190126, + 0.00375524815171957, + 0.003095455700531602, + 0.002299673156812787, + 0.0025268171448260546, + 0.003051369683817029, + 0.0018820727709680796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.350376158952713 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979982316493988 + ], + "opposite_logprob_tokens": [ + -4.634799957275391 + ], + "logprob_reward": -0.00979982316493988, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979982316493988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13364896178245544, + 0.08392971009016037, + 0.06684311479330063, + 0.11920933425426483, + 0.23071350157260895, + 0.10644223541021347, + 0.08404461294412613, + 0.11738593131303787, + 0.034883592277765274, + 0.022899026051163673 + ], + "credit_map": [ + 0.012391882948577404, + 0.0342073030769825, + 0.021396536380052567, + 0.010550410486757755, + 0.015089074149727821, + 0.015050853602588177, + 0.01191239058971405, + 0.013050502166152, + 0.00774782057851553, + 0.012420376762747765, + 0.011411208659410477, + 0.008252991363406181, + 0.012369435280561447, + 0.0094964150339365, + 0.012076176702976227, + 0.010155276395380497, + 0.010397424921393394, + 0.006987038068473339, + 0.012633654288947582, + 0.00423772307112813, + 0.005793105810880661, + 0.016549957916140556, + 0.004430869594216347, + 0.005813338793814182, + 0.010326559655368328, + 0.03440842404961586, + 0.020587829872965813, + 0.010880354791879654, + 0.012923698872327805, + 0.008790679275989532, + 0.01061194483190775, + 0.010679828003048897, + 0.011219664476811886, + 0.0410531610250473, + 0.03522457927465439, + 0.02352546714246273, + 0.04617086797952652, + 0.01324953231960535, + 0.03848149999976158, + 0.02178872562944889, + 0.007275667041540146, + 0.0305657796561718, + 0.01694381795823574, + 0.009682427160441875, + 0.012769519351422787, + 0.00933930929750204, + 0.01338879857212305, + 0.0064769163727760315, + 0.008850973099470139, + 0.012776198796927929, + 0.013498914428055286, + 0.010076813399791718, + 0.007947209291160107, + 0.010595917701721191, + 0.010127800516784191, + 0.010170775465667248, + 0.014191697351634502, + 0.014567363075911999, + 0.019637739285826683, + 0.017192624509334564, + 0.011702803894877434, + 0.014330711215734482, + 0.011577656492590904, + 0.014185328967869282, + 0.002497120527550578, + 0.004863070789724588, + 0.005024353973567486, + 0.005288330372422934, + 0.00437468895688653, + 0.0024178733583539724, + 0.006532628554850817, + 0.003885523183271289, + 0.002348084934055805, + 0.00305664143525064, + 0.00354058132506907, + 0.003154020057991147, + 0.001821416080929339, + 0.003487575566396117, + 0.003081958508118987, + 0.002408746862784028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.00979982316493988 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030989348888397 + ], + "opposite_logprob_tokens": [ + -2.3503098487854004 + ], + "logprob_reward": -0.10030989348888397, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030989348888397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16178475320339203, + 0.056172389537096024, + 0.15409831702709198, + 0.09708596765995026, + 0.12911474704742432, + 0.15410830080509186, + 0.07231738418340683, + 0.0757628083229065, + 0.05302157998085022, + 0.04653368145227432 + ], + "credit_map": [ + 0.015794264152646065, + 0.03256687894463539, + 0.022908495739102364, + 0.012763298116624355, + 0.02065245434641838, + 0.020638367161154747, + 0.017336629331111908, + 0.019124358892440796, + 0.003908765036612749, + 0.009467348456382751, + 0.007959856651723385, + 0.0048673441633582115, + 0.00876294169574976, + 0.006963771302253008, + 0.007678365800529718, + 0.006563995499163866, + 0.020715266466140747, + 0.013845101930201054, + 0.02752550318837166, + 0.013351498171687126, + 0.01299628708511591, + 0.041224170476198196, + 0.01203572191298008, + 0.012404770590364933, + 0.006412615533918142, + 0.02335391193628311, + 0.01526306476444006, + 0.012560197152197361, + 0.008743749931454659, + 0.00794840045273304, + 0.012585579417645931, + 0.010218452662229538, + 0.006553965620696545, + 0.021542124450206757, + 0.019324366003274918, + 0.012464100494980812, + 0.025148684158921242, + 0.009869998320937157, + 0.023423010483384132, + 0.01078849658370018, + 0.013836277648806572, + 0.027438977733254433, + 0.0226737093180418, + 0.014807010069489479, + 0.018239282071590424, + 0.020010877400636673, + 0.020357176661491394, + 0.016744986176490784, + 0.008673544973134995, + 0.012549543753266335, + 0.011808326467871666, + 0.006843534763902426, + 0.00770683353766799, + 0.007575429044663906, + 0.00823210459202528, + 0.008928065188229084, + 0.006957461591809988, + 0.008018902502954006, + 0.010315571911633015, + 0.007129944860935211, + 0.012633170932531357, + 0.009544887579977512, + 0.011480715125799179, + 0.009682153351604939, + 0.005396644584834576, + 0.006299985107034445, + 0.009045948274433613, + 0.006133954506367445, + 0.005567013286054134, + 0.007051445543766022, + 0.006909951101988554, + 0.006616636179387569, + 0.0031500705517828465, + 0.005933681968599558, + 0.007310336921364069, + 0.006526274140924215, + 0.005842950660735369, + 0.003616760252043605, + 0.009333177469670773, + 0.004820429254323244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.10030989348888397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034728288650513 + ], + "opposite_logprob_tokens": [ + -1.9103472232818604 + ], + "logprob_reward": -0.16034728288650513, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034728288650513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10907799005508423, + 0.11361419409513474, + 0.09089376032352448, + 0.09133601933717728, + 0.07349328696727753, + 0.20512373745441437, + 0.140410915017128, + 0.07000493258237839, + 0.07026655972003937, + 0.035778600722551346 + ], + "credit_map": [ + 0.008467278443276882, + 0.019164565950632095, + 0.01646655797958374, + 0.010129790753126144, + 0.015322468243539333, + 0.013251928612589836, + 0.011853496544063091, + 0.014421901665627956, + 0.010263065807521343, + 0.017613712698221207, + 0.01564551144838333, + 0.009215318597853184, + 0.014963637106120586, + 0.016418132930994034, + 0.01582794263958931, + 0.013666867278516293, + 0.01012753788381815, + 0.01142942812293768, + 0.01578338071703911, + 0.009714256040751934, + 0.009080087766051292, + 0.015204655937850475, + 0.009232528507709503, + 0.010321890003979206, + 0.006971485912799835, + 0.022835271432995796, + 0.015476332046091557, + 0.011256110854446888, + 0.00932239554822445, + 0.008466169238090515, + 0.010597215965390205, + 0.0064110360108315945, + 0.005953607615083456, + 0.011729477904736996, + 0.009626519866287708, + 0.006552973762154579, + 0.010672708041965961, + 0.01033683679997921, + 0.011159654706716537, + 0.007461512926965952, + 0.01382231991738081, + 0.028517641127109528, + 0.03376045450568199, + 0.019942115992307663, + 0.03359474986791611, + 0.015299122780561447, + 0.03762218356132507, + 0.02256515622138977, + 0.01196736004203558, + 0.022938322275877, + 0.021428178995847702, + 0.01927347667515278, + 0.01869850419461727, + 0.010661209933459759, + 0.01847376860678196, + 0.01697009988129139, + 0.005742962472140789, + 0.008608181960880756, + 0.012444917112588882, + 0.007387483958154917, + 0.011032105423510075, + 0.00982570368796587, + 0.007532148621976376, + 0.007431428413838148, + 0.006620087660849094, + 0.00722438283264637, + 0.010974853299558163, + 0.009802225045859814, + 0.009363486431539059, + 0.007357728201895952, + 0.011129903607070446, + 0.007793887984007597, + 0.0035135052166879177, + 0.006072433665394783, + 0.0057386355474591255, + 0.004035047255456448, + 0.004531003534793854, + 0.0029879773501306772, + 0.00547439232468605, + 0.0034256044309586287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.16034728288650513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573431625962257 + ], + "opposite_logprob_tokens": [ + -3.895573377609253 + ], + "logprob_reward": -0.020573431625962257, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573431625962257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1726425439119339, + 0.09560743719339371, + 0.052867721766233444, + 0.0896371379494667, + 0.11990170180797577, + 0.11719800531864166, + 0.122796930372715, + 0.15428249537944794, + 0.0414481982588768, + 0.03361772745847702 + ], + "credit_map": [ + 0.012271302752196789, + 0.018288247287273407, + 0.021211586892604828, + 0.02622056007385254, + 0.022635873407125473, + 0.01298003550618887, + 0.038469139486551285, + 0.02056582272052765, + 0.007885297760367393, + 0.014194592833518982, + 0.011908178217709064, + 0.010506986640393734, + 0.015223760157823563, + 0.009660773910582066, + 0.016043365001678467, + 0.010184495709836483, + 0.005648346617817879, + 0.006470038555562496, + 0.007940391078591347, + 0.006611555349081755, + 0.006188517436385155, + 0.007999450899660587, + 0.007431867532432079, + 0.004577558487653732, + 0.005024233367294073, + 0.017998497933149338, + 0.013793065212666988, + 0.012255080975592136, + 0.008359824307262897, + 0.007802536245435476, + 0.014063999056816101, + 0.010339915752410889, + 0.005889408756047487, + 0.021648656576871872, + 0.017817215994000435, + 0.012081073597073555, + 0.024169551208615303, + 0.008978991769254208, + 0.018668806180357933, + 0.010648008435964584, + 0.010615385137498379, + 0.021240826696157455, + 0.01784549467265606, + 0.009589106775820255, + 0.012785734608769417, + 0.016112037003040314, + 0.016239885240793228, + 0.012769555673003197, + 0.0118019487708807, + 0.02017030119895935, + 0.018677901476621628, + 0.01656937040388584, + 0.013991374522447586, + 0.010128662921488285, + 0.015924617648124695, + 0.015532777644693851, + 0.010337742045521736, + 0.016346139833331108, + 0.024248702451586723, + 0.01526985876262188, + 0.03154538571834564, + 0.015371385961771011, + 0.025420550256967545, + 0.015742741525173187, + 0.0029806566890329123, + 0.0052581848576664925, + 0.006700583267956972, + 0.005823416169732809, + 0.005380592308938503, + 0.0035843036603182554, + 0.006811287719756365, + 0.004909177776426077, + 0.003784832777455449, + 0.005107035394757986, + 0.006114968564361334, + 0.00305130030028522, + 0.0033245307859033346, + 0.005557299125939608, + 0.00352084101177752, + 0.0031569215934723616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.020573431625962257 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, horizontal, static shot focuses squarely on an absolutely pure, unadulterated white background. In the center, a child-like bullfrog, drawn in simple, bold black lines, stands at the apex of its roar. Its outline is rendered in minimal, bold black lines, devoid of color fill or shading. Its body leans back slightly from exertion, its limbs splayed outward in a tense pose. Its signature cheeks bulge dramatically to the sides, forming two perfectly round air pockets even larger than its head, dominating the frame. Its mouth gapes into a black O, as if unleashing a silent yet earth-shaking roar. The composition is simple, with stark black and white. Through the extreme exaggeration of the dynamic moment, the simplest lines successfully convey a comical yet explosive visual impact, creating a unique and intense atmosphere.", + "seed": 4479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_0/index_4166_seed_4479_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": -0.11119849886745214, + "vlm_logprob": -0.11119849886745214, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.11119849886745214, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2527216374874115 + ], + "opposite_logprob_tokens": [ + -1.5027216672897339 + ], + "logprob_reward": -0.2527216374874115, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2527216374874115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05144976079463959, + 0.05126643553376198, + 0.15057983994483948, + 0.0645151138305664, + 0.366272509098053, + 0.2394004613161087, + 0.007720280438661575, + 0.006299628410488367, + 0.009703547693789005, + 0.05279240757226944 + ], + "credit_map": [ + 0.012964151799678802, + 0.004129437264055014, + 0.009257092140614986, + 0.005082421936094761, + 0.003360313829034567, + 0.006677380762994289, + 0.005356483161449432, + 0.004622484557330608, + 0.009019846096634865, + 0.004757100250571966, + 0.008085614070296288, + 0.005253449082374573, + 0.005236527416855097, + 0.006651578936725855, + 0.0046235243789851665, + 0.00763879157602787, + 0.03753991052508354, + 0.00981663167476654, + 0.02987576834857464, + 0.00934458989650011, + 0.010266106575727463, + 0.028341922909021378, + 0.011812527664005756, + 0.0135823804885149, + 0.013121092692017555, + 0.007532436866313219, + 0.009756969287991524, + 0.006839968729764223, + 0.007248405367136002, + 0.008226527832448483, + 0.005934464279562235, + 0.00585524644702673, + 0.07779181003570557, + 0.026244590058922768, + 0.0599646233022213, + 0.02748262882232666, + 0.03509240597486496, + 0.05284832417964935, + 0.05719476565718651, + 0.029653353616595268, + 0.07026326656341553, + 0.01433609053492546, + 0.05016184598207474, + 0.011518455110490322, + 0.015924064442515373, + 0.03595080226659775, + 0.02421962469816208, + 0.017026299610733986, + 0.0018077155109494925, + 0.0004692385555244982, + 0.0015720304800197482, + 0.0004819949099328369, + 0.0005531738861463964, + 0.001339731621555984, + 0.0007501394720748067, + 0.000746255973353982, + 0.0010899631306529045, + 0.0006553989951498806, + 0.0008655102574266493, + 0.0006035827682353556, + 0.000825919327326119, + 0.0008088467875495553, + 0.0006393752410076559, + 0.0008110319613479078, + 0.0012266954872757196, + 0.0010284637100994587, + 0.0012060232693329453, + 0.0007677770918235183, + 0.0018480215221643448, + 0.0012404666049405932, + 0.0008299234323203564, + 0.001556176575832069, + 0.011967738159000874, + 0.003101362381130457, + 0.009170674718916416, + 0.004611216951161623, + 0.003840374993160367, + 0.007818699814379215, + 0.00749979168176651, + 0.004782553296536207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.2527216374874115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054989274591207504 + ], + "opposite_logprob_tokens": [ + -2.9299893379211426 + ], + "logprob_reward": -0.054989274591207504, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054989274591207504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12060027569532394, + 0.14492186903953552, + 0.10365190356969833, + 0.04567346349358559, + 0.24261295795440674, + 0.1904030293226242, + 0.032934051007032394, + 0.015735376626253128, + 0.038135431706905365, + 0.06533161550760269 + ], + "credit_map": [ + 0.032518357038497925, + 0.00929789338260889, + 0.020776977762579918, + 0.010158644057810307, + 0.00975253526121378, + 0.014211932197213173, + 0.013875088654458523, + 0.01000887155532837, + 0.02473694086074829, + 0.015956617891788483, + 0.02078574150800705, + 0.013502722606062889, + 0.01589423231780529, + 0.02271345630288124, + 0.009270141832530499, + 0.022062044590711594, + 0.02212897688150406, + 0.010962367989122868, + 0.0180423054844141, + 0.007939941249787807, + 0.008687430992722511, + 0.019077032804489136, + 0.007376269903033972, + 0.009437591768801212, + 0.009828257374465466, + 0.0045737046748399734, + 0.006915991194546223, + 0.004810452461242676, + 0.005051982589066029, + 0.006037557031959295, + 0.004175586625933647, + 0.004279934801161289, + 0.051527079194784164, + 0.017103329300880432, + 0.039087336510419846, + 0.018786700442433357, + 0.022786371409893036, + 0.03298291936516762, + 0.0409560389816761, + 0.01938319392502308, + 0.052993327379226685, + 0.012383820489048958, + 0.03800315037369728, + 0.009540438652038574, + 0.012789584696292877, + 0.029732055962085724, + 0.02141072414815426, + 0.013549938797950745, + 0.007771571632474661, + 0.003103263210505247, + 0.004761851858347654, + 0.002552835736423731, + 0.003948193043470383, + 0.00547070475295186, + 0.0027808386366814375, + 0.002544794697314501, + 0.002812259830534458, + 0.00155651627574116, + 0.002350493101403117, + 0.0014876769855618477, + 0.0020994425285607576, + 0.001930915517732501, + 0.0015684247482568026, + 0.0019296498503535986, + 0.004789234604686499, + 0.004839673172682524, + 0.00490647554397583, + 0.003788173198699951, + 0.007209980394691229, + 0.004399375524371862, + 0.003874977817758918, + 0.004327545873820782, + 0.011514929123222828, + 0.005260584410279989, + 0.008555999025702477, + 0.007845746353268623, + 0.005180755164474249, + 0.007479298859834671, + 0.011483012698590755, + 0.008011302910745144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.054989274591207504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380881205201149 + ], + "opposite_logprob_tokens": [ + -3.288088083267212 + ], + "logprob_reward": -0.0380881205201149, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380881205201149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17596888542175293, + 0.12787850201129913, + 0.08815564960241318, + 0.0806155577301979, + 0.0952828973531723, + 0.18335311114788055, + 0.02859988994896412, + 0.0661916732788086, + 0.09678258001804352, + 0.05717120319604874 + ], + "credit_map": [ + 0.035721711814403534, + 0.018617935478687286, + 0.025151509791612625, + 0.016069870442152023, + 0.012509971857070923, + 0.030894862487912178, + 0.01928051747381687, + 0.017722509801387787, + 0.018572324886918068, + 0.015940850600600243, + 0.016714634373784065, + 0.010563159361481667, + 0.02301955781877041, + 0.019501609727740288, + 0.009674004279077053, + 0.013892379589378834, + 0.011050725355744362, + 0.00937639083713293, + 0.011097300797700882, + 0.010590305551886559, + 0.011583598330616951, + 0.014422441832721233, + 0.008131979033350945, + 0.011902912519872189, + 0.016805753111839294, + 0.007748532108962536, + 0.011304464191198349, + 0.00646529346704483, + 0.011853058822453022, + 0.011813070625066757, + 0.006488661747425795, + 0.008136728778481483, + 0.022555522620677948, + 0.00728125823661685, + 0.015142274089157581, + 0.008143080398440361, + 0.008916343562304974, + 0.01242123730480671, + 0.012732444331049919, + 0.00809074379503727, + 0.05301682651042938, + 0.012062698602676392, + 0.036527350544929504, + 0.009649451822042465, + 0.012336237356066704, + 0.02796083316206932, + 0.018756059929728508, + 0.013043666258454323, + 0.006104056257754564, + 0.0027548500802367926, + 0.003912496380507946, + 0.0022002144251018763, + 0.0033431597985327244, + 0.005682428367435932, + 0.0018973962869495153, + 0.0027052899822592735, + 0.010636874474585056, + 0.006682353559881449, + 0.010251601226627827, + 0.007258289027959108, + 0.007947005331516266, + 0.008554494008421898, + 0.006404132582247257, + 0.00845692865550518, + 0.011335106566548347, + 0.012843428179621696, + 0.013441720046103, + 0.008896729908883572, + 0.019498713314533234, + 0.011267220601439476, + 0.008644516579806805, + 0.010855147615075111, + 0.011730623431503773, + 0.005216375924646854, + 0.009196899831295013, + 0.006982884369790554, + 0.004176400601863861, + 0.006925913505256176, + 0.005428604781627655, + 0.00751350587233901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0380881205201149 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155435800552368 + ], + "opposite_logprob_tokens": [ + -1.701554298400879 + ], + "logprob_reward": -0.20155435800552368, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155435800552368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11586860567331314, + 0.0881904736161232, + 0.0824243500828743, + 0.061839766800403595, + 0.1312970519065857, + 0.08990436792373657, + 0.027616143226623535, + 0.13184501230716705, + 0.2320433259010315, + 0.0389709435403347 + ], + "credit_map": [ + 0.026948288083076477, + 0.010630333796143532, + 0.01681259460747242, + 0.008840218186378479, + 0.01887996308505535, + 0.011304618790745735, + 0.011661701835691929, + 0.010790874250233173, + 0.011760727502405643, + 0.011587590910494328, + 0.011555804871022701, + 0.008805211633443832, + 0.01610209420323372, + 0.00993068516254425, + 0.008474159985780716, + 0.009974192827939987, + 0.01832646317780018, + 0.008650451898574829, + 0.01375068910419941, + 0.005951858591288328, + 0.009923029690980911, + 0.010179081000387669, + 0.00788399949669838, + 0.007758768275380135, + 0.007796506397426128, + 0.00818778108805418, + 0.008689402602612972, + 0.005193146876990795, + 0.013635302893817425, + 0.0073063126765191555, + 0.004157614428550005, + 0.006873690988868475, + 0.028068456798791885, + 0.010636118240654469, + 0.020428355783224106, + 0.008298047818243504, + 0.01409707311540842, + 0.019929053261876106, + 0.01897464692592621, + 0.010865285061299801, + 0.025090979412198067, + 0.005885824095457792, + 0.018257062882184982, + 0.004498218186199665, + 0.00641381461173296, + 0.014540916308760643, + 0.0077672721818089485, + 0.0074502695351839066, + 0.005764595698565245, + 0.002485644770786166, + 0.0046674227342009544, + 0.002097328659147024, + 0.003460567444562912, + 0.0044973790645599365, + 0.0020367561373859644, + 0.0026064442936331034, + 0.0227680541574955, + 0.012078600935637951, + 0.02015489898622036, + 0.0122894998639822, + 0.02447843924164772, + 0.013161643408238888, + 0.013592619448900223, + 0.01332124974578619, + 0.02524562180042267, + 0.0277719683945179, + 0.032767463475465775, + 0.02004718780517578, + 0.049788668751716614, + 0.0242938045412302, + 0.018506120890378952, + 0.03362245857715607, + 0.007304167374968529, + 0.0033771295566111803, + 0.0052729034796357155, + 0.003954633139073849, + 0.0033915138337761164, + 0.005888964980840683, + 0.004540862515568733, + 0.005240766331553459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.20155435800552368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, unadulterated white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a child-like bullfrog rendered in simple, bold black lines without color fill or shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullfrog's body lean back slightly with limbs splayed outward in a tense pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bullfrog's cheeks bulging dramatically to the sides forming two perfectly round air pockets larger than its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bullfrog's mouth gaped into a black O shape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863910373300314 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.00863910373300314, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863910373300314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09518472105264664, + 0.0724266767501831, + 0.09426089376211166, + 0.05970999598503113, + 0.27585893869400024, + 0.24132557213306427, + 0.01200516615062952, + 0.021089915186166763, + 0.08627524226903915, + 0.04186286777257919 + ], + "credit_map": [ + 0.01855887845158577, + 0.00824848935008049, + 0.015749836340546608, + 0.01197502389550209, + 0.007932350039482117, + 0.010220295749604702, + 0.012112515978515148, + 0.010387331247329712, + 0.014340043067932129, + 0.006573401857167482, + 0.010912950150668621, + 0.007242996245622635, + 0.008416893891990185, + 0.009335750713944435, + 0.005515760742127895, + 0.010088879615068436, + 0.02088453620672226, + 0.005935682449489832, + 0.018090730533003807, + 0.00703687546774745, + 0.0065064458176493645, + 0.01771874725818634, + 0.008352275937795639, + 0.009735602885484695, + 0.012658130377531052, + 0.005880707874894142, + 0.00945532787591219, + 0.005601540673524141, + 0.007512979209423065, + 0.007866492494940758, + 0.005216953344643116, + 0.0055178627371788025, + 0.060145024210214615, + 0.019155314192175865, + 0.04597415775060654, + 0.019608190283179283, + 0.025705531239509583, + 0.04045227915048599, + 0.042799949645996094, + 0.022018469870090485, + 0.06877516210079193, + 0.015283738262951374, + 0.04890989139676094, + 0.011951999738812447, + 0.016318464651703835, + 0.03778737410902977, + 0.025114912539720535, + 0.017184020951390266, + 0.0029895894695073366, + 0.00098518212325871, + 0.002290791366249323, + 0.0007994209881871939, + 0.0009713097242638469, + 0.0013285005697980523, + 0.0014291841071099043, + 0.0012111873365938663, + 0.0037553803995251656, + 0.0017195763066411018, + 0.0031032150145620108, + 0.0018318233778700233, + 0.004253133665770292, + 0.0024385827127844095, + 0.0018522075843065977, + 0.0021359974052757025, + 0.008113838732242584, + 0.008840853348374367, + 0.011863249354064465, + 0.009784121066331863, + 0.014257251285016537, + 0.008171824738383293, + 0.009469417855143547, + 0.01577468402683735, + 0.007551900111138821, + 0.003150870092213154, + 0.005905925761908293, + 0.004815930034965277, + 0.003915758337825537, + 0.005328979343175888, + 0.0063296351581811905, + 0.004863869398832321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.00863910373300314 + } + ], + "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/0000128/group_1/index_tensor([4968, 4968])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_1/index_tensor([4968, 4968])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9e1713a13164fac958fcc843e76d5829c2223724 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_1/index_tensor([4968, 4968])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping aerial shot unfolds, the camera slowly panning right, overlooking a majestic mountain range blanketed in white snow. In the soft golden morning sunlight, two parallel railroad tracks, sharply outlined against the thick snow, wind through a sparse, frost-covered coniferous forest. Two classic black cast-iron steam locomotives, one heading deep into the frame, the other approaching, their bodies reflecting a faint metallic sheen in the sunlight. As they precisely intersect in the center of the frame, their massive drive wheels and complex linkages operate steadily and powerfully, while thick, billowing white steam billows from their towering chimneys, rapidly condensing into spectacular clouds in the frigid air. The slanting sunlight gave the steam cloud a warm golden edge, which contrasted sharply with the long blue shadows on the snow. The whole scene, in an epic magnificence and tranquility, showed the harmonious blend of the power of the industrial age and the beauty of nature, creating a magnificent and slightly nostalgic atmosphere.", + "seed": 20411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_1/index_4968_seed_20411_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": -0.07423583790659904, + "vlm_logprob": -0.07423583790659904, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.07423583790659904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19536429643630981 + ], + "opposite_logprob_tokens": [ + -1.820364236831665 + ], + "logprob_reward": -0.19536429643630981, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19536429643630981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08797746151685715, + 0.08244319260120392, + 0.056525737047195435, + 0.1936480700969696, + 0.19629770517349243, + 0.04634682089090347, + 0.07924671471118927, + 0.04495307430624962, + 0.1285310536623001, + 0.0840301662683487 + ], + "credit_map": [ + 0.011312160640954971, + 0.01275543961673975, + 0.009652805514633656, + 0.01254713349044323, + 0.008784088306128979, + 0.011609950102865696, + 0.00826991070061922, + 0.0130459638312459, + 0.014177244156599045, + 0.011726374737918377, + 0.017345402389764786, + 0.009110217913985252, + 0.0082955751568079, + 0.006927328649908304, + 0.006521589122712612, + 0.008339460007846355, + 0.007562353275716305, + 0.00798374880105257, + 0.005894807167351246, + 0.008601565845310688, + 0.006119508296251297, + 0.0056910947896540165, + 0.005903333425521851, + 0.0087693240493536, + 0.038095295429229736, + 0.021599600091576576, + 0.022026536986231804, + 0.02072160132229328, + 0.01827499456703663, + 0.024243546649813652, + 0.019388193264603615, + 0.029298309236764908, + 0.04989994689822197, + 0.008993596769869328, + 0.011002638377249241, + 0.011263909749686718, + 0.01370542123913765, + 0.011024451814591885, + 0.07344917207956314, + 0.016958551481366158, + 0.007589937187731266, + 0.005222782958298922, + 0.00540669122710824, + 0.007941062562167645, + 0.006766486447304487, + 0.004097635857760906, + 0.0060678645968437195, + 0.003254355164244771, + 0.013232397846877575, + 0.009434059262275696, + 0.006690193898975849, + 0.007408876437693834, + 0.013486565090715885, + 0.010942673310637474, + 0.010450921021401882, + 0.007601021323353052, + 0.007076808251440525, + 0.0053457035683095455, + 0.0033947215415537357, + 0.00508597306907177, + 0.007311863359063864, + 0.005825443658977747, + 0.00689725810661912, + 0.004015299491584301, + 0.011711184866726398, + 0.0208715982735157, + 0.01382885780185461, + 0.01371355913579464, + 0.014353561215102673, + 0.01173006184399128, + 0.02840113267302513, + 0.013921093195676804, + 0.00844948273152113, + 0.012396234087646008, + 0.005933478008955717, + 0.004433446563780308, + 0.02119699865579605, + 0.007982266135513783, + 0.009284808300435543, + 0.014353450387716293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.19536429643630981 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036779195070267 + ], + "opposite_logprob_tokens": [ + -2.350367784500122 + ], + "logprob_reward": -0.10036779195070267, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036779195070267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040348418056964874, + 0.08957222104072571, + 0.06310237944126129, + 0.06247122585773468, + 0.358755499124527, + 0.04003285989165306, + 0.04771030321717262, + 0.026677202433347702, + 0.06408490985631943, + 0.20724499225616455 + ], + "credit_map": [ + 0.004675291944295168, + 0.00747370719909668, + 0.004835342522710562, + 0.007082206662744284, + 0.0033677590545266867, + 0.0036303969100117683, + 0.004082824569195509, + 0.005200887098908424, + 0.019805198535323143, + 0.010590183548629284, + 0.023992439731955528, + 0.010070839896798134, + 0.009041267447173595, + 0.004962816834449768, + 0.005280977580696344, + 0.005828500725328922, + 0.012904766015708447, + 0.007472538389265537, + 0.006441469304263592, + 0.004631411284208298, + 0.007023666054010391, + 0.00557662034407258, + 0.0113630136474967, + 0.0076888916082680225, + 0.012356474064290524, + 0.008303900249302387, + 0.005645656492561102, + 0.004324937704950571, + 0.008753214031457901, + 0.009343137964606285, + 0.004771050531417131, + 0.008972857147455215, + 0.09119674563407898, + 0.012943430803716183, + 0.020432885736227036, + 0.01689881831407547, + 0.02327222377061844, + 0.015409220941364765, + 0.14416509866714478, + 0.03443704545497894, + 0.006911103148013353, + 0.006559889763593674, + 0.0029692798852920532, + 0.003925895318388939, + 0.007089175283908844, + 0.006747214123606682, + 0.003148440970107913, + 0.0026818576734513044, + 0.006796292494982481, + 0.005311730783432722, + 0.0031981240026652813, + 0.004462315700948238, + 0.008459464646875858, + 0.00896652415394783, + 0.005793591495603323, + 0.004722259473055601, + 0.004005364142358303, + 0.003542996244505048, + 0.0021575111895799637, + 0.001941160298883915, + 0.005811340641230345, + 0.004141880664974451, + 0.0028585295658558607, + 0.0022184194531291723, + 0.012416856363415718, + 0.009460990317165852, + 0.004620993044227362, + 0.00736198341473937, + 0.0063660466112196445, + 0.012170521542429924, + 0.005198323167860508, + 0.006489197723567486, + 0.020852848887443542, + 0.03017422743141651, + 0.013509665615856647, + 0.006896673236042261, + 0.060312844812870026, + 0.020145846530795097, + 0.017155537381768227, + 0.038197360932826996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.10036779195070267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067168444395065 + ], + "opposite_logprob_tokens": [ + -3.168067216873169 + ], + "logprob_reward": -0.043067168444395065, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067168444395065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07751021534204483, + 0.06856254488229752, + 0.056061096489429474, + 0.06384104490280151, + 0.260100394487381, + 0.03256608545780182, + 0.036485426127910614, + 0.029097450897097588, + 0.04457715153694153, + 0.3311986029148102 + ], + "credit_map": [ + 0.01024781446903944, + 0.015271020121872425, + 0.00833157543092966, + 0.012499505653977394, + 0.011988082900643349, + 0.006598929408937693, + 0.005338369403034449, + 0.007234914228320122, + 0.012386777438223362, + 0.01076729316264391, + 0.01220797561109066, + 0.011473683640360832, + 0.008747700601816177, + 0.0034420015290379524, + 0.005292822606861591, + 0.004244283307343721, + 0.0077954544685781, + 0.00862942822277546, + 0.006194981746375561, + 0.007518323604017496, + 0.007625214289873838, + 0.007297792471945286, + 0.005827842280268669, + 0.005172059405595064, + 0.014938210137188435, + 0.007969736121594906, + 0.009436347521841526, + 0.00822480209171772, + 0.00418118154630065, + 0.0070803784765303135, + 0.0048530809581279755, + 0.007157310377806425, + 0.048216674476861954, + 0.01751849055290222, + 0.02139252796769142, + 0.02582917921245098, + 0.0181113351136446, + 0.010215534828603268, + 0.09361416846513748, + 0.025202468037605286, + 0.004834971856325865, + 0.004869105760008097, + 0.003792696865275502, + 0.006709073670208454, + 0.0036808832082897425, + 0.002954094670712948, + 0.0035401606000959873, + 0.0021850960329174995, + 0.005777033977210522, + 0.00431815255433321, + 0.0026094913482666016, + 0.0046440535224974155, + 0.006241133436560631, + 0.005005888640880585, + 0.004931988660246134, + 0.0029576816596090794, + 0.005388438235968351, + 0.0028026914224028587, + 0.0038572857156395912, + 0.0029558155220001936, + 0.004532639868557453, + 0.004347593057900667, + 0.00331204105168581, + 0.001900946139357984, + 0.00651208171620965, + 0.005303467623889446, + 0.004782519768923521, + 0.005395174957811832, + 0.00523058557882905, + 0.005497826263308525, + 0.007652693428099155, + 0.004202804993838072, + 0.02616390399634838, + 0.04930531978607178, + 0.015300283208489418, + 0.012426061555743217, + 0.10818979889154434, + 0.0358705036342144, + 0.016752144321799278, + 0.06719060987234116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.043067168444395065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200471341609955 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200471341609955, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200471341609955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04414643347263336, + 0.022582823410630226, + 0.06352292746305466, + 0.06008065491914749, + 0.4225645959377289, + 0.024708271026611328, + 0.026003560051321983, + 0.02983826771378517, + 0.037414588034152985, + 0.2691379189491272 + ], + "credit_map": [ + 0.005802930798381567, + 0.006392974406480789, + 0.004434382542967796, + 0.00623743049800396, + 0.007728144992142916, + 0.005145195405930281, + 0.003704866860061884, + 0.004700507037341595, + 0.003041306044906378, + 0.0033639606554061174, + 0.004011640790849924, + 0.003329448401927948, + 0.0027282144874334335, + 0.0019069061381742358, + 0.0020506924483925104, + 0.0021506536286324263, + 0.006351207382977009, + 0.009708324447274208, + 0.00489258486777544, + 0.007214091718196869, + 0.0051567587070167065, + 0.0056334552355110645, + 0.012830459512770176, + 0.011736046522855759, + 0.010539757087826729, + 0.012081731110811234, + 0.006434567738324404, + 0.005571888294070959, + 0.00703422212973237, + 0.00906628929078579, + 0.005204109940677881, + 0.004148092120885849, + 0.10868887603282928, + 0.019619472324848175, + 0.023902541026473045, + 0.026207251474261284, + 0.026824329048395157, + 0.01493890956044197, + 0.16462840139865875, + 0.037754788994789124, + 0.003356975270435214, + 0.0034418560098856688, + 0.0019882600754499435, + 0.0028623882681131363, + 0.0035673489328473806, + 0.0038716518320143223, + 0.002821114147081971, + 0.002798675559461117, + 0.00400301069021225, + 0.0028624783735722303, + 0.0018088624347001314, + 0.0031901129987090826, + 0.004547672346234322, + 0.004063657019287348, + 0.0033644125796854496, + 0.0021633529104292393, + 0.00513407913967967, + 0.003497048281133175, + 0.004985107108950615, + 0.0022309059277176857, + 0.006460092030465603, + 0.0036628814414143562, + 0.0023259317968040705, + 0.0015422250144183636, + 0.005618376657366753, + 0.0049825371243059635, + 0.00436239130795002, + 0.0029970528557896614, + 0.0034038820303976536, + 0.007384267635643482, + 0.003978129476308823, + 0.0046879504807293415, + 0.021000025793910027, + 0.04210122674703598, + 0.012847666628658772, + 0.008564089424908161, + 0.0872679278254509, + 0.027772080153226852, + 0.014827394858002663, + 0.05475752055644989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.014200471341609955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179461359977722 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.018179461359977722, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018179461359977722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05952443927526474, + 0.041900962591171265, + 0.044068269431591034, + 0.014470070600509644, + 0.037885431200265884, + 0.021573500707745552, + 0.019825169816613197, + 0.03954051434993744, + 0.059373367577791214, + 0.6618383526802063 + ], + "credit_map": [ + 0.007787234615534544, + 0.010298240929841995, + 0.006456020753830671, + 0.012654545716941357, + 0.0059612649492919445, + 0.00559272151440382, + 0.004603726789355278, + 0.006170677486807108, + 0.0111016146838665, + 0.003537900745868683, + 0.014455871656537056, + 0.003943385556340218, + 0.0023056373465806246, + 0.0024913507513701916, + 0.002100336132571101, + 0.0019648598972707987, + 0.005903246812522411, + 0.0065390728414058685, + 0.0051523251459002495, + 0.009562499821186066, + 0.004343814216554165, + 0.0033341445960104465, + 0.0045631397515535355, + 0.004670017398893833, + 0.0023152045905590057, + 0.0014896178618073463, + 0.0015446109464392066, + 0.0017802207730710506, + 0.0014832611195743084, + 0.0027708832640200853, + 0.0014011214952915907, + 0.0016851479886099696, + 0.005931629333645105, + 0.004599433857947588, + 0.0029001084621995687, + 0.004989258944988251, + 0.005850797984749079, + 0.005301335360854864, + 0.005710742902010679, + 0.0026021201629191637, + 0.0033260351046919823, + 0.002694804919883609, + 0.0016543874517083168, + 0.0026045034173876047, + 0.004984632134437561, + 0.0025864243507385254, + 0.002130315639078617, + 0.001592395594343543, + 0.00309464312158525, + 0.0020781781058758497, + 0.0017008825670927763, + 0.001950963749550283, + 0.003321111435070634, + 0.004073354881256819, + 0.001925383461639285, + 0.0016806499334052205, + 0.006629315670579672, + 0.00444798031821847, + 0.004508187994360924, + 0.004865655675530434, + 0.0075288619846105576, + 0.005483204033225775, + 0.004047986585646868, + 0.0020293176639825106, + 0.010570860467851162, + 0.010629107244312763, + 0.004310889169573784, + 0.0031598855275660753, + 0.009881226345896721, + 0.010411202907562256, + 0.004588347859680653, + 0.0058218431659042835, + 0.056345488876104355, + 0.10493281483650208, + 0.03278814256191254, + 0.02121029421687126, + 0.2121400088071823, + 0.06227418780326843, + 0.043731674551963806, + 0.12841567397117615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.018179461359977722 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial shot unfolds, the camera slowly panning right, overlooking a majestic mountain range blanketed in white snow. In the soft golden morning sunlight, two parallel railroad tracks, sharply outlined against the thick snow, wind through a sparse, frost-covered coniferous forest. Two classic black cast-iron steam locomotives, one heading deep into the frame, the other approaching, their bodies reflecting a faint metallic sheen in the sunlight. As they precisely intersect in the center of the frame, their massive drive wheels and complex linkages operate steadily and powerfully, while thick, billowing white steam billows from their towering chimneys, rapidly condensing into spectacular clouds in the frigid air. The slanting sunlight gave the steam cloud a warm golden edge, which contrasted sharply with the long blue shadows on the snow. The whole scene, in an epic magnificence and tranquility, showed the harmonious blend of the power of the industrial age and the beauty of nature, creating a magnificent and slightly nostalgic atmosphere.", + "seed": 20412, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_1/index_4968_seed_20412_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": -0.0598272942006588, + "vlm_logprob": -0.0598272942006588, + "vlm_yes_no_margin": 3.374999976158142, + "vlm_token_credit": -0.0598272942006588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15548381209373474 + ], + "opposite_logprob_tokens": [ + -2.1554837226867676 + ], + "logprob_reward": -0.15548381209373474, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.15548381209373474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06729193031787872, + 0.3995536267757416, + 0.04102463647723198, + 0.08500780165195465, + 0.14412133395671844, + 0.04075230285525322, + 0.08825282007455826, + 0.04162496700882912, + 0.02790132723748684, + 0.06446923315525055 + ], + "credit_map": [ + 0.006345128174871206, + 0.006087271962314844, + 0.01259810384362936, + 0.010197967290878296, + 0.006618835031986237, + 0.007879362441599369, + 0.00950518250465393, + 0.008060079999268055, + 0.030892441049218178, + 0.015475261956453323, + 0.08981776982545853, + 0.16436541080474854, + 0.009932373650372028, + 0.018607666715979576, + 0.037826042622327805, + 0.03263666108250618, + 0.002966355998069048, + 0.0036115283146500587, + 0.010517612099647522, + 0.0066559044644236565, + 0.0026891634333878756, + 0.004595001228153706, + 0.004906064365059137, + 0.0050830054096877575, + 0.006863532587885857, + 0.006569890305399895, + 0.015561090782284737, + 0.013701756484806538, + 0.007173564750701189, + 0.01287001557648182, + 0.010354464873671532, + 0.011913487687706947, + 0.00807647593319416, + 0.005832856986671686, + 0.03687574341893196, + 0.008761376142501831, + 0.008228572085499763, + 0.020551491528749466, + 0.0211133174598217, + 0.03468150645494461, + 0.004030809737741947, + 0.004935832228511572, + 0.006378957536071539, + 0.005602913908660412, + 0.005164965055882931, + 0.00518350675702095, + 0.004472574684768915, + 0.004982744809240103, + 0.008025816641747952, + 0.018103528767824173, + 0.014222814701497555, + 0.010076438076794147, + 0.010287833400070667, + 0.008447603322565556, + 0.005723634269088507, + 0.013365155085921288, + 0.0053480640053749084, + 0.008286968804895878, + 0.006835958454757929, + 0.006156386341899633, + 0.004909252282232046, + 0.004001534078270197, + 0.0031280959956347942, + 0.002958705648779869, + 0.0029319366440176964, + 0.0017394552705809474, + 0.003678947454318404, + 0.0017318861791864038, + 0.0028226219583302736, + 0.005902604199945927, + 0.0015361045952886343, + 0.007557770702987909, + 0.008655195124447346, + 0.005363364703953266, + 0.0053656031377613544, + 0.006613162811845541, + 0.017290208488702774, + 0.009314915165305138, + 0.006387484259903431, + 0.005479305051267147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.15548381209373474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908209949731827 + ], + "opposite_logprob_tokens": [ + -2.4640820026397705 + ], + "logprob_reward": -0.08908209949731827, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908209949731827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04868905246257782, + 0.15735505521297455, + 0.2022131383419037, + 0.13217021524906158, + 0.11113213747739792, + 0.07454289495944977, + 0.11281390488147736, + 0.04655316099524498, + 0.04284358397126198, + 0.07168689370155334 + ], + "credit_map": [ + 0.0065457457676529884, + 0.004093029070645571, + 0.006235125008970499, + 0.005906947422772646, + 0.0026439125649631023, + 0.008041176944971085, + 0.005803912412375212, + 0.00941920280456543, + 0.025878943502902985, + 0.007099775597453117, + 0.025203894823789597, + 0.042265377938747406, + 0.006898409221321344, + 0.015510397963225842, + 0.01992054097354412, + 0.014577717520296574, + 0.027670882642269135, + 0.010115181095898151, + 0.03550249710679054, + 0.024608179926872253, + 0.011759932152926922, + 0.051344383507966995, + 0.019317755475640297, + 0.021894332021474838, + 0.008771040476858616, + 0.010878680273890495, + 0.021335601806640625, + 0.022978950291872025, + 0.010387134738266468, + 0.021256206557154655, + 0.019446255639195442, + 0.01711633801460266, + 0.00572127103805542, + 0.005743755027651787, + 0.014023064635694027, + 0.01036075223237276, + 0.007528678514063358, + 0.022645553573966026, + 0.029864301905035973, + 0.015244762413203716, + 0.007099552545696497, + 0.007170898374170065, + 0.011467165313661098, + 0.006207093130797148, + 0.009725476615130901, + 0.017795316874980927, + 0.006523044314235449, + 0.008554345928132534, + 0.010801499709486961, + 0.011862412095069885, + 0.015586075372993946, + 0.01329189632087946, + 0.014454434625804424, + 0.013091832399368286, + 0.009276819415390491, + 0.02444893680512905, + 0.004924677312374115, + 0.007124058902263641, + 0.007166064810007811, + 0.006515091750770807, + 0.005947809666395187, + 0.00716678611934185, + 0.003243313869461417, + 0.00446535786613822, + 0.0046594878658652306, + 0.004517000634223223, + 0.007378019858151674, + 0.004182412289083004, + 0.007568638306111097, + 0.0054872483015060425, + 0.005057846661657095, + 0.003992932382971048, + 0.009259724989533424, + 0.006903903558850288, + 0.006787369027733803, + 0.008920491673052311, + 0.016528427600860596, + 0.00939773116260767, + 0.008636018261313438, + 0.005253229755908251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.08908209949731827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371158614754677 + ], + "opposite_logprob_tokens": [ + -3.4087116718292236 + ], + "logprob_reward": -0.03371158614754677, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371158614754677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0636758953332901, + 0.4262997806072235, + 0.09699863940477371, + 0.08075322210788727, + 0.10151490569114685, + 0.05298616737127304, + 0.04726735129952431, + 0.024373186752200127, + 0.030869098380208015, + 0.07526173442602158 + ], + "credit_map": [ + 0.006387015338987112, + 0.007183561567217112, + 0.007011005189269781, + 0.00907981302589178, + 0.00459628738462925, + 0.011809511110186577, + 0.00684129586443305, + 0.010767406783998013, + 0.03568490967154503, + 0.016115164384245872, + 0.1108417734503746, + 0.15230338275432587, + 0.01388621237128973, + 0.01937822625041008, + 0.04270928353071213, + 0.035380855202674866, + 0.005407487042248249, + 0.006905864924192429, + 0.017793994396924973, + 0.02102476730942726, + 0.006247164215892553, + 0.007310580927878618, + 0.021651994436979294, + 0.010656788945198059, + 0.004779597278684378, + 0.006684913765639067, + 0.01458279974758625, + 0.014459537342190742, + 0.0064121936447918415, + 0.00800891499966383, + 0.012993617914617062, + 0.012831650674343109, + 0.0062127201817929745, + 0.007348967250436544, + 0.02067040465772152, + 0.010731104761362076, + 0.008066680282354355, + 0.013476520776748657, + 0.013860195875167847, + 0.021148309111595154, + 0.006016010884195566, + 0.00926316250115633, + 0.008182791993021965, + 0.008182885125279427, + 0.006860647816210985, + 0.005493940785527229, + 0.004590345546603203, + 0.004396379459649324, + 0.005410327576100826, + 0.008801119402050972, + 0.008150694891810417, + 0.005353088956326246, + 0.00639333575963974, + 0.004286638926714659, + 0.00369122251868248, + 0.0051809209398925304, + 0.002862179884687066, + 0.0037222332321107388, + 0.0032539027743041515, + 0.0034845115151256323, + 0.0033634311985224485, + 0.0031568778213113546, + 0.002269844291731715, + 0.0022602055687457323, + 0.003734670579433441, + 0.0037117425817996264, + 0.0033436070661991835, + 0.0034588747657835484, + 0.0062917605973780155, + 0.004014947917312384, + 0.003516723634675145, + 0.002796770306304097, + 0.007283326704055071, + 0.009987587109208107, + 0.009374449029564857, + 0.012976990081369877, + 0.009127787314355373, + 0.013494719751179218, + 0.006987467873841524, + 0.006029404234141111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.03371158614754677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082051321864128 + ], + "opposite_logprob_tokens": [ + -4.511082172393799 + ], + "logprob_reward": -0.011082051321864128, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082051321864128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038045257329940796, + 0.17900238931179047, + 0.051461733877658844, + 0.08734258264303207, + 0.2689763903617859, + 0.1549111306667328, + 0.03628065437078476, + 0.08439202606678009, + 0.048414621502161026, + 0.05117321014404297 + ], + "credit_map": [ + 0.0038072289898991585, + 0.003940461203455925, + 0.004644175060093403, + 0.00587182492017746, + 0.002202490111812949, + 0.007504749111831188, + 0.0034935700241476297, + 0.006580756511539221, + 0.012719693593680859, + 0.006532054860144854, + 0.0500037744641304, + 0.06324339658021927, + 0.007163234055042267, + 0.008891028352081776, + 0.01717393659055233, + 0.013275256380438805, + 0.0055496278218925, + 0.0030314032919704914, + 0.005959099158644676, + 0.007105994038283825, + 0.0032421634532511234, + 0.013937637209892273, + 0.006167802028357983, + 0.00646800734102726, + 0.005769453477114439, + 0.007198132574558258, + 0.009921458549797535, + 0.016249200329184532, + 0.009536773897707462, + 0.013895452953875065, + 0.015644673258066177, + 0.009127438068389893, + 0.013386040925979614, + 0.012654244899749756, + 0.06632919609546661, + 0.018771052360534668, + 0.013393758796155453, + 0.03294337913393974, + 0.015151373110711575, + 0.09634733945131302, + 0.017564166337251663, + 0.026470843702554703, + 0.0229315347969532, + 0.021718356758356094, + 0.01820356398820877, + 0.020344054326415062, + 0.013350948691368103, + 0.014327655546367168, + 0.004126776475459337, + 0.005648692138493061, + 0.004436852876096964, + 0.005951225757598877, + 0.004839560482650995, + 0.0038780327886343002, + 0.0031121056526899338, + 0.004287408664822578, + 0.010764497332274914, + 0.01647309400141239, + 0.013613207265734673, + 0.012790198437869549, + 0.010339315980672836, + 0.008285591378808022, + 0.005530704744160175, + 0.00659541692584753, + 0.0041531771421432495, + 0.0030011916533112526, + 0.004974115639925003, + 0.002841998590156436, + 0.007790330797433853, + 0.011819450184702873, + 0.0038788558449596167, + 0.009955501183867455, + 0.0058236210606992245, + 0.007112292572855949, + 0.006132109556347132, + 0.007903646677732468, + 0.006091305986046791, + 0.00904140342026949, + 0.004448901861906052, + 0.004619930870831013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.011082051321864128 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776921942830086 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009776921942830086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776921942830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04533974826335907, + 0.32126954197883606, + 0.07063805311918259, + 0.04051462933421135, + 0.05822959914803505, + 0.06699643284082413, + 0.24726422131061554, + 0.04039531201124191, + 0.07403282076120377, + 0.035319630056619644 + ], + "credit_map": [ + 0.0048448508605360985, + 0.0050704642198979855, + 0.0057471212930977345, + 0.007005948573350906, + 0.00304959574714303, + 0.007039973512291908, + 0.004672867711633444, + 0.007908930070698261, + 0.0261999499052763, + 0.012527856044471264, + 0.0712205097079277, + 0.13057756423950195, + 0.00843969639390707, + 0.015378120355308056, + 0.030710509046912193, + 0.026215313002467155, + 0.006629506591707468, + 0.006574003491550684, + 0.011917077004909515, + 0.008764222264289856, + 0.006139352452009916, + 0.01196756400167942, + 0.00866515003144741, + 0.009981179609894753, + 0.0028926203958690166, + 0.002998803509399295, + 0.006952864583581686, + 0.0048665511421859264, + 0.004258934408426285, + 0.008047453127801418, + 0.004273535218089819, + 0.00622386671602726, + 0.004545378498733044, + 0.003223062725737691, + 0.01672179438173771, + 0.0036926392931491137, + 0.004299723077565432, + 0.008050105534493923, + 0.004852514714002609, + 0.012844381853938103, + 0.009935016743838787, + 0.009483585134148598, + 0.007984301075339317, + 0.007660937029868364, + 0.00907911080867052, + 0.010153220035135746, + 0.0053991698659956455, + 0.007301089819520712, + 0.028870750218629837, + 0.02107580564916134, + 0.02920583076775074, + 0.022975919768214226, + 0.04520242661237717, + 0.0506441667675972, + 0.020038072019815445, + 0.02925124391913414, + 0.004138180986046791, + 0.006537226028740406, + 0.005653904750943184, + 0.005679907277226448, + 0.00600192928686738, + 0.005486069712787867, + 0.0033801612444221973, + 0.0035179343540221453, + 0.005980427842587233, + 0.004003120586276054, + 0.00858202576637268, + 0.0037824627943336964, + 0.009606290608644485, + 0.018010837957262993, + 0.004761882126331329, + 0.019305771216750145, + 0.004101448226720095, + 0.005259887780994177, + 0.004207873251289129, + 0.005339363589882851, + 0.0051568010821938515, + 0.0043181609362363815, + 0.002870087279006839, + 0.004066006746143103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.009776921942830086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial shot unfolds, the camera slowly panning right, overlooking a majestic mountain range blanketed in white snow. In the soft golden morning sunlight, two parallel railroad tracks, sharply outlined against the thick snow, wind through a sparse, frost-covered coniferous forest. Two classic black cast-iron steam locomotives, one heading deep into the frame, the other approaching, their bodies reflecting a faint metallic sheen in the sunlight. As they precisely intersect in the center of the frame, their massive drive wheels and complex linkages operate steadily and powerfully, while thick, billowing white steam billows from their towering chimneys, rapidly condensing into spectacular clouds in the frigid air. The slanting sunlight gave the steam cloud a warm golden edge, which contrasted sharply with the long blue shadows on the snow. The whole scene, in an epic magnificence and tranquility, showed the harmonious blend of the power of the industrial age and the beauty of nature, creating a magnificent and slightly nostalgic atmosphere.", + "seed": 20413, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_1/index_4968_seed_20413_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": -0.052081137895584106, + "vlm_logprob": -0.052081137895584106, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.052081137895584106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1582019180059433 + ], + "opposite_logprob_tokens": [ + -2.0332019329071045 + ], + "logprob_reward": -0.1582019180059433, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1582019180059433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20528699457645416, + 0.09616255015134811, + 0.04525545984506607, + 0.04999247565865517, + 0.0804356038570404, + 0.13907556235790253, + 0.10532069206237793, + 0.04955402761697769, + 0.059720467776060104, + 0.16919605433940887 + ], + "credit_map": [ + 0.02047703042626381, + 0.016713468357920647, + 0.038514696061611176, + 0.030665982514619827, + 0.011669612489640713, + 0.03288337588310242, + 0.01515096053481102, + 0.039211876690387726, + 0.011917756870388985, + 0.012526779435575008, + 0.017887823283672333, + 0.013664796017110348, + 0.00911432784050703, + 0.013204854913055897, + 0.006712650414556265, + 0.011133573018014431, + 0.004698935896158218, + 0.006315658334642649, + 0.006060606800019741, + 0.004726849030703306, + 0.004341727588325739, + 0.008006550371646881, + 0.003451972035691142, + 0.007653162349015474, + 0.007342602591961622, + 0.006184964906424284, + 0.005916321184486151, + 0.007903956808149815, + 0.004975763615220785, + 0.005614868830889463, + 0.005954449996352196, + 0.006099553778767586, + 0.010284180752933025, + 0.014334495179355145, + 0.011375650763511658, + 0.008663885295391083, + 0.011603843420743942, + 0.006696202326565981, + 0.00908281747251749, + 0.008394535630941391, + 0.024633517488837242, + 0.016918804496526718, + 0.016848420724272728, + 0.012491848319768906, + 0.01764986477792263, + 0.020259935408830643, + 0.009424437768757343, + 0.02084875851869583, + 0.012037193402647972, + 0.008329415693879128, + 0.03187021240592003, + 0.009586041793227196, + 0.009637903422117233, + 0.015318145975470543, + 0.008913863450288773, + 0.009627917781472206, + 0.00992016401141882, + 0.005771256051957607, + 0.007040038239210844, + 0.003905713092535734, + 0.007978429086506367, + 0.006291541736572981, + 0.004037226550281048, + 0.004609665833413601, + 0.008663327433168888, + 0.008000925183296204, + 0.010333162732422352, + 0.004678850527852774, + 0.00589978089556098, + 0.009308261796832085, + 0.007580365985631943, + 0.005255802534520626, + 0.014743641950190067, + 0.020576516166329384, + 0.015614183619618416, + 0.035565335303545, + 0.027823328971862793, + 0.013010307215154171, + 0.01860327087342739, + 0.023259488865733147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.1582019180059433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054996609687805176 + ], + "opposite_logprob_tokens": [ + -2.9299964904785156 + ], + "logprob_reward": -0.054996609687805176, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054996609687805176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.138209730386734, + 0.11188401281833649, + 0.054412052035331726, + 0.043389156460762024, + 0.08718511462211609, + 0.11762159317731857, + 0.04510995000600815, + 0.11465178430080414, + 0.11235415190458298, + 0.1751825213432312 + ], + "credit_map": [ + 0.015906676650047302, + 0.014748319052159786, + 0.014250453561544418, + 0.01697005145251751, + 0.013804749585688114, + 0.020169897004961967, + 0.015837933868169785, + 0.026521654799580574, + 0.010956809855997562, + 0.01225732546299696, + 0.017057014629244804, + 0.024539293721318245, + 0.008658108301460743, + 0.01804812252521515, + 0.006930605508387089, + 0.013436740264296532, + 0.006001665722578764, + 0.005967516452074051, + 0.007527423091232777, + 0.0034760357812047005, + 0.006216682493686676, + 0.012540149502456188, + 0.004042124841362238, + 0.00864045787602663, + 0.005165225826203823, + 0.005803652573376894, + 0.005044897552579641, + 0.004239109810441732, + 0.006035546772181988, + 0.007447062525898218, + 0.004065407440066338, + 0.005588251166045666, + 0.01134453434497118, + 0.014479855075478554, + 0.009527512826025486, + 0.007958559319376945, + 0.010559871792793274, + 0.009193255566060543, + 0.01250719465315342, + 0.01161432545632124, + 0.012252823449671268, + 0.009951906278729439, + 0.018739836290478706, + 0.007715232670307159, + 0.012758486904203892, + 0.01952451467514038, + 0.007688143756240606, + 0.028990652412176132, + 0.006427096668630838, + 0.002548068296164274, + 0.015607420355081558, + 0.003776228055357933, + 0.002421260578557849, + 0.006777564529329538, + 0.003578868228942156, + 0.003973442129790783, + 0.023916415870189667, + 0.015307273715734482, + 0.016831396147608757, + 0.008699539117515087, + 0.014346396550536156, + 0.010474022477865219, + 0.011566770263016224, + 0.013509957119822502, + 0.013602497056126595, + 0.013264273293316364, + 0.013046729378402233, + 0.015598666854202747, + 0.009898891672492027, + 0.015889601781964302, + 0.020284872502088547, + 0.010768616572022438, + 0.015063660219311714, + 0.019128767773509026, + 0.015544114634394646, + 0.027009081095457077, + 0.04689513519406319, + 0.01598544791340828, + 0.01484230998903513, + 0.020713994279503822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.054996609687805176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355136185884476 + ], + "opposite_logprob_tokens": [ + -3.651355028152466 + ], + "logprob_reward": -0.026355136185884476, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355136185884476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1652592420578003, + 0.06913026422262192, + 0.06195662170648575, + 0.04165000468492508, + 0.07318998873233795, + 0.05653864145278931, + 0.11141593009233475, + 0.12637820839881897, + 0.13691434264183044, + 0.15756675601005554 + ], + "credit_map": [ + 0.012890820391476154, + 0.01520997192710638, + 0.043524403125047684, + 0.033441562205553055, + 0.007916907779872417, + 0.019900493323802948, + 0.017565438523888588, + 0.01480963733047247, + 0.005755425896495581, + 0.00875853095203638, + 0.008933108299970627, + 0.017074013128876686, + 0.0056993355974555016, + 0.007391932886093855, + 0.005328911356627941, + 0.010189010761678219, + 0.009914100170135498, + 0.011075105518102646, + 0.008451155386865139, + 0.006678218487650156, + 0.005945340730249882, + 0.006453369744122028, + 0.005805864930152893, + 0.007633473724126816, + 0.005009485874325037, + 0.006077030673623085, + 0.0043143038637936115, + 0.005576279945671558, + 0.0044523365795612335, + 0.005784689914435148, + 0.00460183946415782, + 0.0058340393006801605, + 0.006669260561466217, + 0.009233342483639717, + 0.008441431447863579, + 0.008972429670393467, + 0.009219841100275517, + 0.010414816439151764, + 0.010283676907420158, + 0.009955191053450108, + 0.011404681019484997, + 0.005446156021207571, + 0.008901666849851608, + 0.004198295995593071, + 0.004185029771178961, + 0.005067987833172083, + 0.0033508986234664917, + 0.013983918353915215, + 0.01661510020494461, + 0.010307414457201958, + 0.026818718761205673, + 0.010997440665960312, + 0.014115814119577408, + 0.010273827239871025, + 0.01095333881676197, + 0.011334281414747238, + 0.027545763179659843, + 0.012115288525819778, + 0.014232834801077843, + 0.013439321890473366, + 0.0165703222155571, + 0.011436435393989086, + 0.014481551013886929, + 0.01655668206512928, + 0.017961924895644188, + 0.009640848264098167, + 0.015559228137135506, + 0.015273920260369778, + 0.012667413800954819, + 0.02156057581305504, + 0.027076369151473045, + 0.017174066975712776, + 0.019644999876618385, + 0.019516954198479652, + 0.016899697482585907, + 0.02817724645137787, + 0.021485015749931335, + 0.013239681720733643, + 0.02423783950507641, + 0.014365327544510365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.026355136185884476 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1552739441394806, + 0.043131519109010696, + 0.1553071141242981, + 0.08084466308355331, + 0.16008053719997406, + 0.035891298204660416, + 0.18916136026382446, + 0.1164315715432167, + 0.022593487054109573, + 0.04128444567322731 + ], + "credit_map": [ + 0.013795343227684498, + 0.014865458942949772, + 0.0422140508890152, + 0.0334492065012455, + 0.0073447441682219505, + 0.019312145188450813, + 0.011751797050237656, + 0.012541213072836399, + 0.003067278303205967, + 0.0062667070887982845, + 0.0052774082869291306, + 0.007170646917074919, + 0.003734170226380229, + 0.005368498153984547, + 0.003804862732067704, + 0.00844194833189249, + 0.019155846908688545, + 0.03122376836836338, + 0.03624290972948074, + 0.011180764995515347, + 0.012218920513987541, + 0.014958524145185947, + 0.01243369746953249, + 0.017892686650156975, + 0.009785414673388004, + 0.010450217872858047, + 0.011810981668531895, + 0.01387244462966919, + 0.010003754869103432, + 0.007246879860758781, + 0.009731857106089592, + 0.00794310588389635, + 0.023387528955936432, + 0.030311478301882744, + 0.018805071711540222, + 0.01364833116531372, + 0.01856025494635105, + 0.02241411991417408, + 0.015667418017983437, + 0.017286337912082672, + 0.005904452875256538, + 0.0034095312003046274, + 0.005274851806461811, + 0.0033819281961768866, + 0.003371234517544508, + 0.004094231408089399, + 0.0022957914043217897, + 0.008159278891980648, + 0.03180599585175514, + 0.017199717462062836, + 0.04682746157050133, + 0.01852823980152607, + 0.022542232647538185, + 0.016357433050870895, + 0.018986325711011887, + 0.016913950443267822, + 0.02724343352019787, + 0.015530708245933056, + 0.013254686258733273, + 0.010691022500395775, + 0.0160213690251112, + 0.011126304045319557, + 0.011122015304863453, + 0.011442032642662525, + 0.003088979749009013, + 0.0022089644335210323, + 0.0026546583976596594, + 0.0028685969300568104, + 0.0016363495960831642, + 0.0038834060542285442, + 0.0034568847622722387, + 0.002795645734295249, + 0.0033581724856048822, + 0.0037055197171866894, + 0.0038914489559829235, + 0.006546391174197197, + 0.005837798584252596, + 0.004776050336658955, + 0.008167256601154804, + 0.005001810844987631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.00978388637304306 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068139225244522 + ], + "opposite_logprob_tokens": [ + -4.511068344116211 + ], + "logprob_reward": -0.011068139225244522, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068139225244522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3256712555885315, + 0.049845725297927856, + 0.03944593295454979, + 0.03888125717639923, + 0.13478274643421173, + 0.048415929079055786, + 0.04084809124469757, + 0.14487940073013306, + 0.09534366428852081, + 0.08188597857952118 + ], + "credit_map": [ + 0.03256320580840111, + 0.026221642270684242, + 0.09020107984542847, + 0.056699540466070175, + 0.012782646343111992, + 0.024833928793668747, + 0.018477903679013252, + 0.06389129906892776, + 0.004948690999299288, + 0.00662260502576828, + 0.006776980124413967, + 0.008781355805695057, + 0.004846930969506502, + 0.006115735974162817, + 0.0042371912859380245, + 0.007516240235418081, + 0.0041038901545107365, + 0.008477717638015747, + 0.009955711662769318, + 0.003615999361500144, + 0.003779731225222349, + 0.003343419637531042, + 0.003017402021214366, + 0.00315206334926188, + 0.004403961822390556, + 0.006319005973637104, + 0.006356090307235718, + 0.003945289645344019, + 0.004473081324249506, + 0.004780187271535397, + 0.003282153280451894, + 0.005321489181369543, + 0.018317634239792824, + 0.02507847733795643, + 0.01424475573003292, + 0.01461866032332182, + 0.01632191799581051, + 0.019765499979257584, + 0.014839273877441883, + 0.011596526950597763, + 0.0082663893699646, + 0.004844177979975939, + 0.007609217427670956, + 0.0036067310720682144, + 0.004109140019863844, + 0.005700749345123768, + 0.004297202453017235, + 0.009982317686080933, + 0.00478427205234766, + 0.0045118508860468864, + 0.00926978886127472, + 0.004385408014059067, + 0.004554011393338442, + 0.005250678397715092, + 0.004015328362584114, + 0.00407675513997674, + 0.03428150713443756, + 0.01996598020195961, + 0.02248399518430233, + 0.012249129824340343, + 0.015262885019183159, + 0.012187061831355095, + 0.012691935524344444, + 0.015756912529468536, + 0.012617615051567554, + 0.014640092849731445, + 0.011036386713385582, + 0.01431870087981224, + 0.009816835634410381, + 0.015635374933481216, + 0.009181031957268715, + 0.008097630925476551, + 0.008731226436793804, + 0.007073855027556419, + 0.0071154809556901455, + 0.014585699886083603, + 0.0075874426402151585, + 0.008773411624133587, + 0.016184654086828232, + 0.011834201402962208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.011068139225244522 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial shot unfolds, the camera slowly panning right, overlooking a majestic mountain range blanketed in white snow. In the soft golden morning sunlight, two parallel railroad tracks, sharply outlined against the thick snow, wind through a sparse, frost-covered coniferous forest. Two classic black cast-iron steam locomotives, one heading deep into the frame, the other approaching, their bodies reflecting a faint metallic sheen in the sunlight. As they precisely intersect in the center of the frame, their massive drive wheels and complex linkages operate steadily and powerfully, while thick, billowing white steam billows from their towering chimneys, rapidly condensing into spectacular clouds in the frigid air. The slanting sunlight gave the steam cloud a warm golden edge, which contrasted sharply with the long blue shadows on the snow. The whole scene, in an epic magnificence and tranquility, showed the harmonious blend of the power of the industrial age and the beauty of nature, creating a magnificent and slightly nostalgic atmosphere.", + "seed": 20414, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_1/index_4968_seed_20414_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": -0.04630680605769157, + "vlm_logprob": -0.04630680605769157, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.04630680605769157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09513191133737564 + ], + "opposite_logprob_tokens": [ + -2.8451318740844727 + ], + "logprob_reward": -0.09513191133737564, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09513191133737564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02464645355939865, + 0.009817562066018581, + 0.007626090198755264, + 0.008688228204846382, + 0.008985379710793495, + 0.013289864175021648, + 0.046247612684965134, + 0.02748771198093891, + 0.015669921413064003, + 0.8375411033630371 + ], + "credit_map": [ + 0.0024512296076864004, + 0.00395343080163002, + 0.002265309216454625, + 0.00338415103033185, + 0.0038007486145943403, + 0.003437900682911277, + 0.0026736564468592405, + 0.002680031815543771, + 0.000807511736638844, + 0.0009869846981018782, + 0.000766092212870717, + 0.0008628203067928553, + 0.0013021033955737948, + 0.0015999876195564866, + 0.0014577774563804269, + 0.002034285804256797, + 0.0008587166666984558, + 0.0012597177410498261, + 0.0006633211742155254, + 0.0006665443652309477, + 0.0013137497007846832, + 0.001335995621047914, + 0.0009655804024077952, + 0.000562465051189065, + 0.0006704235565848649, + 0.0006506221834570169, + 0.0007144527626223862, + 0.0005655995919369161, + 0.0019263416761532426, + 0.0011784969829022884, + 0.0018831462366506457, + 0.001099146087653935, + 0.0009674836182966828, + 0.001150447758845985, + 0.0007912016590125859, + 0.0006412208313122392, + 0.0016776557313278317, + 0.0014712609117850661, + 0.0012822030112147331, + 0.0010039069456979632, + 0.0015229493146762252, + 0.0026579531840980053, + 0.0017092800699174404, + 0.0016308752819895744, + 0.00164342625066638, + 0.0017094239592552185, + 0.001383211463689804, + 0.0010327448835596442, + 0.002856138162314892, + 0.0016217674128711224, + 0.005312760826200247, + 0.0023813305888324976, + 0.019141843542456627, + 0.005975824780762196, + 0.007162643130868673, + 0.0017953045899048448, + 0.0027266668621450663, + 0.004054788500070572, + 0.0027169620152562857, + 0.002670127432793379, + 0.004387131426483393, + 0.005106505937874317, + 0.0036926104221493006, + 0.002132922876626253, + 0.0020159261766821146, + 0.0016596066998317838, + 0.0016497516771778464, + 0.0017250574892386794, + 0.001515539945103228, + 0.003429546020925045, + 0.001505821943283081, + 0.0021686735562980175, + 0.058636151254177094, + 0.03540617600083351, + 0.23060369491577148, + 0.049606937915086746, + 0.043207790702581406, + 0.271734356880188, + 0.01983289048075676, + 0.1285131424665451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.09513191133737564 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209130585193634 + ], + "opposite_logprob_tokens": [ + -2.81209135055542 + ], + "logprob_reward": -0.06209130585193634, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209130585193634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05561070889234543, + 0.03707050532102585, + 0.0270921029150486, + 0.013044912368059158, + 0.016806507483124733, + 0.019572287797927856, + 0.028253313153982162, + 0.026368379592895508, + 0.02598968893289566, + 0.7501916289329529 + ], + "credit_map": [ + 0.005174732767045498, + 0.008720789104700089, + 0.005225969012826681, + 0.0077243768610060215, + 0.00876670703291893, + 0.01074784342199564, + 0.00518623786047101, + 0.004064056556671858, + 0.0036734608002007008, + 0.0038936035707592964, + 0.0037947590462863445, + 0.0035646765027195215, + 0.0047712624073028564, + 0.00656479736790061, + 0.007267097011208534, + 0.0035408460535109043, + 0.0025670030154287815, + 0.004157275427132845, + 0.002399495104327798, + 0.0020982485730201006, + 0.005961772985756397, + 0.005036808084696531, + 0.0017168269259855151, + 0.0031546743120998144, + 0.0011924252612516284, + 0.001662840018980205, + 0.0011625471524894238, + 0.0012005803873762488, + 0.002225379226729274, + 0.0023923926055431366, + 0.001897932030260563, + 0.001310814986936748, + 0.0013956555631011724, + 0.0013568360591307282, + 0.001555320923216641, + 0.0012417077086865902, + 0.00408711563795805, + 0.0030746643897145987, + 0.0019372801762074232, + 0.0021579277236014605, + 0.0019061986822634935, + 0.005604199133813381, + 0.0023705228231847286, + 0.001744803274050355, + 0.0023960997350513935, + 0.0022936740424484015, + 0.001668915618211031, + 0.0015878754202276468, + 0.0018907864578068256, + 0.001320305629633367, + 0.0013699346454814076, + 0.0016186523716896772, + 0.009699525311589241, + 0.00291647226549685, + 0.007683261763304472, + 0.0017543761059641838, + 0.003006146987900138, + 0.003106190823018551, + 0.0034446122590452433, + 0.0026683311443775892, + 0.004325345158576965, + 0.005466730799525976, + 0.0021799965761601925, + 0.002171027474105358, + 0.003062610048800707, + 0.002129786880686879, + 0.0026524621061980724, + 0.0024868303444236517, + 0.0038598193787038326, + 0.004362168256193399, + 0.0027139794547110796, + 0.004722035024315119, + 0.05261945351958275, + 0.03251286968588829, + 0.20566219091415405, + 0.044859591871500015, + 0.03378065302968025, + 0.24627412855625153, + 0.0183839313685894, + 0.11609876155853271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.06209130585193634 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205455958843231 + ], + "opposite_logprob_tokens": [ + -2.8120546340942383 + ], + "logprob_reward": -0.06205455958843231, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205455958843231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10524285584688187, + 0.03920397907495499, + 0.051975369453430176, + 0.02989552542567253, + 0.025959493592381477, + 0.07882527261972427, + 0.058254778385162354, + 0.12085665762424469, + 0.01585082896053791, + 0.4739352762699127 + ], + "credit_map": [ + 0.010023417882621288, + 0.012083286419510841, + 0.013101701624691486, + 0.016467316076159477, + 0.016128405928611755, + 0.021740753203630447, + 0.007877192460000515, + 0.007820786908268929, + 0.004139425233006477, + 0.005171551369130611, + 0.004345777910202742, + 0.005236887373030186, + 0.005430842284113169, + 0.005671103484928608, + 0.004922797903418541, + 0.0042855930514633656, + 0.004671603441238403, + 0.007820927537977695, + 0.004724225495010614, + 0.004060768522322178, + 0.010816847905516624, + 0.009208256378769875, + 0.004427505191415548, + 0.006245231721550226, + 0.002594349207356572, + 0.004126383922994137, + 0.0029519388917833567, + 0.0022528597619384527, + 0.005530346650630236, + 0.004679497331380844, + 0.0047094328328967094, + 0.003050717292353511, + 0.002558494918048382, + 0.00415080226957798, + 0.0027453419752418995, + 0.00246661645360291, + 0.003933216445147991, + 0.0045936452224850655, + 0.0023395551834255457, + 0.0031718218233436346, + 0.00952820386737585, + 0.016963880509138107, + 0.009936951100826263, + 0.00950504932552576, + 0.009640800766646862, + 0.009749533608555794, + 0.0073722186498343945, + 0.006128635723143816, + 0.004118910059332848, + 0.0029314293060451746, + 0.0020448488648980856, + 0.0030855974182486534, + 0.019171210005879402, + 0.006093565374612808, + 0.016723819077014923, + 0.004085398279130459, + 0.012966138310730457, + 0.01170162670314312, + 0.014260941185057163, + 0.010671871714293957, + 0.014116374775767326, + 0.032874997705221176, + 0.013992130756378174, + 0.010272573679685593, + 0.0023942068219184875, + 0.0012214331654831767, + 0.0016103886300697923, + 0.0022467065136879683, + 0.0016740337014198303, + 0.0022558807395398617, + 0.001925532124005258, + 0.002522645751014352, + 0.03278784081339836, + 0.021590592339634895, + 0.12741142511367798, + 0.027764203026890755, + 0.018009264022111893, + 0.1599966287612915, + 0.011900429613888264, + 0.07447486370801926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.06205455958843231 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643949404358864 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008643949404358864, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643949404358864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04996141791343689, + 0.03661659359931946, + 0.02691253274679184, + 0.017285004258155823, + 0.01659386232495308, + 0.0482107475399971, + 0.0191329512745142, + 0.17696595191955566, + 0.017668047919869423, + 0.5906528830528259 + ], + "credit_map": [ + 0.005301633384078741, + 0.0065225400030612946, + 0.0063731432892382145, + 0.00628467183560133, + 0.0066607254557311535, + 0.010484735481441021, + 0.004281443078070879, + 0.004052526783198118, + 0.002818590495735407, + 0.0032828906551003456, + 0.0030133018735796213, + 0.002379676792770624, + 0.003801789367571473, + 0.007892363704741001, + 0.003789986949414015, + 0.009637993760406971, + 0.002089644782245159, + 0.0030120774172246456, + 0.002481515286490321, + 0.0015490340301766992, + 0.00794627983123064, + 0.004854817874729633, + 0.002163904719054699, + 0.0028152575250715017, + 0.001295452588237822, + 0.0026189940981566906, + 0.0015099982265383005, + 0.0016530717257410288, + 0.003027860540896654, + 0.0023149275220930576, + 0.0029971327167004347, + 0.0018675666069611907, + 0.0016011649277061224, + 0.002320184838026762, + 0.0019782420713454485, + 0.001396216219291091, + 0.0022349096834659576, + 0.0025446959771215916, + 0.002669940935447812, + 0.0018485074397176504, + 0.006250597536563873, + 0.01109133567661047, + 0.0058665028773248196, + 0.004373530391603708, + 0.005725893657654524, + 0.006002525798976421, + 0.004102623555809259, + 0.004797742702066898, + 0.0018149387324228883, + 0.0018241474172100425, + 0.0034751202911138535, + 0.001578367780894041, + 0.00500111561268568, + 0.002342603402212262, + 0.001847385661676526, + 0.0012492716778069735, + 0.01837341859936714, + 0.017168588936328888, + 0.03189772740006447, + 0.016352498903870583, + 0.028955752030014992, + 0.03669743984937668, + 0.01239418052136898, + 0.015126352198421955, + 0.001720758737064898, + 0.001305855461396277, + 0.0018344426061958075, + 0.001531498390249908, + 0.0021557356230914593, + 0.005234776064753532, + 0.001506409258581698, + 0.00237857224419713, + 0.04140464588999748, + 0.024774309247732162, + 0.16328772902488708, + 0.03488269820809364, + 0.03181587904691696, + 0.19026774168014526, + 0.014136221259832382, + 0.09008365124464035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.008643949404358864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an aerial view that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two parallel railroad tracks visible winding through a frost-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two classic black cast-iron steam locomotives present, one moving away and one approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the locomotives intersect in the center of the frame with visible billowing white steam and metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a mountain range blanketed in white snow under soft golden morning sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036123041063547134 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.0036123041063547134, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036123041063547134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06814289093017578, + 0.02318168617784977, + 0.021600689738988876, + 0.011696520261466503, + 0.016418814659118652, + 0.025521136820316315, + 0.07362332195043564, + 0.04670063778758049, + 0.041405871510505676, + 0.6717084646224976 + ], + "credit_map": [ + 0.006584683433175087, + 0.007828389294445515, + 0.007562045473605394, + 0.006604194175451994, + 0.010816610418260098, + 0.010543407872319221, + 0.00673301424831152, + 0.011470544151961803, + 0.0025958886835724115, + 0.0028152447193861008, + 0.002131313318386674, + 0.0017103791469708085, + 0.0027777848299592733, + 0.004217670764774084, + 0.0021175288129597902, + 0.0048158783465623856, + 0.002412335015833378, + 0.0019627108704298735, + 0.001921561430208385, + 0.001154670026153326, + 0.004815947264432907, + 0.00542489206418395, + 0.0015213664155453444, + 0.0023872083984315395, + 0.0012363382847979665, + 0.0009852403309196234, + 0.0010746404295787215, + 0.0008801480871625245, + 0.001609827857464552, + 0.002289109630510211, + 0.001804252970032394, + 0.0018169638933613896, + 0.002278405474498868, + 0.0018664438975974917, + 0.001365188043564558, + 0.0010195381473749876, + 0.002114230068400502, + 0.004243454895913601, + 0.0019215315114706755, + 0.0016100237844511867, + 0.002555797342211008, + 0.005405517760664225, + 0.003334721317514777, + 0.003333723871037364, + 0.002745747333392501, + 0.003472068812698126, + 0.0026276533026248217, + 0.002045906614512205, + 0.004481709562242031, + 0.0026801330968737602, + 0.007342364173382521, + 0.003840353572741151, + 0.030459031462669373, + 0.0090617211535573, + 0.013359801843762398, + 0.0023982091806828976, + 0.005114655941724777, + 0.004097159020602703, + 0.0075585199519991875, + 0.0040132408030331135, + 0.006967426277697086, + 0.008951284922659397, + 0.0044824848882853985, + 0.0055158669129014015, + 0.003883559722453356, + 0.0025817861314862967, + 0.004007564391940832, + 0.0033672689460217953, + 0.0055459411814808846, + 0.012507215142250061, + 0.0031332094222307205, + 0.006379329599440098, + 0.04782227799296379, + 0.0292019322514534, + 0.18473583459854126, + 0.03968626633286476, + 0.031632546335458755, + 0.219711571931839, + 0.016043638810515404, + 0.10287436097860336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0036123041063547134 + } + ], + "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/0000128/group_10/index_tensor([6559, 6559])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_10/index_tensor([6559, 6559])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f95860f89187e225f5d8d0287592e0c38af1510a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_10/index_tensor([6559, 6559])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning display of high-contrast photo-realism, a slow, sweeping, low-angle tracking shot glides over a majestic, solitary oak tree, standing defiantly on a grassy knoll beneath a vast, clear blue sky. Its ancient, heavy, gnarled branches, covered in deeply fissured and richly textured bark, snake skyward, their gnarled crown densely laden with ripe acorns. A powerful, unseen gust of wind suddenly strikes the tree, causing its entire massive trunk to bear up and its heavy branches to sway and shake violently. The movement is violent and energetic, with leaves swaying in a chaotic dance. The scene is defined by harsh, brilliant sunlight, creating a stark interplay of light and shadow; deep, inky shadows shift swiftly across the tree's richly textured bark and the sun-kissed leaves, which shimmer between vibrant emerald and near-black, generating a powerful and dramatic primal atmosphere that captures raw, untamed nature with visceral clarity.", + "seed": 5229, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_10/index_6559_seed_5229_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": -0.2128661091439426, + "vlm_logprob": -0.2128661091439426, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.2128661091439426, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0731685608625412 + ], + "opposite_logprob_tokens": [ + -2.6981685161590576 + ], + "logprob_reward": -0.0731685608625412, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0731685608625412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1300560086965561, + 0.17291507124900818, + 0.11243928223848343, + 0.18801464140415192, + 0.0857536643743515, + 0.059571798890829086, + 0.019004350528120995, + 0.02805161662399769, + 0.1299615353345871, + 0.07423200458288193 + ], + "credit_map": [ + 0.017170554026961327, + 0.012127064168453217, + 0.019137026742100716, + 0.01653958298265934, + 0.01589500531554222, + 0.016762826591730118, + 0.01723945140838623, + 0.015184500254690647, + 0.040492232888936996, + 0.014984751120209694, + 0.027850138023495674, + 0.019704541191458702, + 0.012602142058312893, + 0.015579712577164173, + 0.017284663394093513, + 0.02441689744591713, + 0.014256647787988186, + 0.01475945208221674, + 0.011635633185505867, + 0.023062104359269142, + 0.01668807864189148, + 0.013742633163928986, + 0.010841138660907745, + 0.007453594356775284, + 0.06866190582513809, + 0.02429361455142498, + 0.014392160810530186, + 0.014013894833624363, + 0.01200911682099104, + 0.009324753656983376, + 0.02008301205933094, + 0.02523619309067726, + 0.017735498026013374, + 0.009204482659697533, + 0.011205238290131092, + 0.008806063793599606, + 0.007396232336759567, + 0.006959862541407347, + 0.012315471656620502, + 0.012130813673138618, + 0.007996215485036373, + 0.007188747171312571, + 0.008650428615510464, + 0.007373171858489513, + 0.007704996038228273, + 0.008723526261746883, + 0.0068956282921135426, + 0.005039084702730179, + 0.0028478449676185846, + 0.0020736949518322945, + 0.002116735093295574, + 0.0020899558439850807, + 0.002040244173258543, + 0.003050000173971057, + 0.0025259978137910366, + 0.0022598751820623875, + 0.004047501366585493, + 0.0035682395100593567, + 0.0039167036302387714, + 0.004158699419349432, + 0.0029284076299518347, + 0.0030023264698684216, + 0.003927765414118767, + 0.002501972019672394, + 0.02947383187711239, + 0.016364172101020813, + 0.011728874407708645, + 0.0205217432230711, + 0.00921826995909214, + 0.012146762572228909, + 0.016476545482873917, + 0.01403134223073721, + 0.012132925912737846, + 0.007240115664899349, + 0.00726158544421196, + 0.0069076563231647015, + 0.006931424140930176, + 0.013280494138598442, + 0.012657969258725643, + 0.007819831371307373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0731685608625412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811749815940857 + ], + "opposite_logprob_tokens": [ + -3.2881174087524414 + ], + "logprob_reward": -0.03811749815940857, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811749815940857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03460508957505226, + 0.05671239644289017, + 0.03692290186882019, + 0.27826938033103943, + 0.20326006412506104, + 0.040931232273578644, + 0.07917042821645737, + 0.059909217059612274, + 0.11335860192775726, + 0.09686071425676346 + ], + "credit_map": [ + 0.005823876243084669, + 0.0030709421262145042, + 0.004887604620307684, + 0.0038934098556637764, + 0.004031292162835598, + 0.005090843420475721, + 0.004173063673079014, + 0.00363405654206872, + 0.010513022541999817, + 0.0063171107321977615, + 0.005927288439124823, + 0.007006743922829628, + 0.009227104485034943, + 0.007077060639858246, + 0.006074231583625078, + 0.004569833632558584, + 0.0066793025471270084, + 0.0054366448894143105, + 0.005411062855273485, + 0.006251900922507048, + 0.0029364563524723053, + 0.0027796137146651745, + 0.004743163473904133, + 0.0026847573462873697, + 0.10535076260566711, + 0.03750753030180931, + 0.01976245641708374, + 0.015493806451559067, + 0.017999647185206413, + 0.01188772264868021, + 0.031882818788290024, + 0.038384631276130676, + 0.05114448443055153, + 0.023701883852481842, + 0.019362712278962135, + 0.017224520444869995, + 0.01344925444573164, + 0.0187041237950325, + 0.02394787035882473, + 0.03572520241141319, + 0.008096815086901188, + 0.005148417316377163, + 0.005553094670176506, + 0.0041767568327486515, + 0.0031794931273907423, + 0.0038090848829597235, + 0.00679214159026742, + 0.004175430163741112, + 0.014508040621876717, + 0.01131310872733593, + 0.007213061675429344, + 0.012111260555684566, + 0.006584626622498035, + 0.008230229839682579, + 0.012806298211216927, + 0.0064038001000881195, + 0.00828627496957779, + 0.009988892823457718, + 0.009442797861993313, + 0.009039108641445637, + 0.005217674653977156, + 0.00549389747902751, + 0.006363373249769211, + 0.006077200174331665, + 0.02489352785050869, + 0.014732328243553638, + 0.009824411012232304, + 0.018051758408546448, + 0.008029053919017315, + 0.010346162132918835, + 0.014344936236739159, + 0.013136429712176323, + 0.02274545282125473, + 0.012620652094483376, + 0.008197079412639141, + 0.01097849290817976, + 0.008340579457581043, + 0.009785490110516548, + 0.013976296409964561, + 0.010216670110821724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.03811749815940857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25206857919692993 + ], + "opposite_logprob_tokens": [ + -1.5020685195922852 + ], + "logprob_reward": -0.25206857919692993, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25206857919692993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15619657933712006, + 0.058314286172389984, + 0.07832589000463486, + 0.42216822504997253, + 0.058414287865161896, + 0.026941003277897835, + 0.034395650029182434, + 0.04188999533653259, + 0.04001125320792198, + 0.08334282785654068 + ], + "credit_map": [ + 0.018787790089845657, + 0.015684451907873154, + 0.024971405044198036, + 0.0210835300385952, + 0.016955476254224777, + 0.024131761863827705, + 0.018458520993590355, + 0.01612364500761032, + 0.008825466968119144, + 0.006405648775398731, + 0.010676359757781029, + 0.009260090999305248, + 0.0042519462294876575, + 0.006903410889208317, + 0.006964368745684624, + 0.0050269984640181065, + 0.019254548475146294, + 0.009228026494383812, + 0.010773137211799622, + 0.0072020962834358215, + 0.006193941924721003, + 0.006474226247519255, + 0.00856064260005951, + 0.01063927635550499, + 0.15380187332630157, + 0.055647313594818115, + 0.03502486273646355, + 0.024728789925575256, + 0.028420349583029747, + 0.022174566984176636, + 0.04843807965517044, + 0.05393240973353386, + 0.00671852333471179, + 0.007363625802099705, + 0.007734789978712797, + 0.006017239298671484, + 0.010296562686562538, + 0.007064111530780792, + 0.0070826392620801926, + 0.00613680062815547, + 0.0036231998819857836, + 0.003775827120989561, + 0.004706544801592827, + 0.0034945278894156218, + 0.0026239589788019657, + 0.0036987136118113995, + 0.002807093784213066, + 0.0022111372090876102, + 0.008005485869944096, + 0.003246455220505595, + 0.005546019412577152, + 0.004384726285934448, + 0.002034679753705859, + 0.003252958180382848, + 0.004933423828333616, + 0.0029919021762907505, + 0.008124771527945995, + 0.004894082434475422, + 0.005541869439184666, + 0.006238123401999474, + 0.003509909613057971, + 0.00362706882879138, + 0.005654327571392059, + 0.004299841821193695, + 0.010997297242283821, + 0.0037744371220469475, + 0.005228164605796337, + 0.00541317043825984, + 0.003022294957190752, + 0.003426718758419156, + 0.004499661270529032, + 0.003649509744718671, + 0.01125276181846857, + 0.009032337926328182, + 0.008069917559623718, + 0.011073500849306583, + 0.011016816832125187, + 0.012962300330400467, + 0.011924882419407368, + 0.00801031943410635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.25206857919692993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6933444738388062 + ], + "opposite_logprob_tokens": [ + -0.6933444738388062 + ], + "logprob_reward": -0.6933444738388062, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933444738388062, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10067842155694962, + 0.08978786319494247, + 0.24223816394805908, + 0.279985636472702, + 0.05820431932806969, + 0.05780269578099251, + 0.0329163633286953, + 0.0350993312895298, + 0.04612113535404205, + 0.05716618150472641 + ], + "credit_map": [ + 0.014353501610457897, + 0.009311901405453682, + 0.01639747992157936, + 0.015071098692715168, + 0.010161629877984524, + 0.011509707197546959, + 0.014899016357958317, + 0.008974076248705387, + 0.021359790116548538, + 0.007864625193178654, + 0.01575869508087635, + 0.010329670272767544, + 0.006485871504992247, + 0.008692899718880653, + 0.007056232541799545, + 0.012240066193044186, + 0.058056820183992386, + 0.023458097130060196, + 0.03865902125835419, + 0.031859513372182846, + 0.0192088820040226, + 0.016394317150115967, + 0.02095632627606392, + 0.033645156770944595, + 0.09926758706569672, + 0.037418264895677567, + 0.022816304117441177, + 0.02115044742822647, + 0.018241630867123604, + 0.012509423308074474, + 0.03082135133445263, + 0.03776059299707413, + 0.009873779490590096, + 0.01035238429903984, + 0.006539716385304928, + 0.005683045834302902, + 0.008659539744257927, + 0.00595470005646348, + 0.005890632513910532, + 0.00525051262229681, + 0.01482222881168127, + 0.006454764399677515, + 0.007694854401051998, + 0.005472815129905939, + 0.00426470348611474, + 0.004331050906330347, + 0.008032447658479214, + 0.006729821674525738, + 0.006937049329280853, + 0.004515734501183033, + 0.0026812327560037374, + 0.004536840133368969, + 0.0032582716085016727, + 0.00305348914116621, + 0.005375675391405821, + 0.002558063715696335, + 0.005043606273829937, + 0.004045515321195126, + 0.004292935132980347, + 0.005062551703304052, + 0.0038088869769126177, + 0.004944781772792339, + 0.004188845865428448, + 0.0037122059147804976, + 0.010958311147987843, + 0.004356998018920422, + 0.00482267327606678, + 0.00835797656327486, + 0.0033046535681933165, + 0.003893601708114147, + 0.0055861216969788074, + 0.004840793553739786, + 0.011783418245613575, + 0.005759008228778839, + 0.0061925551854074, + 0.006633399520069361, + 0.004371357150375843, + 0.006102932151407003, + 0.007181256078183651, + 0.009142247028648853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.6933444738388062 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1289624124765396, + 0.05998821184039116, + 0.1758573055267334, + 0.22097209095954895, + 0.07240496575832367, + 0.1274581402540207, + 0.012272508814930916, + 0.02276155911386013, + 0.12755121290683746, + 0.051771607249975204 + ], + "credit_map": [ + 0.020373838022351265, + 0.012269407510757446, + 0.021725714206695557, + 0.01639137789607048, + 0.011607110500335693, + 0.018673371523618698, + 0.017239093780517578, + 0.01068251021206379, + 0.01664510741829872, + 0.0063681998290121555, + 0.0064149596728384495, + 0.005367330275475979, + 0.006790322717279196, + 0.005045868456363678, + 0.0054354541935026646, + 0.007920967414975166, + 0.030549203976988792, + 0.017809763550758362, + 0.021868854761123657, + 0.03615111485123634, + 0.020252730697393417, + 0.015605279244482517, + 0.018940089270472527, + 0.014680266380310059, + 0.07871926575899124, + 0.02889641560614109, + 0.017729377374053, + 0.014683416113257408, + 0.014482111670076847, + 0.009469160810112953, + 0.02568872831761837, + 0.03130361810326576, + 0.0159451924264431, + 0.009529643692076206, + 0.008106271736323833, + 0.007747998461127281, + 0.006336995866149664, + 0.006543003022670746, + 0.008724396117031574, + 0.009471462108194828, + 0.03404955193400383, + 0.013327015563845634, + 0.011107378639280796, + 0.011025951243937016, + 0.011576974764466286, + 0.009911680594086647, + 0.021213719621300697, + 0.015245865099132061, + 0.0020913337357342243, + 0.0014991351636126637, + 0.0014795397873967886, + 0.0019391529494896531, + 0.0012186251115053892, + 0.0012790851760655642, + 0.0015728286234661937, + 0.0011928087333217263, + 0.005612513050436974, + 0.0026664191391319036, + 0.0023850062862038612, + 0.002932269126176834, + 0.0017271592514589429, + 0.0017960056429728866, + 0.003197904210537672, + 0.0024442814756184816, + 0.02735326997935772, + 0.015873553231358528, + 0.012927078641951084, + 0.0184857826679945, + 0.00968057382851839, + 0.012807666324079037, + 0.01656770333647728, + 0.013855582103133202, + 0.008223241195082664, + 0.006829366087913513, + 0.00494556687772274, + 0.008191329427063465, + 0.0043852380476891994, + 0.006397665478289127, + 0.006824574898928404, + 0.005974625702947378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.007631433662027121 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a stunning display of high-contrast photo-realism, a slow, sweeping, low-angle tracking shot glides over a majestic, solitary oak tree, standing defiantly on a grassy knoll beneath a vast, clear blue sky. Its ancient, heavy, gnarled branches, covered in deeply fissured and richly textured bark, snake skyward, their gnarled crown densely laden with ripe acorns. A powerful, unseen gust of wind suddenly strikes the tree, causing its entire massive trunk to bear up and its heavy branches to sway and shake violently. The movement is violent and energetic, with leaves swaying in a chaotic dance. The scene is defined by harsh, brilliant sunlight, creating a stark interplay of light and shadow; deep, inky shadows shift swiftly across the tree's richly textured bark and the sun-kissed leaves, which shimmer between vibrant emerald and near-black, generating a powerful and dramatic primal atmosphere that captures raw, untamed nature with visceral clarity.", + "seed": 5230, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_10/index_6559_seed_5230_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": -0.2909584637731314, + "vlm_logprob": -0.2909584637731314, + "vlm_yes_no_margin": 2.200000023841858, + "vlm_token_credit": -0.2909584637731314, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8996182680130005 + ], + "opposite_logprob_tokens": [ + -0.5246182680130005 + ], + "logprob_reward": -0.8996182680130005, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8996182680130005, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16760073602199554, + 0.1443248689174652, + 0.26594552397727966, + 0.06721548736095428, + 0.07895135879516602, + 0.07876494526863098, + 0.03220074623823166, + 0.09170743077993393, + 0.024488046765327454, + 0.04880087822675705 + ], + "credit_map": [ + 0.02485804073512554, + 0.01952841691672802, + 0.015477008186280727, + 0.01598467119038105, + 0.03913990408182144, + 0.014027480967342854, + 0.01995438151061535, + 0.018630821257829666, + 0.016881264746189117, + 0.022258495911955833, + 0.010592126287519932, + 0.012215415015816689, + 0.03124701790511608, + 0.017330141738057137, + 0.018432559445500374, + 0.015367846935987473, + 0.03548699617385864, + 0.049819353967905045, + 0.023184066638350487, + 0.01977919600903988, + 0.03860083594918251, + 0.03558297082781792, + 0.031209830194711685, + 0.032282255589962006, + 0.008282250724732876, + 0.012755605392158031, + 0.00721635390073061, + 0.005417626351118088, + 0.015435347333550453, + 0.0055382149294018745, + 0.006771237123757601, + 0.005798851139843464, + 0.019489409402012825, + 0.011062228120863438, + 0.007925949059426785, + 0.00830018799751997, + 0.010563078336417675, + 0.007943296805024147, + 0.006904768291860819, + 0.006762434262782335, + 0.008310076780617237, + 0.012788902968168259, + 0.006439553573727608, + 0.007535628043115139, + 0.014168537221848965, + 0.01407836563885212, + 0.00943575520068407, + 0.006008124444633722, + 0.0057977731339633465, + 0.005650112871080637, + 0.0031096262391656637, + 0.0031390509102493525, + 0.004757911432534456, + 0.003888259408995509, + 0.0031719959806650877, + 0.0026860127691179514, + 0.01648341678082943, + 0.009219820611178875, + 0.00725364638492465, + 0.005857645068317652, + 0.014666191302239895, + 0.013867822475731373, + 0.009068039245903492, + 0.015290837734937668, + 0.003207914298400283, + 0.0033459579572081566, + 0.002686053514480591, + 0.0018203894142061472, + 0.003943872172385454, + 0.004045411478728056, + 0.0022098731715232134, + 0.003228571964427829, + 0.004889972507953644, + 0.005483934190124273, + 0.006392157170921564, + 0.008804390206933022, + 0.006489594466984272, + 0.008422637358307838, + 0.004417972639203072, + 0.00390021875500679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.8996182680130005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011095019057393074 + ], + "opposite_logprob_tokens": [ + -4.51109504699707 + ], + "logprob_reward": -0.011095019057393074, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011095019057393074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10246869921684265, + 0.07731809467077255, + 0.486718088388443, + 0.06769302487373352, + 0.10797566920518875, + 0.07175277173519135, + 0.022741591557860374, + 0.017421159893274307, + 0.02194729633629322, + 0.02396373637020588 + ], + "credit_map": [ + 0.012385878711938858, + 0.011498764157295227, + 0.012320653535425663, + 0.010157535783946514, + 0.020380187779664993, + 0.009057514369487762, + 0.012507958337664604, + 0.014160188846290112, + 0.008875343948602676, + 0.011580866761505604, + 0.006510696839541197, + 0.006860856432467699, + 0.010879001580178738, + 0.010326672345399857, + 0.011768571101129055, + 0.010516076348721981, + 0.06673381477594376, + 0.09071771055459976, + 0.04305596649646759, + 0.03682103753089905, + 0.0718732476234436, + 0.0623062327504158, + 0.05770391225814819, + 0.05750608816742897, + 0.00780489994212985, + 0.01122156623750925, + 0.010338392108678818, + 0.005545290187001228, + 0.013460853137075901, + 0.006320551969110966, + 0.006839529145509005, + 0.00616192864254117, + 0.015255521982908249, + 0.021185193210840225, + 0.01234695315361023, + 0.013124441727995872, + 0.015372869558632374, + 0.008912905119359493, + 0.01158944796770811, + 0.010188326239585876, + 0.007977665401995182, + 0.012306592427194118, + 0.0065725864842534065, + 0.005552690476179123, + 0.010224681347608566, + 0.01791846938431263, + 0.006473080720752478, + 0.00472699198871851, + 0.004230644553899765, + 0.0038861825596541166, + 0.0025199619121849537, + 0.0022698906250298023, + 0.0032454123720526695, + 0.0029044717084616423, + 0.0020933395717293024, + 0.0015916862757876515, + 0.0018658863846212626, + 0.002166164806112647, + 0.0019740969873964787, + 0.0016443963395431638, + 0.002870061667636037, + 0.003835513722151518, + 0.0012553933775052428, + 0.001809645676985383, + 0.0028749138582497835, + 0.002744675148278475, + 0.0029280264861881733, + 0.001825683400966227, + 0.003397732973098755, + 0.0034693898633122444, + 0.0020956455264240503, + 0.0026112261693924665, + 0.0024132446851581335, + 0.002887079492211342, + 0.003333121072500944, + 0.003922198433429003, + 0.0032731848768889904, + 0.004066580906510353, + 0.002116154180839658, + 0.0019521708600223064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.011095019057393074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428011804819107 + ], + "opposite_logprob_tokens": [ + -2.017801284790039 + ], + "logprob_reward": -0.1428011804819107, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428011804819107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06604461371898651, + 0.09197816252708435, + 0.3055839240550995, + 0.03615405037999153, + 0.04329618811607361, + 0.07267484813928604, + 0.04642033949494362, + 0.2553251385688782, + 0.05098281055688858, + 0.03153989464044571 + ], + "credit_map": [ + 0.011586272157728672, + 0.007668757811188698, + 0.008744762279093266, + 0.006012157071381807, + 0.008698887191712856, + 0.007925097830593586, + 0.006140060257166624, + 0.009268623776733875, + 0.01088014803826809, + 0.014058275148272514, + 0.009748219512403011, + 0.005478882230818272, + 0.021189691498875618, + 0.011333763599395752, + 0.011473406106233597, + 0.007815785706043243, + 0.04098493978381157, + 0.05634544789791107, + 0.02702421136200428, + 0.02371719479560852, + 0.04487184062600136, + 0.04120505601167679, + 0.03430983051657677, + 0.037125423550605774, + 0.004614589270204306, + 0.00610725162550807, + 0.004625583067536354, + 0.0035338029265403748, + 0.005146749317646027, + 0.004217207431793213, + 0.0033385648857802153, + 0.004570304416120052, + 0.008153006434440613, + 0.007618734147399664, + 0.004271417856216431, + 0.0042748530395329, + 0.005585063248872757, + 0.0044210683554410934, + 0.005055238027125597, + 0.003916812129318714, + 0.009951040148735046, + 0.009037901647388935, + 0.008644520305097103, + 0.00813765823841095, + 0.008783791214227676, + 0.007407098542898893, + 0.009685743600130081, + 0.011027097702026367, + 0.0073280674405395985, + 0.005663676653057337, + 0.004121412057429552, + 0.00488103786483407, + 0.008426823653280735, + 0.007013216149061918, + 0.004607882350683212, + 0.0043782261200249195, + 0.04314827919006348, + 0.02913687378168106, + 0.023099210113286972, + 0.021447552368044853, + 0.04851244390010834, + 0.03824581950902939, + 0.023678699508309364, + 0.02805631048977375, + 0.005893413908779621, + 0.005378618836402893, + 0.008732481859624386, + 0.006998618133366108, + 0.005418315529823303, + 0.007026350125670433, + 0.0047124591656029224, + 0.006822557654231787, + 0.003956940025091171, + 0.0025742969010025263, + 0.003533534472808242, + 0.006887372583150864, + 0.0029150028713047504, + 0.005141942296177149, + 0.003112283069640398, + 0.003418526379391551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.1428011804819107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870786726474762 + ], + "opposite_logprob_tokens": [ + -1.1370786428451538 + ], + "logprob_reward": -0.3870786726474762, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870786726474762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13481348752975464, + 0.14024558663368225, + 0.10492730885744095, + 0.08281276375055313, + 0.08787044137716293, + 0.1661890596151352, + 0.04301397129893303, + 0.10070785135030746, + 0.07527022063732147, + 0.06414930522441864 + ], + "credit_map": [ + 0.015377095900475979, + 0.01578338071703911, + 0.018563639372587204, + 0.014572597108781338, + 0.022825095802545547, + 0.012475910596549511, + 0.016515234485268593, + 0.01870052143931389, + 0.016506800428032875, + 0.020706109702587128, + 0.010300688445568085, + 0.011939004994928837, + 0.030889835208654404, + 0.01606602780520916, + 0.01937009207904339, + 0.014467029832303524, + 0.013520671054720879, + 0.022551782429218292, + 0.008608382195234299, + 0.008840170688927174, + 0.021440736949443817, + 0.010595346800982952, + 0.008172988891601562, + 0.011197224259376526, + 0.010100017301738262, + 0.01504810806363821, + 0.009381731040775776, + 0.006203183438628912, + 0.020357007160782814, + 0.00653228722512722, + 0.008500928990542889, + 0.0066895028576254845, + 0.017190083861351013, + 0.013257970102131367, + 0.009965542703866959, + 0.008206932805478573, + 0.01155531033873558, + 0.009898441843688488, + 0.008305689319968224, + 0.009490467607975006, + 0.020116310566663742, + 0.021821802482008934, + 0.01904902420938015, + 0.012779413722455502, + 0.017525766044855118, + 0.03719247877597809, + 0.014152426272630692, + 0.02355184778571129, + 0.006970164366066456, + 0.007364014629274607, + 0.0039146216586232185, + 0.005851021967828274, + 0.0062052225694060326, + 0.005039785988628864, + 0.00370878167450428, + 0.003960362169891596, + 0.019351404160261154, + 0.012778203003108501, + 0.006698936689645052, + 0.007278387434780598, + 0.016992105171084404, + 0.01895168423652649, + 0.010130425915122032, + 0.008526706136763096, + 0.00995886605232954, + 0.007643674034625292, + 0.013254641555249691, + 0.009687133133411407, + 0.008285116404294968, + 0.010864251293241978, + 0.007236468140035868, + 0.008340063504874706, + 0.006912878714501858, + 0.006989570800215006, + 0.007608726620674133, + 0.01208407711237669, + 0.00628773495554924, + 0.011629548855125904, + 0.00620869779959321, + 0.0064280773513019085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.3870786726474762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199178665876389 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199178665876389, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199178665876389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13619056344032288, + 0.17790727317333221, + 0.15929484367370605, + 0.03433812037110329, + 0.06743484735488892, + 0.20105253159999847, + 0.053012095391750336, + 0.06764073669910431, + 0.04109448194503784, + 0.0620344914495945 + ], + "credit_map": [ + 0.026971373707056046, + 0.015649860724806786, + 0.02047969028353691, + 0.008722344413399696, + 0.0298574510961771, + 0.011395690031349659, + 0.010858964174985886, + 0.012255200184881687, + 0.027788132429122925, + 0.027731485664844513, + 0.0173586867749691, + 0.00916161946952343, + 0.05156745761632919, + 0.016718821600079536, + 0.0155266048386693, + 0.01205445546656847, + 0.021632986143231392, + 0.028558235615491867, + 0.015953518450260162, + 0.01166642364114523, + 0.024315157905220985, + 0.019686883315443993, + 0.01931367814540863, + 0.018167944625020027, + 0.0038373603019863367, + 0.004903126507997513, + 0.00438383873552084, + 0.0034567678812891245, + 0.005772228352725506, + 0.00323724327608943, + 0.004010009113699198, + 0.0047375475987792015, + 0.01592337153851986, + 0.01031782291829586, + 0.005078976042568684, + 0.004815469495952129, + 0.00878228060901165, + 0.008233489468693733, + 0.007808384019881487, + 0.006475044414401054, + 0.02347400039434433, + 0.03221217170357704, + 0.02169826254248619, + 0.014284343458712101, + 0.018290136009454727, + 0.06087198480963707, + 0.013190281577408314, + 0.017031343653798103, + 0.010474571026861668, + 0.007499367464333773, + 0.0073774224147200584, + 0.003410136792808771, + 0.0066178240813314915, + 0.007089161779731512, + 0.004971775226294994, + 0.005571838468313217, + 0.014906765893101692, + 0.006813321262598038, + 0.0061070481315255165, + 0.004146846942603588, + 0.012314291670918465, + 0.00903289020061493, + 0.00607963977381587, + 0.008239938877522945, + 0.005876587238162756, + 0.004296505823731422, + 0.006493539549410343, + 0.003767274785786867, + 0.005338040180504322, + 0.00622300710529089, + 0.004088905639946461, + 0.005010619293898344, + 0.007238025777041912, + 0.005435805767774582, + 0.010309898294508457, + 0.010120738297700882, + 0.005455134902149439, + 0.012322488240897655, + 0.005875918548554182, + 0.00527648301795125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.014199178665876389 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning display of high-contrast photo-realism, a slow, sweeping, low-angle tracking shot glides over a majestic, solitary oak tree, standing defiantly on a grassy knoll beneath a vast, clear blue sky. Its ancient, heavy, gnarled branches, covered in deeply fissured and richly textured bark, snake skyward, their gnarled crown densely laden with ripe acorns. A powerful, unseen gust of wind suddenly strikes the tree, causing its entire massive trunk to bear up and its heavy branches to sway and shake violently. The movement is violent and energetic, with leaves swaying in a chaotic dance. The scene is defined by harsh, brilliant sunlight, creating a stark interplay of light and shadow; deep, inky shadows shift swiftly across the tree's richly textured bark and the sun-kissed leaves, which shimmer between vibrant emerald and near-black, generating a powerful and dramatic primal atmosphere that captures raw, untamed nature with visceral clarity.", + "seed": 5231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_10/index_6559_seed_5231_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": -0.20705166943371295, + "vlm_logprob": -0.20705166943371295, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.20705166943371295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20536823570728302 + ], + "opposite_logprob_tokens": [ + -1.7053682804107666 + ], + "logprob_reward": -0.20536823570728302, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20536823570728302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1844835877418518, + 0.462767094373703, + 0.033999573439359665, + 0.10431864112615585, + 0.04530828446149826, + 0.04854055121541023, + 0.0471920482814312, + 0.017095670104026794, + 0.009310781955718994, + 0.04698386788368225 + ], + "credit_map": [ + 0.018409905955195427, + 0.03515877202153206, + 0.010432264767587185, + 0.01124060433357954, + 0.05202750489115715, + 0.023848654702305794, + 0.020743926987051964, + 0.012621922418475151, + 0.04095578193664551, + 0.15691818296909332, + 0.02957790717482567, + 0.021755000576376915, + 0.08805662393569946, + 0.01819135993719101, + 0.03648245707154274, + 0.07082969695329666, + 0.004779561422765255, + 0.006321159191429615, + 0.004215294495224953, + 0.0030016128439456224, + 0.005258433986455202, + 0.00445492472499609, + 0.0027945791371166706, + 0.003173999721184373, + 0.008747542276978493, + 0.02141118422150612, + 0.009351599030196667, + 0.007744763046503067, + 0.012980778701603413, + 0.021894190460443497, + 0.01012144424021244, + 0.012067116796970367, + 0.011663091368973255, + 0.0036213500425219536, + 0.003998341038823128, + 0.003562947968021035, + 0.004270288627594709, + 0.0035138507373631, + 0.011171981692314148, + 0.0035064255353063345, + 0.007562318351119757, + 0.002724098740145564, + 0.004422828555107117, + 0.008451621979475021, + 0.009480757638812065, + 0.0075353821739554405, + 0.004379636142402887, + 0.003983896225690842, + 0.011466116644442081, + 0.002823940245434642, + 0.0030369816813617945, + 0.004399868659675121, + 0.015340916812419891, + 0.0026342228520661592, + 0.004536185413599014, + 0.0029538043309003115, + 0.0019416621653363109, + 0.001491307863034308, + 0.0016053528524935246, + 0.0018324392149224877, + 0.001905289595015347, + 0.002704166341573, + 0.003396803978830576, + 0.0022186439018696547, + 0.0012947425711899996, + 0.0013179857051, + 0.001282937591895461, + 0.0011312247952446342, + 0.00119401840493083, + 0.0011496178340166807, + 0.0009800887200981379, + 0.0009601641213521361, + 0.0037334843073040247, + 0.004945623222738504, + 0.006539683789014816, + 0.005890244618058205, + 0.006407583132386208, + 0.006796484813094139, + 0.006418292410671711, + 0.006252460181713104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.20536823570728302 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042008996009827 + ], + "opposite_logprob_tokens": [ + -1.9104200601577759 + ], + "logprob_reward": -0.16042008996009827, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042008996009827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15972937643527985, + 0.33049869537353516, + 0.04207422584295273, + 0.1192118376493454, + 0.11834883689880371, + 0.04132576286792755, + 0.045467961579561234, + 0.03336845338344574, + 0.039637885987758636, + 0.0703369751572609 + ], + "credit_map": [ + 0.01847214810550213, + 0.02991933561861515, + 0.008198467083275318, + 0.007248581852763891, + 0.048409588634967804, + 0.01259818859398365, + 0.012627633288502693, + 0.022255439311265945, + 0.03057822212576866, + 0.10760243237018585, + 0.02326257899403572, + 0.017815733328461647, + 0.05808611586689949, + 0.014003715477883816, + 0.027905547991394997, + 0.05124436691403389, + 0.005096839740872383, + 0.008239376358687878, + 0.004409133922308683, + 0.0038129121530801058, + 0.0055159940384328365, + 0.005287460517138243, + 0.0037433176767081022, + 0.005969190038740635, + 0.01122459676116705, + 0.026184991002082825, + 0.0069916038773953915, + 0.0113611975684762, + 0.019098829478025436, + 0.014408187009394169, + 0.01332072913646698, + 0.016621703281998634, + 0.033975619822740555, + 0.008942957036197186, + 0.008479555137455463, + 0.007072974927723408, + 0.01143569778650999, + 0.007423035800457001, + 0.029779963195323944, + 0.011239037849009037, + 0.006324971094727516, + 0.003072547260671854, + 0.0031733443029224873, + 0.00513621186837554, + 0.010964570567011833, + 0.006186513230204582, + 0.0037395681720227003, + 0.002728037303313613, + 0.010023458860814571, + 0.0028754102531820536, + 0.002568391850218177, + 0.003630140097811818, + 0.015614837408065796, + 0.003079051151871681, + 0.0043631684966385365, + 0.0033135025296360254, + 0.004049265291541815, + 0.004858394153416157, + 0.004217042587697506, + 0.003273692447692156, + 0.0036811886820942163, + 0.003740317188203335, + 0.005901067052036524, + 0.00364748970605433, + 0.007126617711037397, + 0.005054594483226538, + 0.0069207013584673405, + 0.0034162600059062243, + 0.00555063970386982, + 0.005189104937016964, + 0.002859405241906643, + 0.0035205655731260777, + 0.005337980575859547, + 0.010209106840193272, + 0.007192987482994795, + 0.009453052654862404, + 0.010164512321352959, + 0.01077866367995739, + 0.008250413462519646, + 0.008950257673859596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.16042008996009827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700213760137558 + ], + "opposite_logprob_tokens": [ + -2.695021390914917 + ], + "logprob_reward": -0.0700213760137558, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700213760137558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03362935781478882, + 0.16758732497692108, + 0.040189605206251144, + 0.17211204767227173, + 0.10615120828151703, + 0.047192610800266266, + 0.11186457425355911, + 0.015783607959747314, + 0.2786693871021271, + 0.026820270344614983 + ], + "credit_map": [ + 0.004427335225045681, + 0.005347353406250477, + 0.0029545584693551064, + 0.0030560484156012535, + 0.005337517708539963, + 0.0026838479097932577, + 0.005312735214829445, + 0.004509962163865566, + 0.015544213354587555, + 0.05797693133354187, + 0.010804158635437489, + 0.007420605514198542, + 0.032207194715738297, + 0.0071731493808329105, + 0.013355954550206661, + 0.023105129599571228, + 0.003574889153242111, + 0.00966479443013668, + 0.008690610527992249, + 0.0035950590390712023, + 0.004509155172854662, + 0.00395971117541194, + 0.002441302640363574, + 0.0037540877237915993, + 0.012241615913808346, + 0.018451275303959846, + 0.008039014413952827, + 0.013189626857638359, + 0.019503282383084297, + 0.04583577439188957, + 0.04223378375172615, + 0.012617668136954308, + 0.023984985426068306, + 0.01394602283835411, + 0.010925504378974438, + 0.012419121339917183, + 0.010905446484684944, + 0.009310295805335045, + 0.018518943339586258, + 0.00614088773727417, + 0.013628832995891571, + 0.003327957820147276, + 0.003910965286195278, + 0.004468629136681557, + 0.012535999529063702, + 0.0027704581152647734, + 0.0042215329594910145, + 0.0023282403126358986, + 0.022850455716252327, + 0.010752888396382332, + 0.010540476068854332, + 0.014217144809663296, + 0.02591518871486187, + 0.009313588961958885, + 0.010678056627511978, + 0.0075967879965901375, + 0.0015388549072667956, + 0.001240079989656806, + 0.001466240850277245, + 0.0011792355217039585, + 0.002030265284702182, + 0.001398423220962286, + 0.004995440598577261, + 0.001935067237354815, + 0.04671122878789902, + 0.033793747425079346, + 0.05590158700942993, + 0.026063693687319756, + 0.03771725669503212, + 0.03684068098664284, + 0.01632171869277954, + 0.025319505482912064, + 0.0029136536177247763, + 0.003958000801503658, + 0.002874300116673112, + 0.003174292854964733, + 0.00398437911644578, + 0.004958210978657007, + 0.002848696429282427, + 0.0021087394561618567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.0700213760137558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761513710021973 + ], + "opposite_logprob_tokens": [ + -0.8261513710021973 + ], + "logprob_reward": -0.5761513710021973, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761513710021973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040970832109451294, + 0.19076378643512726, + 0.03352194279432297, + 0.1380896121263504, + 0.10133766382932663, + 0.05493723973631859, + 0.11489184200763702, + 0.01976829022169113, + 0.25028303265571594, + 0.055435772985219955 + ], + "credit_map": [ + 0.005080352537333965, + 0.0066070714965462685, + 0.0042853448539972305, + 0.00348776881583035, + 0.0067452313378453255, + 0.004270211327821016, + 0.006371369119733572, + 0.004123477730900049, + 0.017037592828273773, + 0.06253749877214432, + 0.01322260219603777, + 0.010092493146657944, + 0.03525374457240105, + 0.008022589609026909, + 0.014194012619554996, + 0.030403241515159607, + 0.003686022711917758, + 0.004812970757484436, + 0.0048793223686516285, + 0.0029060393571853638, + 0.006521423812955618, + 0.004225561860948801, + 0.0028535062447190285, + 0.003637094981968403, + 0.010971146635711193, + 0.02140033058822155, + 0.009340924210846424, + 0.00984772015362978, + 0.015398161485791206, + 0.032777752727270126, + 0.024492789059877396, + 0.013860776089131832, + 0.023869551718235016, + 0.011864555068314075, + 0.009878892451524734, + 0.008651294745504856, + 0.009610052220523357, + 0.008579600602388382, + 0.021429182961583138, + 0.0074545275419950485, + 0.013278852216899395, + 0.004235010594129562, + 0.004331931006163359, + 0.006871496327221394, + 0.012535328976809978, + 0.0051067606545984745, + 0.0045479219406843185, + 0.004029933828860521, + 0.02821982279419899, + 0.0049229571595788, + 0.004553775768727064, + 0.007028546184301376, + 0.04351600632071495, + 0.006810600403696299, + 0.012670138850808144, + 0.007169990334659815, + 0.00223620911128819, + 0.0024499413557350636, + 0.002319341991096735, + 0.0022403900511562824, + 0.00213793502189219, + 0.0028254524804651737, + 0.003417490515857935, + 0.0021415287628769875, + 0.04250442609190941, + 0.03044516034424305, + 0.0502408966422081, + 0.023573901504278183, + 0.03398836404085159, + 0.03276034817099571, + 0.014727740548551083, + 0.02204217202961445, + 0.0042755152098834515, + 0.007729196455329657, + 0.005237021949142218, + 0.007149220909923315, + 0.008897061459720135, + 0.008393093943595886, + 0.006230099126696587, + 0.00752456346526742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.5761513710021973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297274485230446 + ], + "opposite_logprob_tokens": [ + -3.7732973098754883 + ], + "logprob_reward": -0.023297274485230446, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297274485230446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10811597108840942, + 0.18153659999370575, + 0.15699689090251923, + 0.0863892138004303, + 0.05364569276571274, + 0.13198308646678925, + 0.04073644056916237, + 0.058068543672561646, + 0.16264519095420837, + 0.01988239772617817 + ], + "credit_map": [ + 0.013105389662086964, + 0.017549613490700722, + 0.007280283607542515, + 0.005383610259741545, + 0.031122470274567604, + 0.010069980286061764, + 0.010301998816430569, + 0.013302627019584179, + 0.018058938905596733, + 0.06077996641397476, + 0.011814531870186329, + 0.008679011836647987, + 0.03531282767653465, + 0.007821495644748211, + 0.015297497622668743, + 0.023772336542606354, + 0.013088331557810307, + 0.042088210582733154, + 0.02739734575152397, + 0.010716430842876434, + 0.026786889880895615, + 0.014449480921030045, + 0.009899054653942585, + 0.012571130879223347, + 0.0063702077604830265, + 0.017681119963526726, + 0.007538446690887213, + 0.006187852006405592, + 0.009053273126482964, + 0.02206057496368885, + 0.008808628655970097, + 0.008689108304679394, + 0.00836933683604002, + 0.00478771235793829, + 0.007882659323513508, + 0.005066244397312403, + 0.007023578509688377, + 0.005094409454613924, + 0.0069138347171247005, + 0.008507913909852505, + 0.034664493054151535, + 0.010821019299328327, + 0.007661531679332256, + 0.011473732069134712, + 0.043017879128456116, + 0.0066982451826334, + 0.011593528091907501, + 0.006052655167877674, + 0.008867649361491203, + 0.0029461586382240057, + 0.0034556975588202477, + 0.004883023910224438, + 0.010763649828732014, + 0.0035232501104474068, + 0.003513443050906062, + 0.002783568575978279, + 0.00759576540440321, + 0.008142451755702496, + 0.006878959946334362, + 0.006236645393073559, + 0.007113039493560791, + 0.00638999929651618, + 0.010681736283004284, + 0.005029947962611914, + 0.027555424720048904, + 0.02022469975054264, + 0.032467011362314224, + 0.015697676688432693, + 0.021922066807746887, + 0.021165400743484497, + 0.008954758755862713, + 0.014658153988420963, + 0.0019680759869515896, + 0.0028663757257163525, + 0.002163135213777423, + 0.0023277588188648224, + 0.002726996084675193, + 0.0025192180182784796, + 0.0029039569199085236, + 0.0024068814236670732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.023297274485230446 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning display of high-contrast photo-realism, a slow, sweeping, low-angle tracking shot glides over a majestic, solitary oak tree, standing defiantly on a grassy knoll beneath a vast, clear blue sky. Its ancient, heavy, gnarled branches, covered in deeply fissured and richly textured bark, snake skyward, their gnarled crown densely laden with ripe acorns. A powerful, unseen gust of wind suddenly strikes the tree, causing its entire massive trunk to bear up and its heavy branches to sway and shake violently. The movement is violent and energetic, with leaves swaying in a chaotic dance. The scene is defined by harsh, brilliant sunlight, creating a stark interplay of light and shadow; deep, inky shadows shift swiftly across the tree's richly textured bark and the sun-kissed leaves, which shimmer between vibrant emerald and near-black, generating a powerful and dramatic primal atmosphere that captures raw, untamed nature with visceral clarity.", + "seed": 5232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_10/index_6559_seed_5232_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": -0.27708019912242887, + "vlm_logprob": -0.27708019912242887, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.27708019912242887, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0554429292678833 + ], + "opposite_logprob_tokens": [ + -0.4304429590702057 + ], + "logprob_reward": -1.0554429292678833, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0554429292678833, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2673420011997223, + 0.1281736046075821, + 0.09668787568807602, + 0.02282743528485298, + 0.0320478193461895, + 0.11447564512491226, + 0.10828202962875366, + 0.04335048422217369, + 0.06532878428697586, + 0.12148439139127731 + ], + "credit_map": [ + 0.046171028167009354, + 0.03253616765141487, + 0.02877349592745304, + 0.018190138041973114, + 0.04760078340768814, + 0.03228959068655968, + 0.03334393724799156, + 0.02843686379492283, + 0.02428380213677883, + 0.047078829258680344, + 0.009764575399458408, + 0.010676387697458267, + 0.015974439680576324, + 0.007633436005562544, + 0.00866820476949215, + 0.004093932453542948, + 0.007146102841943502, + 0.016251493245363235, + 0.007494793739169836, + 0.013922566547989845, + 0.018896929919719696, + 0.009634662419557571, + 0.015758881345391273, + 0.007582443300634623, + 0.0028128032572567463, + 0.0032636732794344425, + 0.0017429229337722063, + 0.005257264710962772, + 0.003392269369214773, + 0.0018252604641020298, + 0.002733442699536681, + 0.0017997974064201117, + 0.0037746771704405546, + 0.006577152758836746, + 0.00340943387709558, + 0.005761317443102598, + 0.003341513220220804, + 0.002612649230286479, + 0.004949256777763367, + 0.0016218216624110937, + 0.008675698190927505, + 0.018015317618846893, + 0.01095374021679163, + 0.021329054608941078, + 0.014038578607141972, + 0.011103031225502491, + 0.018011610954999924, + 0.01234861183911562, + 0.008957631886005402, + 0.018112799152731895, + 0.010678541846573353, + 0.017193365842103958, + 0.016797160729765892, + 0.01279664970934391, + 0.015474164858460426, + 0.0082717165350914, + 0.005832018796354532, + 0.007140509784221649, + 0.006073267199099064, + 0.00841396301984787, + 0.004907307215034962, + 0.00469991285353899, + 0.004136048723012209, + 0.002147456631064415, + 0.016340646892786026, + 0.008315288461744785, + 0.009990865364670753, + 0.012232608161866665, + 0.005074344575405121, + 0.005107608158141375, + 0.004931803327053785, + 0.00333561678417027, + 0.01199500821530819, + 0.022295592352747917, + 0.021027840673923492, + 0.02734393998980522, + 0.009583359584212303, + 0.012525010854005814, + 0.008465428836643696, + 0.008248208090662956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -1.0554429292678833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310210794210434 + ], + "opposite_logprob_tokens": [ + -3.1681020259857178 + ], + "logprob_reward": -0.04310210794210434, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310210794210434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20361167192459106, + 0.24015749990940094, + 0.09214099496603012, + 0.03240884467959404, + 0.031717635691165924, + 0.05695819482207298, + 0.05368164926767349, + 0.0513131357729435, + 0.07364039123058319, + 0.16437003016471863 + ], + "credit_map": [ + 0.03403143584728241, + 0.02780921384692192, + 0.018297463655471802, + 0.011493632569909096, + 0.049738090485334396, + 0.0250431876629591, + 0.022077247500419617, + 0.015121386386454105, + 0.03997392952442169, + 0.08819983899593353, + 0.017041955143213272, + 0.019137738272547722, + 0.03460008278489113, + 0.01575850509107113, + 0.015384720638394356, + 0.010060722939670086, + 0.014147537760436535, + 0.017807764932513237, + 0.009432971477508545, + 0.016904689371585846, + 0.009211848489940166, + 0.005424694158136845, + 0.01352342963218689, + 0.005688057281076908, + 0.0038244507741183043, + 0.0051635270938277245, + 0.002513176528736949, + 0.008447840809822083, + 0.004027593415230513, + 0.0030848938040435314, + 0.0033271063584834337, + 0.002020255895331502, + 0.0033642842900007963, + 0.00548284687101841, + 0.003300964832305908, + 0.005748437251895666, + 0.003511194372549653, + 0.0027538472786545753, + 0.005994871724396944, + 0.001561189303174615, + 0.005965136922895908, + 0.008444715291261673, + 0.0065372781828045845, + 0.010468590073287487, + 0.006575403269380331, + 0.00496880616992712, + 0.008661840111017227, + 0.005336420144885778, + 0.0046311947517097, + 0.009117615409195423, + 0.005052837543189526, + 0.008044478483498096, + 0.009004224091768265, + 0.006649107672274113, + 0.007307321298867464, + 0.003874871414154768, + 0.00633193738758564, + 0.008462943136692047, + 0.006922414526343346, + 0.01025985274463892, + 0.006781153846532106, + 0.004771954379975796, + 0.005189819727092981, + 0.002593059791252017, + 0.008095892146229744, + 0.009922241792082787, + 0.012790316715836525, + 0.0186496302485466, + 0.004625327419489622, + 0.0074371895752847195, + 0.006370886228978634, + 0.0057489038445055485, + 0.01789385825395584, + 0.028945058584213257, + 0.028405170887708664, + 0.03689417615532875, + 0.01345355249941349, + 0.015669116750359535, + 0.011565161868929863, + 0.011543931439518929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.04310210794210434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067626655101776 + ], + "opposite_logprob_tokens": [ + -3.168067693710327 + ], + "logprob_reward": -0.043067626655101776, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067626655101776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08872207999229431, + 0.184306338429451, + 0.04517839103937149, + 0.03267761319875717, + 0.0366852842271328, + 0.11233873665332794, + 0.10267297178506851, + 0.13495029509067535, + 0.21448880434036255, + 0.047979485243558884 + ], + "credit_map": [ + 0.010547657497227192, + 0.016785766929388046, + 0.008568717166781425, + 0.008967668749392033, + 0.008205393329262733, + 0.012868793681263924, + 0.012505865655839443, + 0.010272222571074963, + 0.02350286953151226, + 0.06675346940755844, + 0.010662724263966084, + 0.014768246561288834, + 0.033669184893369675, + 0.016812706366181374, + 0.010781758464872837, + 0.007355382665991783, + 0.005710352677851915, + 0.006340102292597294, + 0.004522448871284723, + 0.007185711991041899, + 0.005999189801514149, + 0.003996188286691904, + 0.007688639219850302, + 0.003735756268724799, + 0.0038788740057498217, + 0.005170198157429695, + 0.003565732855349779, + 0.0075852638110518456, + 0.0030549876391887665, + 0.0023996199015527964, + 0.0036370770540088415, + 0.003385859774425626, + 0.005696445237845182, + 0.006968789733946323, + 0.0043308325111866, + 0.005436934996396303, + 0.004073590971529484, + 0.0038214772939682007, + 0.004129152279347181, + 0.002228062367066741, + 0.00977066159248352, + 0.018138596788048744, + 0.010163718834519386, + 0.018553612753748894, + 0.01561648491770029, + 0.010623211041092873, + 0.01669379137456417, + 0.01277865469455719, + 0.009724892675876617, + 0.015985220670700073, + 0.013737759552896023, + 0.02000931091606617, + 0.0109111238270998, + 0.00989323016256094, + 0.011418730951845646, + 0.010992702096700668, + 0.01703350432217121, + 0.0255082119256258, + 0.016361169517040253, + 0.023644646629691124, + 0.018090631812810898, + 0.011438150890171528, + 0.015763740986585617, + 0.007110236678272486, + 0.03882488235831261, + 0.03886273503303528, + 0.03447418659925461, + 0.04160711169242859, + 0.01517950464040041, + 0.017472241073846817, + 0.014134779572486877, + 0.013933352194726467, + 0.006608461495488882, + 0.007122870068997145, + 0.0076996623538434505, + 0.012040888890624046, + 0.0036546357441693544, + 0.0037068764213472605, + 0.00347269419580698, + 0.003673398634418845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.043067626655101776 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.225591778755188 + ], + "opposite_logprob_tokens": [ + -1.600591778755188 + ], + "logprob_reward": -0.225591778755188, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.225591778755188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06838498264551163, + 0.096063993871212, + 0.07843618094921112, + 0.22309693694114685, + 0.04236337170004845, + 0.05769113451242447, + 0.09851252287626266, + 0.09309712052345276, + 0.19394037127494812, + 0.04841335490345955 + ], + "credit_map": [ + 0.007940642535686493, + 0.008860526606440544, + 0.007594233378767967, + 0.004572994075715542, + 0.012756061740219593, + 0.010155590251088142, + 0.007160252425819635, + 0.009344684891402721, + 0.017710285261273384, + 0.032280441373586655, + 0.0074669667519629, + 0.007895546965301037, + 0.010187470354139805, + 0.00784263014793396, + 0.009251308627426624, + 0.00342934369109571, + 0.01378527469933033, + 0.010265450924634933, + 0.0113642867654562, + 0.010847736150026321, + 0.009964117780327797, + 0.007165498565882444, + 0.009272023104131222, + 0.0057717920280992985, + 0.038770925253629684, + 0.032378900796175, + 0.015574484132230282, + 0.04650717228651047, + 0.0244514849036932, + 0.024239474907517433, + 0.025599585846066475, + 0.01557491160929203, + 0.003969445358961821, + 0.007456093095242977, + 0.0037618507631123066, + 0.007085857447236776, + 0.005210882984101772, + 0.00396987097337842, + 0.008284452371299267, + 0.002624920569360256, + 0.007189157884567976, + 0.009959986433386803, + 0.008688166737556458, + 0.00628920691087842, + 0.007656492758542299, + 0.005799500737339258, + 0.007095303852111101, + 0.005013317335397005, + 0.008778927847743034, + 0.01591017283499241, + 0.010078289546072483, + 0.0187897477298975, + 0.0127538638189435, + 0.00894097238779068, + 0.014189540408551693, + 0.009071014821529388, + 0.009886658750474453, + 0.014664418995380402, + 0.012995623983442783, + 0.01541775930672884, + 0.015480189584195614, + 0.010598409920930862, + 0.009034516289830208, + 0.005019542761147022, + 0.033281367272138596, + 0.03086431697010994, + 0.02830199897289276, + 0.041854966431856155, + 0.01120014674961567, + 0.02004457637667656, + 0.017652451992034912, + 0.010740536265075207, + 0.006356684025377035, + 0.007008283864706755, + 0.007201184052973986, + 0.00956711731851101, + 0.004291241057217121, + 0.005302868317812681, + 0.004132889676839113, + 0.004553085658699274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.225591778755188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary oak tree with ancient, gnarled branches and a crown laden with ripe acorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tree standing on a grassy knoll beneath a vast, clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, sweeping, low-angle tracking shot gliding over the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sudden gust of wind cause the tree's trunk to bear up and its branches to sway and shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, brilliant sunlight creating stark, deep shadows and high contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196552991867065 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196552991867065, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196552991867065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09031295776367188, + 0.07009280472993851, + 0.06866737455129623, + 0.0640307143330574, + 0.05317765101790428, + 0.1286262422800064, + 0.04857517406344414, + 0.17098528146743774, + 0.09451106190681458, + 0.21102078258991241 + ], + "credit_map": [ + 0.013149222359061241, + 0.011443021707236767, + 0.007471765857189894, + 0.008435181342065334, + 0.019124895334243774, + 0.010856549255549908, + 0.011118100956082344, + 0.008714216761291027, + 0.010490700602531433, + 0.016387056559324265, + 0.008256079629063606, + 0.004285251721739769, + 0.017204422503709793, + 0.006364329718053341, + 0.00453002005815506, + 0.002574944868683815, + 0.01256402675062418, + 0.012504441663622856, + 0.010053049772977829, + 0.009354858659207821, + 0.0067773559130728245, + 0.005362706258893013, + 0.007137073203921318, + 0.004913860931992531, + 0.01130982767790556, + 0.010276627726852894, + 0.004513775929808617, + 0.012954160571098328, + 0.007068573031574488, + 0.008351103402674198, + 0.0050495695322752, + 0.004507072735577822, + 0.008263935334980488, + 0.009229638613760471, + 0.0058264220133423805, + 0.006721018347889185, + 0.005419076420366764, + 0.00798275787383318, + 0.004570283927023411, + 0.005164520815014839, + 0.01780039444565773, + 0.019315628334879875, + 0.019684532657265663, + 0.01957419328391552, + 0.013818135485053062, + 0.011770817451179028, + 0.014859341084957123, + 0.011803191155195236, + 0.003547752508893609, + 0.006870802957564592, + 0.006073138676583767, + 0.010979843325912952, + 0.005401606671512127, + 0.005340880248695612, + 0.005543768871575594, + 0.004817378707230091, + 0.020228561013936996, + 0.026157567277550697, + 0.024621564894914627, + 0.023302046582102776, + 0.035671334713697433, + 0.015423781238496304, + 0.016066687181591988, + 0.009513731114566326, + 0.018157383427023888, + 0.012939332984387875, + 0.01448098849505186, + 0.018389452248811722, + 0.008291970007121563, + 0.009335910901427269, + 0.006725038401782513, + 0.0061909849755465984, + 0.022016126662492752, + 0.036069050431251526, + 0.035877589136362076, + 0.053483445197343826, + 0.01544551458209753, + 0.01791580393910408, + 0.014748933725059032, + 0.01546430867165327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.018196552991867065 + } + ], + "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/0000128/group_12/index_tensor([292, 292])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_12/index_tensor([292, 292])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b425c22561eb4518023361c27260f9e20fc1b051 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_12/index_tensor([292, 292])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle shot rapidly moves upward from a ruined asphalt road to reveal a massive mecha against the backdrop of the scarred Tokyo Tower at dusk. The mecha's design blends realism and fantasy, constructed from polished white and dark blue metal panels, now worn and charred. Pulsating cyan light emanates from its vents and radiates along its panel lines, casting an ethereal glow over the surrounding smoke. The animation emphasizes its immense weight; with ponderous movements, the mecha extends its massive feet, causing the ground to tremble and crack. Visible through a translucent holographic display, within the cockpit lies a 16-year-old boy with sharp, stylized anime features, spiky blue hair, and a resolute gaze, gripping the controls, his knuckles white. As he decisively presses the throttle forward, the mecha's thrusters roar and ignite, sending it hurtling forward, its fists a blur of speed, trailed by sharp lines of motion through the air. The punch connects with the unseen enemy, triggering a catastrophic explosion; the scene is instantly engulfed in the impact frame of blinding white light, followed by a violent outward blast of fire, smoke, and flying debris, illuminating the intense, high-stakes atmosphere of the Dusk Battle.", + "seed": 32693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_12/index_292_seed_32693_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": -0.03842509146779775, + "vlm_logprob": -0.03842509146779775, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.03842509146779775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0237870030105114 + ], + "opposite_logprob_tokens": [ + -3.898787021636963 + ], + "logprob_reward": -0.0237870030105114, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0237870030105114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40708479285240173, + 0.054731015115976334, + 0.10058289766311646, + 0.09827661514282227, + 0.023403896018862724, + 0.15728721022605896, + 0.06398171186447144, + 0.021854186430573463, + 0.032680902630090714, + 0.040116775780916214 + ], + "credit_map": [ + 0.05965175852179527, + 0.041380178183317184, + 0.07649820297956467, + 0.04050217196345329, + 0.04434553161263466, + 0.024497706443071365, + 0.06921954452991486, + 0.05098966881632805, + 0.005210471339523792, + 0.006126856431365013, + 0.012554428540170193, + 0.005771626718342304, + 0.004583477973937988, + 0.004376666620373726, + 0.0099768852815032, + 0.006130602210760117, + 0.016495468094944954, + 0.01047196239233017, + 0.016586344689130783, + 0.013920804485678673, + 0.01061407383531332, + 0.011961270123720169, + 0.013425610959529877, + 0.00710736308246851, + 0.01538039930164814, + 0.013176308013498783, + 0.018729915842413902, + 0.009295198135077953, + 0.008613056503236294, + 0.012992637231945992, + 0.010344307869672775, + 0.0097447969019413, + 0.0042052315548062325, + 0.0030352012254297733, + 0.0015044996980577707, + 0.0017724281642585993, + 0.0019328721100464463, + 0.003169481409713626, + 0.0029024232644587755, + 0.004881759639829397, + 0.017223693430423737, + 0.020563043653964996, + 0.020415661856532097, + 0.01243763230741024, + 0.02619246020913124, + 0.016550268977880478, + 0.016683192923665047, + 0.02722126618027687, + 0.009202424436807632, + 0.010671310126781464, + 0.008113970048725605, + 0.004824383649975061, + 0.01133011281490326, + 0.006012143567204475, + 0.006374882534146309, + 0.007452491205185652, + 0.0023818861227482557, + 0.0028951517306268215, + 0.0026277415454387665, + 0.0026413823943585157, + 0.0018677134066820145, + 0.0017241378081962466, + 0.0029158038087189198, + 0.004800369031727314, + 0.003610741114243865, + 0.00694564264267683, + 0.003403951646760106, + 0.0031996320467442274, + 0.004345489200204611, + 0.003951257560402155, + 0.0038279693108052015, + 0.003396216779947281, + 0.003190455725416541, + 0.004772939719259739, + 0.0029854655731469393, + 0.00345050310716033, + 0.007276153191924095, + 0.006736949551850557, + 0.005151099991053343, + 0.006553205661475658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0237870030105114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372172638773918 + ], + "opposite_logprob_tokens": [ + -3.408721685409546 + ], + "logprob_reward": -0.03372172638773918, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372172638773918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19226527214050293, + 0.08940308541059494, + 0.13598524034023285, + 0.07701674103736877, + 0.09544526785612106, + 0.12139328569173813, + 0.1093841940164566, + 0.03799733519554138, + 0.09481797367334366, + 0.04629155620932579 + ], + "credit_map": [ + 0.020578661933541298, + 0.01735719107091427, + 0.03351986035704613, + 0.01636413484811783, + 0.024235354736447334, + 0.01583859883248806, + 0.03980187699198723, + 0.024569597095251083, + 0.009468914940953255, + 0.010186896659433842, + 0.01703881099820137, + 0.013383596204221249, + 0.009915612637996674, + 0.0056600808165967464, + 0.016235310584306717, + 0.007513862568885088, + 0.02000538632273674, + 0.01085383165627718, + 0.026962101459503174, + 0.014930718578398228, + 0.011978828348219395, + 0.015593066811561584, + 0.024276379495859146, + 0.01138492301106453, + 0.010970734991133213, + 0.008882886730134487, + 0.01636319048702717, + 0.008640335872769356, + 0.006810359191149473, + 0.012060080654919147, + 0.0063476888462901115, + 0.006941471714526415, + 0.017579548060894012, + 0.0053046829998493195, + 0.00790700875222683, + 0.005093934014439583, + 0.005497519392520189, + 0.021113570779561996, + 0.005311259999871254, + 0.027637744322419167, + 0.014203081838786602, + 0.015319319441914558, + 0.013476802967488766, + 0.01636667549610138, + 0.01592920906841755, + 0.017859499901533127, + 0.014525611884891987, + 0.013713081367313862, + 0.013369789347052574, + 0.009775451384484768, + 0.015351353213191032, + 0.008146075531840324, + 0.014287885278463364, + 0.014398420229554176, + 0.013519338332116604, + 0.02053588256239891, + 0.004782726988196373, + 0.005891508422791958, + 0.004456839058548212, + 0.005644447635859251, + 0.0034313108772039413, + 0.0028549181297421455, + 0.005239767953753471, + 0.005695815198123455, + 0.01281209196895361, + 0.012564333155751228, + 0.00738125666975975, + 0.010223569348454475, + 0.016332417726516724, + 0.014323526062071323, + 0.01145177986472845, + 0.009728997945785522, + 0.003002628218382597, + 0.004312623292207718, + 0.0031264794524759054, + 0.0035243043676018715, + 0.009882597252726555, + 0.007533549796789885, + 0.007375449873507023, + 0.007533923722803593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.03372172638773918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554424814879894 + ], + "opposite_logprob_tokens": [ + -4.38755464553833 + ], + "logprob_reward": -0.012554424814879894, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554424814879894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18070784211158752, + 0.0984245166182518, + 0.0823427215218544, + 0.07038713991641998, + 0.07407811284065247, + 0.06051988899707794, + 0.1818765550851822, + 0.09570526331663132, + 0.07486169785261154, + 0.08109624683856964 + ], + "credit_map": [ + 0.022807259112596512, + 0.018318239599466324, + 0.02569740079343319, + 0.018714943900704384, + 0.03079097345471382, + 0.012585780583322048, + 0.02494296245276928, + 0.026850279420614243, + 0.01266133040189743, + 0.015594649128615856, + 0.014254017733037472, + 0.010961960069835186, + 0.010317958891391754, + 0.010042170062661171, + 0.013403496704995632, + 0.011188938282430172, + 0.010193566791713238, + 0.009460714645683765, + 0.013823996298015118, + 0.008644318208098412, + 0.009755817241966724, + 0.007620652671903372, + 0.013914501294493675, + 0.008929155766963959, + 0.010770370252430439, + 0.00855126790702343, + 0.011446447111666203, + 0.005861970596015453, + 0.0065681287087500095, + 0.013304474763572216, + 0.00707683339715004, + 0.006807652302086353, + 0.014361828565597534, + 0.004375170450657606, + 0.00533342594280839, + 0.005662525072693825, + 0.0041760108433663845, + 0.014601320959627628, + 0.0047151087783277035, + 0.020852725952863693, + 0.007283126935362816, + 0.007717351894825697, + 0.00519470265135169, + 0.004646683111786842, + 0.008701609447598457, + 0.011124461889266968, + 0.007865721359848976, + 0.007986229844391346, + 0.020360972732305527, + 0.018348637968301773, + 0.023625921458005905, + 0.016609719023108482, + 0.024095864966511726, + 0.02335250750184059, + 0.023109564557671547, + 0.03237336874008179, + 0.012094378471374512, + 0.01087185274809599, + 0.010520365089178085, + 0.011261587962508202, + 0.012494560331106186, + 0.012515421956777573, + 0.013187783770263195, + 0.012759307399392128, + 0.0076242778450250626, + 0.01679805852472782, + 0.007122117560356855, + 0.009194246493279934, + 0.008278943598270416, + 0.01009705476462841, + 0.008240588009357452, + 0.007506412919610739, + 0.004612762480974197, + 0.0053134486079216, + 0.004403383005410433, + 0.005408055614680052, + 0.019317517057061195, + 0.013427953235805035, + 0.014075040817260742, + 0.014538091607391834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.012554424814879894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900232821702957 + ], + "opposite_logprob_tokens": [ + -2.5790023803710938 + ], + "logprob_reward": -0.07900232821702957, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900232821702957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1383615881204605, + 0.03773104399442673, + 0.05073424428701401, + 0.046555113047361374, + 0.16656318306922913, + 0.06274942308664322, + 0.30727511644363403, + 0.04422111436724663, + 0.09015508741140366, + 0.0556541383266449 + ], + "credit_map": [ + 0.01574494130909443, + 0.0157034732401371, + 0.017992647364735603, + 0.01597405970096588, + 0.020357882604002953, + 0.014250761829316616, + 0.021380135789513588, + 0.01695767045021057, + 0.003700028173625469, + 0.004358308389782906, + 0.007192784454673529, + 0.004400078672915697, + 0.003958508372306824, + 0.002949010580778122, + 0.006546796299517155, + 0.004625525325536728, + 0.005879226606339216, + 0.005321657285094261, + 0.0077412617392838, + 0.008283055387437344, + 0.005822949577122927, + 0.006550690624862909, + 0.0071276589296758175, + 0.0040077390149235725, + 0.005890915170311928, + 0.005655135493725538, + 0.010051409713923931, + 0.003733158577233553, + 0.0042057340033352375, + 0.00803622230887413, + 0.004601922817528248, + 0.004380606114864349, + 0.029806697741150856, + 0.010535449720919132, + 0.012482384219765663, + 0.012645739130675793, + 0.01144903339445591, + 0.03613709285855293, + 0.011811810545623302, + 0.04169495776295662, + 0.007501006126403809, + 0.007302156649529934, + 0.006244631949812174, + 0.0047542862594127655, + 0.008492004126310349, + 0.011552729643881321, + 0.007483604829758406, + 0.009418998844921589, + 0.0389055572450161, + 0.02445121482014656, + 0.04567639157176018, + 0.02401239238679409, + 0.03628544881939888, + 0.042203307151794434, + 0.037723615765571594, + 0.05801716446876526, + 0.004161539487540722, + 0.007896663621068, + 0.005628641694784164, + 0.00661437725648284, + 0.0039686961099505424, + 0.002784902695566416, + 0.004821246489882469, + 0.008345039561390877, + 0.008303853683173656, + 0.02445882000029087, + 0.00860362034291029, + 0.011081690900027752, + 0.00754885608330369, + 0.010168151929974556, + 0.010543030686676502, + 0.009447058662772179, + 0.004838850814849138, + 0.007120933849364519, + 0.005597967188805342, + 0.006944203749299049, + 0.008484004065394402, + 0.00901807565242052, + 0.006006747484207153, + 0.007643349468708038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.07900232821702957 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043059974908828735 + ], + "opposite_logprob_tokens": [ + -3.168060064315796 + ], + "logprob_reward": -0.043059974908828735, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043059974908828735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10807125270366669, + 0.09959816932678223, + 0.04494168236851692, + 0.026041880249977112, + 0.030242152512073517, + 0.3070432245731354, + 0.07232166826725006, + 0.030963746830821037, + 0.10619426518678665, + 0.17458194494247437 + ], + "credit_map": [ + 0.011160545982420444, + 0.008946684189140797, + 0.018341807648539543, + 0.011672922410070896, + 0.013074898160994053, + 0.006516367197036743, + 0.02501770481467247, + 0.013340325094759464, + 0.00868377648293972, + 0.013961770571768284, + 0.021701214835047722, + 0.009955045767128468, + 0.010559058748185635, + 0.00732611957937479, + 0.018825609236955643, + 0.008585566654801369, + 0.005680248606950045, + 0.005549262743443251, + 0.008524134755134583, + 0.006479648407548666, + 0.0031940648332238197, + 0.0026915231719613075, + 0.008765047416090965, + 0.004057752434164286, + 0.004117549397051334, + 0.0027173454873263836, + 0.004387566354125738, + 0.002801832277327776, + 0.0021759127266705036, + 0.0033299929928034544, + 0.002113461261615157, + 0.0043982211500406265, + 0.003993747290223837, + 0.00270868674851954, + 0.002358390251174569, + 0.0033335131593048573, + 0.0025283924769610167, + 0.0046876403503119946, + 0.0031981116626411676, + 0.007433671038597822, + 0.03647420182824135, + 0.03870531544089317, + 0.03862506151199341, + 0.02721255086362362, + 0.047266170382499695, + 0.02969605103135109, + 0.03445085510611534, + 0.05461303889751434, + 0.007515189703553915, + 0.007689906284213066, + 0.00907896924763918, + 0.007978891022503376, + 0.010147975757718086, + 0.008078760467469692, + 0.00969715416431427, + 0.012134815566241741, + 0.0027619244065135717, + 0.0051085008308291435, + 0.0033285922836512327, + 0.005384530406445265, + 0.003349276026710868, + 0.002595676574856043, + 0.0037378817796707153, + 0.004697366151958704, + 0.01199372112751007, + 0.0240371935069561, + 0.011974604800343513, + 0.012277616187930107, + 0.009421108290553093, + 0.010592383332550526, + 0.013320419937372208, + 0.01257721520960331, + 0.008632191456854343, + 0.00899476744234562, + 0.009491492994129658, + 0.014707493595778942, + 0.03807178512215614, + 0.029075223952531815, + 0.03394505754113197, + 0.0316639170050621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.043059974908828735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot rapidly moves upward from a ruined asphalt road to reveal a massive mecha against the backdrop of the scarred Tokyo Tower at dusk. The mecha's design blends realism and fantasy, constructed from polished white and dark blue metal panels, now worn and charred. Pulsating cyan light emanates from its vents and radiates along its panel lines, casting an ethereal glow over the surrounding smoke. The animation emphasizes its immense weight; with ponderous movements, the mecha extends its massive feet, causing the ground to tremble and crack. Visible through a translucent holographic display, within the cockpit lies a 16-year-old boy with sharp, stylized anime features, spiky blue hair, and a resolute gaze, gripping the controls, his knuckles white. As he decisively presses the throttle forward, the mecha's thrusters roar and ignite, sending it hurtling forward, its fists a blur of speed, trailed by sharp lines of motion through the air. The punch connects with the unseen enemy, triggering a catastrophic explosion; the scene is instantly engulfed in the impact frame of blinding white light, followed by a violent outward blast of fire, smoke, and flying debris, illuminating the intense, high-stakes atmosphere of the Dusk Battle.", + "seed": 32694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_12/index_292_seed_32694_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": -0.030964172445237637, + "vlm_logprob": -0.030964172445237637, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.030964172445237637, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03132764250040054 + ], + "opposite_logprob_tokens": [ + -3.531327724456787 + ], + "logprob_reward": -0.03132764250040054, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03132764250040054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31825414299964905, + 0.13156846165657043, + 0.057171858847141266, + 0.05166669934988022, + 0.2518555521965027, + 0.042561329901218414, + 0.03226736932992935, + 0.04744404926896095, + 0.025992782786488533, + 0.04121778532862663 + ], + "credit_map": [ + 0.0434999018907547, + 0.04118173196911812, + 0.039223890751600266, + 0.04483406990766525, + 0.01923665776848793, + 0.03948390483856201, + 0.030700484290719032, + 0.0600934773683548, + 0.006725760642439127, + 0.00822940468788147, + 0.031438764184713364, + 0.01301109790802002, + 0.00827611330896616, + 0.003245860803872347, + 0.04896296188235283, + 0.011678491719067097, + 0.0046880957670509815, + 0.004555712454020977, + 0.0069686127826571465, + 0.011339210905134678, + 0.0063136909157037735, + 0.005874975118786097, + 0.006183151621371508, + 0.011248409748077393, + 0.0064785173162817955, + 0.014536695554852486, + 0.004206477664411068, + 0.00933209341019392, + 0.005949290003627539, + 0.0035230284556746483, + 0.0045783668756484985, + 0.0030622270423918962, + 0.010910016484558582, + 0.020126255229115486, + 0.012066507712006569, + 0.08004415780305862, + 0.012229346670210361, + 0.01885547861456871, + 0.03692055121064186, + 0.06070322543382645, + 0.006141075398772955, + 0.00549087580293417, + 0.00550197483971715, + 0.005801497958600521, + 0.004645469132810831, + 0.005839219316840172, + 0.004103464540094137, + 0.005037750117480755, + 0.0038502304814755917, + 0.004832413047552109, + 0.005220675840973854, + 0.0025203339755535126, + 0.003644562792032957, + 0.005036374554038048, + 0.002333431039005518, + 0.0048293499276041985, + 0.008114917203783989, + 0.004622624721378088, + 0.006797913461923599, + 0.005513788666576147, + 0.005056004039943218, + 0.0049624270759522915, + 0.005473394878208637, + 0.006902977358549833, + 0.003069722093641758, + 0.0025232385378330946, + 0.0032357927411794662, + 0.004370520822703838, + 0.003284544451162219, + 0.0039233858697116375, + 0.002841315697878599, + 0.0027442644350230694, + 0.001682561356574297, + 0.0033827286679297686, + 0.002790428465232253, + 0.0076707499101758, + 0.009129902347922325, + 0.007963172160089016, + 0.005405894480645657, + 0.003192346775904298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.03132764250040054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812266141176224 + ], + "opposite_logprob_tokens": [ + -3.2881226539611816 + ], + "logprob_reward": -0.03812266141176224, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812266141176224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5035355687141418, + 0.1608513593673706, + 0.08875652402639389, + 0.07008674740791321, + 0.01596042700111866, + 0.025070296600461006, + 0.02916175127029419, + 0.04216165095567703, + 0.011481275781989098, + 0.05293433740735054 + ], + "credit_map": [ + 0.11350685358047485, + 0.08133059740066528, + 0.06438612937927246, + 0.03923085704445839, + 0.016979120671749115, + 0.032519858330488205, + 0.028614919632673264, + 0.12696723639965057, + 0.008003320544958115, + 0.007382791955024004, + 0.06098725646734238, + 0.03364202752709389, + 0.007834022864699364, + 0.003500858088955283, + 0.027984803542494774, + 0.011516297236084938, + 0.006381594110280275, + 0.008856227621436119, + 0.011494368314743042, + 0.01717229187488556, + 0.009088455699384212, + 0.006994387600570917, + 0.013254369609057903, + 0.015514844097197056, + 0.00941272173076868, + 0.01261186134070158, + 0.004621995612978935, + 0.013240068219602108, + 0.0067161074839532375, + 0.0030462746508419514, + 0.015375814400613308, + 0.005061910022050142, + 0.0011427761055529118, + 0.0015074405819177628, + 0.0018960008164867759, + 0.0036100607831031084, + 0.0012237268965691328, + 0.001685935421846807, + 0.0014775281306356192, + 0.003416958265006542, + 0.005620008800178766, + 0.002614903263747692, + 0.0025955706369131804, + 0.003067422891035676, + 0.002465300727635622, + 0.003580917604267597, + 0.0023823014926165342, + 0.002743874443694949, + 0.003281526267528534, + 0.005488796625286341, + 0.00471093924716115, + 0.0024247330147773027, + 0.0037513920105993748, + 0.002958723809570074, + 0.002581260399892926, + 0.003964379895478487, + 0.007760456297546625, + 0.004289586097002029, + 0.005598478950560093, + 0.0051426962018013, + 0.004503605421632528, + 0.004470023792237043, + 0.004753418266773224, + 0.0056433905847370625, + 0.0013422024203464389, + 0.0010120613733306527, + 0.0017060190439224243, + 0.0013489137636497617, + 0.0015898155979812145, + 0.00165432202629745, + 0.0014939422253519297, + 0.0013340000296011567, + 0.003067169338464737, + 0.004345495719462633, + 0.003646071068942547, + 0.008390632458031178, + 0.013025997206568718, + 0.01028341893106699, + 0.005718541331589222, + 0.00445701414719224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03812266141176224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798760525882244 + ], + "opposite_logprob_tokens": [ + -4.634798526763916 + ], + "logprob_reward": -0.009798760525882244, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798760525882244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13136479258537292, + 0.07569107413291931, + 0.052219174802303314, + 0.1540265530347824, + 0.08246874809265137, + 0.10270123183727264, + 0.06580808013677597, + 0.10870996862649918, + 0.13578644394874573, + 0.09122387319803238 + ], + "credit_map": [ + 0.01712391898036003, + 0.01972082629799843, + 0.017174728214740753, + 0.019967131316661835, + 0.012732023373246193, + 0.018522774800658226, + 0.011177013628184795, + 0.014946391806006432, + 0.005496184341609478, + 0.004739302210509777, + 0.010224924422800541, + 0.01649954542517662, + 0.00652867928147316, + 0.004001479130238295, + 0.017621565610170364, + 0.010579394176602364, + 0.004143534693866968, + 0.0058476561680436134, + 0.005999841261655092, + 0.008028209209442139, + 0.0074016940779984, + 0.005795277655124664, + 0.007417149376124144, + 0.007585811894387007, + 0.030392693355679512, + 0.02670699916779995, + 0.013110131025314331, + 0.026750274002552032, + 0.00999093521386385, + 0.009583079256117344, + 0.02952779456973076, + 0.007964652962982655, + 0.005012196023017168, + 0.009283877909183502, + 0.005570890847593546, + 0.013378087431192398, + 0.016059178858995438, + 0.01275898702442646, + 0.011577547527849674, + 0.008827987127006054, + 0.012666781432926655, + 0.015362418256700039, + 0.015317331068217754, + 0.012591524980962276, + 0.014521862380206585, + 0.011384722776710987, + 0.010566579177975655, + 0.010290024802088737, + 0.009514018893241882, + 0.007145883981138468, + 0.010153219103813171, + 0.005199481267482042, + 0.008899188600480556, + 0.009058569557964802, + 0.00620995182543993, + 0.00962777528911829, + 0.021439289674162865, + 0.014926339499652386, + 0.010529977269470692, + 0.016010908409953117, + 0.016241028904914856, + 0.008571607992053032, + 0.011614307761192322, + 0.0093765240162611, + 0.014946765266358852, + 0.01807912066578865, + 0.014131573028862476, + 0.012421500869095325, + 0.03310257941484451, + 0.019029751420021057, + 0.014576015062630177, + 0.009499138221144676, + 0.008058568462729454, + 0.00683830538764596, + 0.007643101271241903, + 0.009023482911288738, + 0.0233963243663311, + 0.0185704343020916, + 0.00792899914085865, + 0.009764672257006168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.009798760525882244 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498318374156952 + ], + "opposite_logprob_tokens": [ + -2.929983139038086 + ], + "logprob_reward": -0.05498318374156952, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498318374156952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1976148933172226, + 0.08829676359891891, + 0.05190812796354294, + 0.11955474317073822, + 0.0676867738366127, + 0.040372125804424286, + 0.040538664907217026, + 0.21471062302589417, + 0.08362200111150742, + 0.09569529443979263 + ], + "credit_map": [ + 0.03181909769773483, + 0.022290928289294243, + 0.030030911788344383, + 0.0235357154160738, + 0.009547783993184566, + 0.013246848247945309, + 0.013328426517546177, + 0.053815193474292755, + 0.004634852986782789, + 0.009457388892769814, + 0.016759568825364113, + 0.00862938817590475, + 0.0057850126177072525, + 0.003002496901899576, + 0.027270253747701645, + 0.012757803313434124, + 0.0036324062384665012, + 0.004779082257300615, + 0.005891331471502781, + 0.009334906935691833, + 0.009519748389720917, + 0.00412167189642787, + 0.009124931879341602, + 0.0055040521547198296, + 0.014874308370053768, + 0.009701103903353214, + 0.00791149027645588, + 0.04196012392640114, + 0.0051448047161102295, + 0.004936582408845425, + 0.029628489166498184, + 0.005397838540375233, + 0.00424253148958087, + 0.008181631565093994, + 0.005163327790796757, + 0.014218037016689777, + 0.007174694444984198, + 0.008208787068724632, + 0.007202122826129198, + 0.013295642100274563, + 0.004580864682793617, + 0.005522863008081913, + 0.005403286777436733, + 0.006479586940258741, + 0.005652153864502907, + 0.00434355903416872, + 0.005026009865105152, + 0.00336380279622972, + 0.005093900486826897, + 0.005457811523228884, + 0.004900810308754444, + 0.004877848085016012, + 0.006274688057601452, + 0.0035459324717521667, + 0.004683146718889475, + 0.005704525858163834, + 0.02262159250676632, + 0.020260121673345566, + 0.021928399801254272, + 0.0255710668861866, + 0.04315662011504173, + 0.02657635137438774, + 0.029440980404615402, + 0.025155486539006233, + 0.007978374138474464, + 0.006729991640895605, + 0.010224716737866402, + 0.008939535357058048, + 0.017478303983807564, + 0.014142855070531368, + 0.010444920510053635, + 0.007683301344513893, + 0.005865312647074461, + 0.009781847707927227, + 0.00748805608600378, + 0.016306687146425247, + 0.015479808673262596, + 0.013340670615434647, + 0.012625772505998611, + 0.0148071413859725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.05498318374156952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058861404657364 + ], + "opposite_logprob_tokens": [ + -3.8955886363983154 + ], + "logprob_reward": -0.02058861404657364, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058861404657364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18722686171531677, + 0.10668329149484634, + 0.04066486656665802, + 0.03706565126776695, + 0.13194338977336884, + 0.020996470004320145, + 0.02339363843202591, + 0.045314811170101166, + 0.09627267718315125, + 0.310438334941864 + ], + "credit_map": [ + 0.03686372563242912, + 0.01858203299343586, + 0.024213137105107307, + 0.018664296716451645, + 0.008209136314690113, + 0.017324216663837433, + 0.013775618746876717, + 0.049594696611166, + 0.004060284234583378, + 0.004870103672146797, + 0.04508335515856743, + 0.013019057922065258, + 0.005319804884493351, + 0.0033445146400481462, + 0.023257002234458923, + 0.007729168515652418, + 0.0030549033544957638, + 0.004435300827026367, + 0.003986273892223835, + 0.008689775131642818, + 0.005806853994727135, + 0.0032063638791441917, + 0.005362019408494234, + 0.006123376544564962, + 0.004747397266328335, + 0.0065520452335476875, + 0.004791476298123598, + 0.005835122894495726, + 0.0030033604707568884, + 0.0032856562174856663, + 0.0066661108285188675, + 0.0021844797302037477, + 0.005687348544597626, + 0.010261006653308868, + 0.009797131642699242, + 0.038051679730415344, + 0.0063413153402507305, + 0.010324113070964813, + 0.01595744676887989, + 0.03552335128188133, + 0.003012614557519555, + 0.0023175489623099566, + 0.002714365255087614, + 0.002280572196468711, + 0.002525137737393379, + 0.0028760128188878298, + 0.0017290514660999179, + 0.003541167825460434, + 0.0022292882204055786, + 0.005142854060977697, + 0.0030840847175568342, + 0.0019352900562807918, + 0.003202318912371993, + 0.001995735103264451, + 0.0028176370542496443, + 0.0029864313546568155, + 0.006467390339821577, + 0.004312202800065279, + 0.005649270955473185, + 0.006646325346082449, + 0.006881986279040575, + 0.0054554445669054985, + 0.0053674569353461266, + 0.004534737206995487, + 0.011444290168583393, + 0.012107804417610168, + 0.01033428218215704, + 0.013844029046595097, + 0.013908511027693748, + 0.014517683535814285, + 0.01071056816726923, + 0.009405507706105709, + 0.01664101518690586, + 0.029698746278882027, + 0.0234312005341053, + 0.04323858395218849, + 0.06669644266366959, + 0.06271643191576004, + 0.0413791760802269, + 0.02663673460483551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.02058861404657364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot rapidly moves upward from a ruined asphalt road to reveal a massive mecha against the backdrop of the scarred Tokyo Tower at dusk. The mecha's design blends realism and fantasy, constructed from polished white and dark blue metal panels, now worn and charred. Pulsating cyan light emanates from its vents and radiates along its panel lines, casting an ethereal glow over the surrounding smoke. The animation emphasizes its immense weight; with ponderous movements, the mecha extends its massive feet, causing the ground to tremble and crack. Visible through a translucent holographic display, within the cockpit lies a 16-year-old boy with sharp, stylized anime features, spiky blue hair, and a resolute gaze, gripping the controls, his knuckles white. As he decisively presses the throttle forward, the mecha's thrusters roar and ignite, sending it hurtling forward, its fists a blur of speed, trailed by sharp lines of motion through the air. The punch connects with the unseen enemy, triggering a catastrophic explosion; the scene is instantly engulfed in the impact frame of blinding white light, followed by a violent outward blast of fire, smoke, and flying debris, illuminating the intense, high-stakes atmosphere of the Dusk Battle.", + "seed": 32695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_12/index_292_seed_32695_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": -0.03878686968237162, + "vlm_logprob": -0.03878686968237162, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.03878686968237162, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02140853926539421 + ], + "opposite_logprob_tokens": [ + -4.021408557891846 + ], + "logprob_reward": -0.02140853926539421, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02140853926539421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2188841700553894, + 0.32096555829048157, + 0.10237573087215424, + 0.09579922258853912, + 0.04324640706181526, + 0.049820754677057266, + 0.07414644211530685, + 0.02679518796503544, + 0.03302677720785141, + 0.034939710050821304 + ], + "credit_map": [ + 0.027507834136486053, + 0.026926828548312187, + 0.03756174445152283, + 0.029332177713513374, + 0.02091868408024311, + 0.018129218369722366, + 0.027245182543992996, + 0.03126250579953194, + 0.030273480340838432, + 0.04707096889615059, + 0.062326423823833466, + 0.045189838856458664, + 0.019428515806794167, + 0.024812322109937668, + 0.04235750064253807, + 0.04950650781393051, + 0.012500308454036713, + 0.011339143849909306, + 0.014795435592532158, + 0.014970473013818264, + 0.007870317436754704, + 0.015397178009152412, + 0.010859193280339241, + 0.014643684960901737, + 0.008123230189085007, + 0.010441282764077187, + 0.014386985450983047, + 0.02177032269537449, + 0.009658199734985828, + 0.01018458977341652, + 0.010162299498915672, + 0.0110723115503788, + 0.005252419039607048, + 0.005660271737724543, + 0.00573046924546361, + 0.0072891185991466045, + 0.00513332849368453, + 0.0049296217039227486, + 0.004848175682127476, + 0.004403002094477415, + 0.009681585244834423, + 0.005100369453430176, + 0.007489423267543316, + 0.004540137480944395, + 0.005227091256529093, + 0.005614234134554863, + 0.0035822237841784954, + 0.008585692383348942, + 0.009747871197760105, + 0.01601765677332878, + 0.008283271454274654, + 0.0070124599151313305, + 0.012451522052288055, + 0.008419394493103027, + 0.006434548180550337, + 0.005779723171144724, + 0.004347766749560833, + 0.0029745143838226795, + 0.0035477811470627785, + 0.002858277875930071, + 0.0032301656901836395, + 0.003945608157664537, + 0.002310647862032056, + 0.0035804258659482002, + 0.0029000379145145416, + 0.0033643231727182865, + 0.0036600003950297832, + 0.002988958964124322, + 0.005705989431589842, + 0.007537542842328548, + 0.0035017114132642746, + 0.0033682130742818117, + 0.003649258753284812, + 0.0042849392630159855, + 0.002944005187600851, + 0.0033696710597723722, + 0.008518428541719913, + 0.0033965311013162136, + 0.004775326699018478, + 0.00400155084207654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02140853926539421 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337347537279129 + ], + "opposite_logprob_tokens": [ + -3.4087347984313965 + ], + "logprob_reward": -0.0337347537279129, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337347537279129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1831590086221695, + 0.5637973546981812, + 0.07212252914905548, + 0.026110343635082245, + 0.025685103610157967, + 0.036236342042684555, + 0.01389430183917284, + 0.023698322474956512, + 0.024433821439743042, + 0.030862905085086823 + ], + "credit_map": [ + 0.030275501310825348, + 0.020305782556533813, + 0.025214243680238724, + 0.023289622738957405, + 0.01611284539103508, + 0.01766975037753582, + 0.02512066625058651, + 0.025170568376779556, + 0.04076381400227547, + 0.10986033082008362, + 0.06990174204111099, + 0.060607846826314926, + 0.05163205415010452, + 0.047109633684158325, + 0.07837400585412979, + 0.10554786771535873, + 0.006277597974985838, + 0.008084098808467388, + 0.016018427908420563, + 0.0119396410882473, + 0.008254487067461014, + 0.0066197034902870655, + 0.00916155707091093, + 0.00576701108366251, + 0.00452478788793087, + 0.002291611395776272, + 0.004145339131355286, + 0.005542905535548925, + 0.0017070381436496973, + 0.0026956447400152683, + 0.002065386390313506, + 0.0031376266852021217, + 0.005745369475334883, + 0.00306538469158113, + 0.0035472537856549025, + 0.002977046649903059, + 0.0021967662032693624, + 0.002778073772788048, + 0.0019190823659300804, + 0.003456124337390065, + 0.0067888349294662476, + 0.004392507951706648, + 0.005270988680422306, + 0.0026912279427051544, + 0.003498577978461981, + 0.004045176785439253, + 0.004041582345962524, + 0.0055074431002140045, + 0.0017548085888847709, + 0.002288570161908865, + 0.001620198367163539, + 0.0011796650942415, + 0.002211062703281641, + 0.0018668151460587978, + 0.0009388522594235837, + 0.0020343272481113672, + 0.0044044386595487595, + 0.0029930698219686747, + 0.0030463607981801033, + 0.0019206193974241614, + 0.0028564033564180136, + 0.00309961405582726, + 0.002743482356891036, + 0.0026343322824686766, + 0.002695307834073901, + 0.003913318272680044, + 0.002080861246213317, + 0.0017095927614718676, + 0.004370726179331541, + 0.005609877407550812, + 0.002013456542044878, + 0.0020406795665621758, + 0.0029425148386508226, + 0.003781162202358246, + 0.0025831384118646383, + 0.0027783256955444813, + 0.007540345191955566, + 0.0035185208544135094, + 0.0043114288710057735, + 0.003407466923817992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0337347537279129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011095491237938404 + ], + "opposite_logprob_tokens": [ + -4.5110955238342285 + ], + "logprob_reward": -0.011095491237938404, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011095491237938404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09762933105230331, + 0.5571067929267883, + 0.05126861482858658, + 0.1064811646938324, + 0.061641987413167953, + 0.024615056812763214, + 0.019018536433577538, + 0.03588547185063362, + 0.02675064653158188, + 0.01960240676999092 + ], + "credit_map": [ + 0.007897349074482918, + 0.011020495556294918, + 0.01678687520325184, + 0.015476370230317116, + 0.010883770883083344, + 0.012142675928771496, + 0.010045373812317848, + 0.013376411981880665, + 0.039717260748147964, + 0.10443060100078583, + 0.07229089736938477, + 0.06139416992664337, + 0.05123613402247429, + 0.04482240229845047, + 0.07816236466169357, + 0.10505293309688568, + 0.005118159111589193, + 0.005383388604968786, + 0.0072838319465518, + 0.00755865965038538, + 0.004516462329775095, + 0.006793586071580648, + 0.00792102050036192, + 0.00669350940734148, + 0.008906559087336063, + 0.014048391953110695, + 0.01753874495625496, + 0.021850144490599632, + 0.010982505045831203, + 0.00978757068514824, + 0.01072337944060564, + 0.012643871828913689, + 0.008000782690942287, + 0.007703718729317188, + 0.007332129869610071, + 0.006299145519733429, + 0.005455476697534323, + 0.010582898743450642, + 0.006125173531472683, + 0.01014266163110733, + 0.003905339865013957, + 0.0026727146469056606, + 0.003998615778982639, + 0.0026155984960496426, + 0.0024660772178322077, + 0.0029015413019806147, + 0.0018534057307988405, + 0.0042017619125545025, + 0.003105508629232645, + 0.0033701341599226, + 0.0020732965786010027, + 0.0017306295922026038, + 0.0022697937674820423, + 0.002589248586446047, + 0.0012386117596179247, + 0.002641313010826707, + 0.006641697604209185, + 0.0031083980575203896, + 0.0051584490574896336, + 0.0037153218872845173, + 0.0031462155748158693, + 0.006462248042225838, + 0.004645033739507198, + 0.0030081067234277725, + 0.002532590413466096, + 0.0031991302967071533, + 0.0026264654006808996, + 0.002338139107450843, + 0.0046237981878221035, + 0.0059884293004870415, + 0.002536483108997345, + 0.0029056111816316843, + 0.0019172501051798463, + 0.0018817076925188303, + 0.0017374810995534062, + 0.0019230650505051017, + 0.005711360368877649, + 0.0023698543664067984, + 0.002343496074900031, + 0.001718190498650074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.011095491237938404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903658598661423 + ], + "opposite_logprob_tokens": [ + -2.5790364742279053 + ], + "logprob_reward": -0.07903658598661423, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903658598661423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11626989394426346, + 0.21627278625965118, + 0.06360583752393723, + 0.18416470289230347, + 0.04226762801408768, + 0.030786117538809776, + 0.10168799757957458, + 0.1104368269443512, + 0.03521138057112694, + 0.09929685294628143 + ], + "credit_map": [ + 0.011628233827650547, + 0.013560235500335693, + 0.01783936098217964, + 0.01778125949203968, + 0.009851831942796707, + 0.014110292308032513, + 0.013706568628549576, + 0.017792105674743652, + 0.018777495250105858, + 0.03579242154955864, + 0.0371895506978035, + 0.0281887948513031, + 0.017955226823687553, + 0.016863469034433365, + 0.027180280536413193, + 0.034325551241636276, + 0.006474883761256933, + 0.007022837642580271, + 0.009771598502993584, + 0.008806937374174595, + 0.005157116334885359, + 0.00961651187390089, + 0.0093152467161417, + 0.007440708577632904, + 0.02334693633019924, + 0.02183922566473484, + 0.030261876061558723, + 0.029111668467521667, + 0.01911357417702675, + 0.019199535250663757, + 0.02108139544725418, + 0.020210497081279755, + 0.007972827181220055, + 0.0054892897605896, + 0.00527916569262743, + 0.004657705780118704, + 0.004630529787391424, + 0.004038474056869745, + 0.0028066588565707207, + 0.00739298015832901, + 0.006178676150739193, + 0.00339868082664907, + 0.0047613950446248055, + 0.0024836682714521885, + 0.0034435114357620478, + 0.003078361740335822, + 0.002923661842942238, + 0.0045181624591350555, + 0.011122604832053185, + 0.018648486584424973, + 0.011975167319178581, + 0.010964740067720413, + 0.012365641072392464, + 0.013699647039175034, + 0.008955797180533409, + 0.013955917209386826, + 0.014962869696319103, + 0.01274084858596325, + 0.01887817494571209, + 0.00994810089468956, + 0.012763544917106628, + 0.014158559031784534, + 0.013334655202925205, + 0.013650062493979931, + 0.0033821887336671352, + 0.005459477659314871, + 0.0038619462866336107, + 0.003781083505600691, + 0.005982575938105583, + 0.0061379531398415565, + 0.0033571103122085333, + 0.003249043133109808, + 0.008272971957921982, + 0.00917416624724865, + 0.007742521818727255, + 0.008232764899730682, + 0.027955126017332077, + 0.01303850393742323, + 0.017039814963936806, + 0.007840996608138084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.07903658598661423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865897819399834 + ], + "opposite_logprob_tokens": [ + -3.048659086227417 + ], + "logprob_reward": -0.04865897819399834, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865897819399834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13961005210876465, + 0.48449617624282837, + 0.10390403866767883, + 0.020452693104743958, + 0.0411255806684494, + 0.023310579359531403, + 0.046607691794633865, + 0.04486790671944618, + 0.06764409691095352, + 0.02798117883503437 + ], + "credit_map": [ + 0.039526667445898056, + 0.023778432980179787, + 0.017483387142419815, + 0.02022932469844818, + 0.006288937758654356, + 0.010737200267612934, + 0.011932711116969585, + 0.009633387438952923, + 0.03998236358165741, + 0.08590124547481537, + 0.08383158594369888, + 0.06316749006509781, + 0.040574152022600174, + 0.03576669469475746, + 0.06313071399927139, + 0.07214196771383286, + 0.009855329990386963, + 0.01145887840539217, + 0.016680674627423286, + 0.014260932803153992, + 0.011660510674118996, + 0.01592261530458927, + 0.010999198071658611, + 0.01306589599698782, + 0.0026659751310944557, + 0.002250828081741929, + 0.0020786544773727655, + 0.003172790165990591, + 0.0023386573884636164, + 0.003334513632580638, + 0.0019578810315579176, + 0.0026533924974501133, + 0.0033381839748471975, + 0.00456006545573473, + 0.004291226156055927, + 0.0051082465797662735, + 0.0042009372264146805, + 0.006877975072711706, + 0.003944243770092726, + 0.008804700337350368, + 0.0038162183482199907, + 0.0028143005911260843, + 0.00357893668115139, + 0.0017910492606461048, + 0.002368266461417079, + 0.0025245454162359238, + 0.0021140638273209333, + 0.004303198307752609, + 0.0077124605886638165, + 0.0066785854287445545, + 0.004960259888321161, + 0.004663607105612755, + 0.004241976421326399, + 0.006520576309412718, + 0.005306266248226166, + 0.006523957010358572, + 0.007976653054356575, + 0.005626704543828964, + 0.006340728607028723, + 0.0034352773800492287, + 0.006268886849284172, + 0.005176405888050795, + 0.005882646422833204, + 0.004160602111369371, + 0.004826715681701899, + 0.009968697093427181, + 0.005951170343905687, + 0.005689313169568777, + 0.012367824092507362, + 0.01315271481871605, + 0.008200578391551971, + 0.00748708751052618, + 0.003140383865684271, + 0.004440991673618555, + 0.002891089767217636, + 0.0030094741377979517, + 0.005296705290675163, + 0.002886677160859108, + 0.0033386480063199997, + 0.002977206837385893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04865897819399834 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot rapidly moves upward from a ruined asphalt road to reveal a massive mecha against the backdrop of the scarred Tokyo Tower at dusk. The mecha's design blends realism and fantasy, constructed from polished white and dark blue metal panels, now worn and charred. Pulsating cyan light emanates from its vents and radiates along its panel lines, casting an ethereal glow over the surrounding smoke. The animation emphasizes its immense weight; with ponderous movements, the mecha extends its massive feet, causing the ground to tremble and crack. Visible through a translucent holographic display, within the cockpit lies a 16-year-old boy with sharp, stylized anime features, spiky blue hair, and a resolute gaze, gripping the controls, his knuckles white. As he decisively presses the throttle forward, the mecha's thrusters roar and ignite, sending it hurtling forward, its fists a blur of speed, trailed by sharp lines of motion through the air. The punch connects with the unseen enemy, triggering a catastrophic explosion; the scene is instantly engulfed in the impact frame of blinding white light, followed by a violent outward blast of fire, smoke, and flying debris, illuminating the intense, high-stakes atmosphere of the Dusk Battle.", + "seed": 32696, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_12/index_292_seed_32696_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": -0.021474533714354038, + "vlm_logprob": -0.021474533714354038, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.021474533714354038, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028547298163175583 + ], + "opposite_logprob_tokens": [ + -3.6535472869873047 + ], + "logprob_reward": -0.028547298163175583, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028547298163175583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4796496033668518, + 0.097199946641922, + 0.021750804036855698, + 0.036375682801008224, + 0.1711590588092804, + 0.02717652916908264, + 0.012712450698018074, + 0.015260213986039162, + 0.10574699938297272, + 0.032968681305646896 + ], + "credit_map": [ + 0.02529003657400608, + 0.026207176968455315, + 0.17248357832431793, + 0.17247731983661652, + 0.01457869540899992, + 0.011911261826753616, + 0.034667883068323135, + 0.022033678367733955, + 0.01686341129243374, + 0.014088558964431286, + 0.020187661051750183, + 0.013999844901263714, + 0.007357622031122446, + 0.010005353949964046, + 0.008931966498494148, + 0.005765533074736595, + 0.0021878397092223167, + 0.002808856777846813, + 0.002546739997342229, + 0.0038084948901087046, + 0.0020779897458851337, + 0.0033771051093935966, + 0.0021936700213700533, + 0.0027501105796545744, + 0.0032260760199278593, + 0.004260356072336435, + 0.006656134966760874, + 0.004576592240482569, + 0.0041658529080450535, + 0.0031934010330587626, + 0.004938274156302214, + 0.005359000060707331, + 0.031214024871587753, + 0.015187550336122513, + 0.030950237065553665, + 0.01653979904949665, + 0.01656300388276577, + 0.0176657997071743, + 0.014500767923891544, + 0.028537878766655922, + 0.0029555403161793947, + 0.0044832658022642136, + 0.004094922449439764, + 0.0032027908600866795, + 0.0038513545878231525, + 0.002380545251071453, + 0.0023381710052490234, + 0.0038699437864124775, + 0.0013562326785176992, + 0.0017208671197295189, + 0.0018715608166530728, + 0.0011776405153796077, + 0.0013403970515355468, + 0.0016676337691023946, + 0.002202362520620227, + 0.001375756342895329, + 0.0017028586007654667, + 0.0023038717918097973, + 0.0019420190947130322, + 0.0011127651669085026, + 0.002388342749327421, + 0.0029458357021212578, + 0.0015509534860029817, + 0.0013135691406205297, + 0.004853858612477779, + 0.0100161824375391, + 0.008428333327174187, + 0.002712743356823921, + 0.003731999546289444, + 0.0381607748568058, + 0.03418814763426781, + 0.0036549712531268597, + 0.003314343746751547, + 0.004376728553324938, + 0.0029771472327411175, + 0.003282561432570219, + 0.0053200917318463326, + 0.005835370626300573, + 0.005331621505320072, + 0.00253081857226789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.028547298163175583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603211596608162 + ], + "opposite_logprob_tokens": [ + -3.8956031799316406 + ], + "logprob_reward": -0.020603211596608162, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603211596608162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3988964855670929, + 0.051865287125110626, + 0.05873563885688782, + 0.05857575312256813, + 0.26217371225357056, + 0.0271657295525074, + 0.020118027925491333, + 0.02390577830374241, + 0.038667697459459305, + 0.059895895421504974 + ], + "credit_map": [ + 0.02184702269732952, + 0.024457192048430443, + 0.14434242248535156, + 0.1375345140695572, + 0.012914000079035759, + 0.010650922544300556, + 0.02893245778977871, + 0.018217986449599266, + 0.00808174628764391, + 0.007616412825882435, + 0.009216292761266232, + 0.005911927204579115, + 0.0060910810716450214, + 0.007573653943836689, + 0.0041710431687533855, + 0.0032031305599957705, + 0.005855985917150974, + 0.0074806478805840015, + 0.005983778275549412, + 0.011258876882493496, + 0.007876395247876644, + 0.009418617002665997, + 0.005476604215800762, + 0.005384733434766531, + 0.005169090814888477, + 0.009489172138273716, + 0.012255617417395115, + 0.006274258252233267, + 0.0071561336517333984, + 0.0034711032640188932, + 0.006885631009936333, + 0.00787474773824215, + 0.053133729845285416, + 0.020413696765899658, + 0.04289739578962326, + 0.024803033098578453, + 0.030187558382749557, + 0.025316869840025902, + 0.022940093651413918, + 0.04248134419322014, + 0.0023444145917892456, + 0.005141983740031719, + 0.003259653691202402, + 0.0035463436506688595, + 0.0028646753635257483, + 0.0031886056531220675, + 0.003409549593925476, + 0.003410501405596733, + 0.0015616602031514049, + 0.0027256053872406483, + 0.0024307991843670607, + 0.0016430162359029055, + 0.002351324073970318, + 0.003234950127080083, + 0.0037054885178804398, + 0.002465184312313795, + 0.0031265313737094402, + 0.0028437406290322542, + 0.002929027657955885, + 0.0015831557102501392, + 0.003983647096902132, + 0.004604528192430735, + 0.002245631068944931, + 0.0025895158760249615, + 0.0038481818046420813, + 0.0032141332048922777, + 0.0023977423552423716, + 0.0020307039376348257, + 0.0019833336118608713, + 0.012357623316347599, + 0.01094304583966732, + 0.0018929342040792108, + 0.008356991223990917, + 0.007365297060459852, + 0.006123272702097893, + 0.004332851152867079, + 0.009708546102046967, + 0.012434625998139381, + 0.007109965197741985, + 0.004464350640773773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.020603211596608162 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421128399670124 + ], + "opposite_logprob_tokens": [ + -4.264211177825928 + ], + "logprob_reward": -0.01421128399670124, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421128399670124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1770617663860321, + 0.059411101043224335, + 0.08424868434667587, + 0.05439666286110878, + 0.10343319177627563, + 0.04950076714158058, + 0.035658642649650574, + 0.02610429748892784, + 0.3444516062736511, + 0.06573326885700226 + ], + "credit_map": [ + 0.015409519895911217, + 0.014684966765344143, + 0.047165531665086746, + 0.04402224346995354, + 0.011882662773132324, + 0.010194766335189342, + 0.01524317730218172, + 0.018458904698491096, + 0.010870820842683315, + 0.008723698556423187, + 0.008094013668596745, + 0.00782553106546402, + 0.005313759669661522, + 0.007733966689556837, + 0.004962124861776829, + 0.005887185223400593, + 0.007238408550620079, + 0.01248848531395197, + 0.00927963200956583, + 0.012056131847202778, + 0.008130286820232868, + 0.014295882545411587, + 0.0084359897300601, + 0.012323863804340363, + 0.006564011797308922, + 0.005549172405153513, + 0.011319875717163086, + 0.0060976785607635975, + 0.005968858953565359, + 0.0059377336874604225, + 0.006493984255939722, + 0.006465352606028318, + 0.020765772089362144, + 0.008387105539441109, + 0.0192671287804842, + 0.009478708729147911, + 0.009875965304672718, + 0.009755520150065422, + 0.00895966961979866, + 0.016943318769335747, + 0.003924335353076458, + 0.007636429741978645, + 0.004180770833045244, + 0.0073172422125935555, + 0.0045095281675457954, + 0.006410306319594383, + 0.007193600293248892, + 0.008328551426529884, + 0.003238043049350381, + 0.004725378472357988, + 0.004517942201346159, + 0.0042654466815292835, + 0.004547002725303173, + 0.003345532575622201, + 0.005668525118380785, + 0.005350771360099316, + 0.0027350434102118015, + 0.0031808242201805115, + 0.0032070043962448835, + 0.0018298507202416658, + 0.0049904282204806805, + 0.004602056462317705, + 0.002507257740944624, + 0.0030518330167979, + 0.020884497091174126, + 0.03176524117588997, + 0.0208184402436018, + 0.008599905297160149, + 0.013007394038140774, + 0.125092014670372, + 0.11177365481853485, + 0.012510462664067745, + 0.007489129900932312, + 0.008394205942749977, + 0.007332304958254099, + 0.007911697961390018, + 0.008409447968006134, + 0.01129092276096344, + 0.007836136035621166, + 0.007069421000778675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.01421128399670124 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981475181877613 + ], + "opposite_logprob_tokens": [ + -3.5298147201538086 + ], + "logprob_reward": -0.02981475181877613, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981475181877613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4131172001361847, + 0.06821902096271515, + 0.053776245564222336, + 0.04208312928676605, + 0.10607016831636429, + 0.033149927854537964, + 0.04332314431667328, + 0.11829429864883423, + 0.07995975762605667, + 0.04200714826583862 + ], + "credit_map": [ + 0.020297713577747345, + 0.02666454017162323, + 0.14272400736808777, + 0.14792628586292267, + 0.014043042436242104, + 0.01067916676402092, + 0.030237674713134766, + 0.020544743165373802, + 0.010835107415914536, + 0.013672162778675556, + 0.007189069874584675, + 0.00710817938670516, + 0.00651239138096571, + 0.007818627171218395, + 0.007733923848718405, + 0.007349555380642414, + 0.0037771055940538645, + 0.00823710486292839, + 0.0046918438747525215, + 0.01174324844032526, + 0.00677755381911993, + 0.009508208371698856, + 0.0042344979010522366, + 0.004806679207831621, + 0.0032952765468508005, + 0.004319528117775917, + 0.00619170069694519, + 0.005570333916693926, + 0.004098651465028524, + 0.0038464886602014303, + 0.005866974126547575, + 0.008894175291061401, + 0.022241933271288872, + 0.008149385452270508, + 0.018927864730358124, + 0.010264103300869465, + 0.00965830683708191, + 0.009137033484876156, + 0.009027140215039253, + 0.018664389848709106, + 0.0037163470406085253, + 0.0045119705609977245, + 0.005576256662607193, + 0.00472478661686182, + 0.0024472398217767477, + 0.003813870018348098, + 0.003440860426053405, + 0.004918592516332865, + 0.003588669700548053, + 0.005126575008034706, + 0.007033586967736483, + 0.0065433369018137455, + 0.004833606071770191, + 0.0067250775173306465, + 0.004748934879899025, + 0.004723353777080774, + 0.011166375130414963, + 0.01293928176164627, + 0.009512647055089474, + 0.005447813309729099, + 0.03039061836898327, + 0.026453880593180656, + 0.011089088395237923, + 0.011294592171907425, + 0.006141357589513063, + 0.006577544845640659, + 0.02112756296992302, + 0.004491649102419615, + 0.013175449334084988, + 0.005915833171457052, + 0.017313657328486443, + 0.005216696765273809, + 0.006903556641191244, + 0.005744343623518944, + 0.0042917681857943535, + 0.004117469768971205, + 0.0052259089425206184, + 0.006783577613532543, + 0.004761915188282728, + 0.004178605042397976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.02981475181877613 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a low-angle shot that rapidly moves upward from a ruined asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mecha visible against the backdrop of a scarred Tokyo Tower at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mecha feature polished white and dark blue metal panels that are worn and charred with pulsating cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a 16-year-old boy with spiky blue hair and sharp anime features visible inside the translucent holographic cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mecha extend its massive feet to cause the ground to tremble and crack before punching forward and triggering a catastrophic explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196122996509075 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196122996509075, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196122996509075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39300599694252014, + 0.0450989305973053, + 0.024457687512040138, + 0.035731665790081024, + 0.08182519674301147, + 0.018256112933158875, + 0.020530426874756813, + 0.027011556550860405, + 0.20659387111663818, + 0.14748850464820862 + ], + "credit_map": [ + 0.016908837482333183, + 0.01610470563173294, + 0.15408362448215485, + 0.15279680490493774, + 0.006594911217689514, + 0.006770635023713112, + 0.027594393119215965, + 0.012152141891419888, + 0.010548930615186691, + 0.006593294441699982, + 0.0087861567735672, + 0.004437871277332306, + 0.004300601314753294, + 0.00555223086848855, + 0.002534331288188696, + 0.0023455151822417974, + 0.002289484953507781, + 0.0029053394682705402, + 0.0031739608384668827, + 0.0037060624454170465, + 0.002554226666688919, + 0.00378632009960711, + 0.002522716298699379, + 0.003519579069688916, + 0.0030384554993361235, + 0.00325217773206532, + 0.008089929819107056, + 0.0036272932775318623, + 0.003908353392034769, + 0.004233244340866804, + 0.004114892799407244, + 0.005467323120683432, + 0.017337094992399216, + 0.006242999341338873, + 0.014666683971881866, + 0.007929067127406597, + 0.007462932262569666, + 0.007735745515674353, + 0.00705929147079587, + 0.013391389511525631, + 0.0014684504130855203, + 0.0025020658504217863, + 0.0022564835380762815, + 0.0026246686466038227, + 0.002021603751927614, + 0.0022438077721744776, + 0.0017971371999010444, + 0.0033418976236134768, + 0.002002004999667406, + 0.002802504226565361, + 0.0031417636200785637, + 0.002253794576972723, + 0.0022916945163160563, + 0.0031037037260830402, + 0.003105986164882779, + 0.0018289765575900674, + 0.0029339008033275604, + 0.003629429033026099, + 0.0032149041071534157, + 0.0021077918354421854, + 0.005072502419352531, + 0.004436524119228125, + 0.002466388512402773, + 0.0031501171179115772, + 0.01647426001727581, + 0.0201801136136055, + 0.010161149315536022, + 0.008312213234603405, + 0.010015218518674374, + 0.06542891263961792, + 0.06731763482093811, + 0.008704389445483685, + 0.014832360669970512, + 0.02082010544836521, + 0.013213500380516052, + 0.013643640093505383, + 0.022779226303100586, + 0.028562432155013084, + 0.021552741527557373, + 0.012084513902664185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.014196122996509075 + } + ], + "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/0000128/group_13/index_tensor([9519, 9519])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_13/index_tensor([9519, 9519])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e1aaddfb168256e46380682d84babf433b2feaa2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_13/index_tensor([9519, 9519])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, overhead tracking shot follows a massive icebreaker as it majestically navigates the vast expanse of Arctic ice. Painted a striking, weathered red, its sturdy steel prow bears the scratches and rust of past voyages, it pushes forward with an unstoppable, steady, and powerful rhythm. Under the vibrant light of Makoto Shinkai's signature aesthetic—a low-hanging sun paints the sky a soft hue of pink, purple, and golden orange—each time the prow collides with the meters-thick ice, the world seems to freeze. As the ice groans and cracks with a bang, countless ice crystals are instantly launched into a spectacular cloud of particles. Under the extreme slow-motion effects, these ice crystals are no longer chaotic fragments, but dance and spin gracefully in the air. Each tiny crystal accurately captures and refracts the dreamy light in the sky, like billions of suspended diamonds, shining with a rainbow spectrum, and drawing out slender and beautiful halo stripes in the lens, combining the tremendous power of machinery with the instantaneous beauty of nature into a visual picture that is both magnificent and full of poetry and emotion.", + "seed": 25437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_13/index_9519_seed_25437_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": -0.02943226331844926, + "vlm_logprob": -0.02943226331844926, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.02943226331844926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012688618153333664 + ], + "opposite_logprob_tokens": [ + -5.012688636779785 + ], + "logprob_reward": -0.012688618153333664, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012688618153333664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08205656707286835, + 0.050562784075737, + 0.034730441868305206, + 0.05987968295812607, + 0.038942039012908936, + 0.053247347474098206, + 0.11898903548717499, + 0.22655294835567474, + 0.17272049188613892, + 0.16231869161128998 + ], + "credit_map": [ + 0.011378089897334576, + 0.01204534899443388, + 0.011438114568591118, + 0.009435119107365608, + 0.008300227113068104, + 0.011261402629315853, + 0.01151820458471775, + 0.006680058781057596, + 0.006952811032533646, + 0.004900186788290739, + 0.005791719537228346, + 0.005123266950249672, + 0.006931760348379612, + 0.009107925929129124, + 0.006623593624681234, + 0.005131518002599478, + 0.004877427127212286, + 0.003353237872943282, + 0.004857214633375406, + 0.00598360737785697, + 0.00323552661575377, + 0.003601636504754424, + 0.004929861519485712, + 0.003891930915415287, + 0.009476950392127037, + 0.0069385371170938015, + 0.008254383690655231, + 0.010303549468517303, + 0.005484430585056543, + 0.005606312770396471, + 0.007608832325786352, + 0.006206688936799765, + 0.004456200171262026, + 0.004493307322263718, + 0.005381803028285503, + 0.007168757263571024, + 0.004465671256184578, + 0.003373576793819666, + 0.005820340942591429, + 0.0037823831662535667, + 0.005361809395253658, + 0.006457052659243345, + 0.0072930799797177315, + 0.008428600616753101, + 0.004894884303212166, + 0.005061328411102295, + 0.009085750207304955, + 0.00666483910754323, + 0.008616308681666851, + 0.009357225149869919, + 0.018376924097537994, + 0.014383193105459213, + 0.011213758029043674, + 0.009598864242434502, + 0.0178389735519886, + 0.029603786766529083, + 0.026272408664226532, + 0.023703934624791145, + 0.03155259042978287, + 0.0313890241086483, + 0.02152358740568161, + 0.0228973887860775, + 0.03442559391260147, + 0.03478841483592987, + 0.01796669140458107, + 0.025228101760149002, + 0.023074770346283913, + 0.0313679538667202, + 0.023122817277908325, + 0.01765117421746254, + 0.01860992982983589, + 0.015699056908488274, + 0.013658461160957813, + 0.020248964428901672, + 0.017831405624747276, + 0.02340885065495968, + 0.03212440013885498, + 0.018139326944947243, + 0.020864883437752724, + 0.016042400151491165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.012688618153333664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904993534088135 + ], + "opposite_logprob_tokens": [ + -2.464049816131592 + ], + "logprob_reward": -0.08904993534088135, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904993534088135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08694044500589371, + 0.09902025759220123, + 0.043389324098825455, + 0.07830090820789337, + 0.040932733565568924, + 0.07613278925418854, + 0.1766016185283661, + 0.14840731024742126, + 0.14908017218112946, + 0.10119449347257614 + ], + "credit_map": [ + 0.01873082108795643, + 0.009668772108852863, + 0.015638338401913643, + 0.007644869852811098, + 0.008379255421459675, + 0.0081973597407341, + 0.009730963036417961, + 0.00895006861537695, + 0.013374133966863155, + 0.007957365363836288, + 0.013137469068169594, + 0.013568945229053497, + 0.008905618451535702, + 0.014305146411061287, + 0.014145158231258392, + 0.013626424595713615, + 0.0057965293526649475, + 0.005413098260760307, + 0.006149783730506897, + 0.008176136761903763, + 0.004469242412596941, + 0.003326965495944023, + 0.005924271885305643, + 0.004133292939513922, + 0.012415855191648006, + 0.005066533572971821, + 0.012264239601790905, + 0.0088304178789258, + 0.0057174754329025745, + 0.009271804243326187, + 0.007380200549960136, + 0.017354382202029228, + 0.003922687843441963, + 0.004713816102594137, + 0.005195631179958582, + 0.006166000384837389, + 0.005192628595978022, + 0.003285794984549284, + 0.007345862220972776, + 0.005110311787575483, + 0.007550465874373913, + 0.007688713725656271, + 0.011349727399647236, + 0.013641398400068283, + 0.006197565235197544, + 0.006556973326951265, + 0.012790764681994915, + 0.010357183404266834, + 0.01296302955597639, + 0.018086690455675125, + 0.024449767544865608, + 0.02803400531411171, + 0.027850355952978134, + 0.018594462424516678, + 0.0199413001537323, + 0.026681998744606972, + 0.015523992478847504, + 0.015880640596151352, + 0.020259395241737366, + 0.019508473575115204, + 0.01585061475634575, + 0.014720020815730095, + 0.022979866713285446, + 0.023684298619627953, + 0.015317731536924839, + 0.02061392553150654, + 0.01626839116215706, + 0.017677508294582367, + 0.020593276247382164, + 0.020631885156035423, + 0.020500613376498222, + 0.017476852983236313, + 0.008074583485722542, + 0.014067346230149269, + 0.008937605656683445, + 0.008388465270400047, + 0.022672707214951515, + 0.019530490040779114, + 0.008899372071027756, + 0.010623916983604431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.08904993534088135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033702824264764786 + ], + "opposite_logprob_tokens": [ + -3.408702850341797 + ], + "logprob_reward": -0.033702824264764786, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033702824264764786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04530582204461098, + 0.019056696444749832, + 0.027633484452962875, + 0.042133718729019165, + 0.03827102854847908, + 0.05151147022843361, + 0.151320680975914, + 0.22392787039279938, + 0.3454156219959259, + 0.05542360246181488 + ], + "credit_map": [ + 0.007269762456417084, + 0.004151001572608948, + 0.007330779917538166, + 0.005451015662401915, + 0.0036640088073909283, + 0.004656871780753136, + 0.008211050182580948, + 0.004571332596242428, + 0.002358743455260992, + 0.0021973277907818556, + 0.002811166225001216, + 0.0024302636738866568, + 0.0019778755959123373, + 0.00220495811663568, + 0.0024775133933871984, + 0.002598847495391965, + 0.0031543844379484653, + 0.002876484766602516, + 0.003364633535966277, + 0.0037181915249675512, + 0.0036955936811864376, + 0.0030482832808047533, + 0.004721744917333126, + 0.003054168075323105, + 0.006363338325172663, + 0.004731963388621807, + 0.005414127837866545, + 0.005098233465105295, + 0.0038278496358543634, + 0.004196109250187874, + 0.00460775475949049, + 0.007894342765212059, + 0.0032326013315469027, + 0.004004262387752533, + 0.00518457405269146, + 0.00528330123052001, + 0.0038978629745543003, + 0.0045105163007974625, + 0.006847941782325506, + 0.005309969186782837, + 0.0037493305280804634, + 0.005027319770306349, + 0.007615182548761368, + 0.009468838572502136, + 0.004463565535843372, + 0.005999093409627676, + 0.007695166394114494, + 0.007492973934859037, + 0.009153216145932674, + 0.013169714249670506, + 0.022872567176818848, + 0.017483990639448166, + 0.020855233073234558, + 0.010756285861134529, + 0.023204728960990906, + 0.033824943006038666, + 0.024633483961224556, + 0.022604087367653847, + 0.031353920698165894, + 0.030094824731349945, + 0.02262386493384838, + 0.021564217284321785, + 0.03464437648653984, + 0.036409102380275726, + 0.030261358246207237, + 0.05293034389615059, + 0.043936144560575485, + 0.08058101683855057, + 0.047408003360033035, + 0.023842142894864082, + 0.0328851081430912, + 0.03357148915529251, + 0.006175821181386709, + 0.006083549465984106, + 0.00838406290858984, + 0.007891686633229256, + 0.008129200898110867, + 0.00605010287836194, + 0.007572521455585957, + 0.005136657040566206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.033702824264764786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628949359059334 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.007628949359059334, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628949359059334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07087896019220352, + 0.038965728133916855, + 0.03102492168545723, + 0.03909965232014656, + 0.04329507425427437, + 0.05481541529297829, + 0.14517851173877716, + 0.2120542675256729, + 0.2380058765411377, + 0.12668156623840332 + ], + "credit_map": [ + 0.014767110347747803, + 0.009901290759444237, + 0.014322238974273205, + 0.00820806622505188, + 0.006709631532430649, + 0.00695446040481329, + 0.005767828784883022, + 0.004248330369591713, + 0.006535927765071392, + 0.00470316968858242, + 0.005033356603235006, + 0.004810663405805826, + 0.0048218839801847935, + 0.005946493707597256, + 0.0040032207034528255, + 0.003111009020358324, + 0.004854760132730007, + 0.002997957868501544, + 0.004488999955356121, + 0.005997327622026205, + 0.0034638000652194023, + 0.0028581616934388876, + 0.003685085102915764, + 0.0026788306422531605, + 0.0057693119160830975, + 0.0039397841319441795, + 0.004740848671644926, + 0.004533083643764257, + 0.004362395498901606, + 0.004272841848433018, + 0.005827887915074825, + 0.005653499625623226, + 0.006344708614051342, + 0.004082228988409042, + 0.006692394148558378, + 0.0068733831867575645, + 0.0033496953547000885, + 0.003573362249881029, + 0.005893517751246691, + 0.006485781632363796, + 0.003816381562501192, + 0.004722216632217169, + 0.007823402062058449, + 0.006550411228090525, + 0.006666412577033043, + 0.00821126438677311, + 0.009488437324762344, + 0.007536889053881168, + 0.014640319161117077, + 0.012700092978775501, + 0.02005671337246895, + 0.02308037132024765, + 0.015304487198591232, + 0.016989657655358315, + 0.0204890388995409, + 0.021917829290032387, + 0.018996866419911385, + 0.021553557366132736, + 0.026462744921445847, + 0.027516711503267288, + 0.01765458658337593, + 0.01828155480325222, + 0.034692566841840744, + 0.046895693987607956, + 0.022915832698345184, + 0.031023217365145683, + 0.035111311823129654, + 0.05307218059897423, + 0.025499438866972923, + 0.019528839737176895, + 0.025972923263907433, + 0.024882130324840546, + 0.011323227547109127, + 0.018025998026132584, + 0.014983410015702248, + 0.014948952943086624, + 0.020928241312503815, + 0.018769605085253716, + 0.012264206074178219, + 0.015437922440469265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.007628949359059334 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090989474207163 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.004090989474207163, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090989474207163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2306385040283203, + 0.11896912753582001, + 0.045643363147974014, + 0.02482115663588047, + 0.03495835140347481, + 0.045673567801713943, + 0.060589492321014404, + 0.1458321511745453, + 0.2007129192352295, + 0.09216140955686569 + ], + "credit_map": [ + 0.05281025171279907, + 0.048199087381362915, + 0.03486073762178421, + 0.024783747270703316, + 0.01601501926779747, + 0.01701056957244873, + 0.015656519681215286, + 0.02130257897078991, + 0.026589680463075638, + 0.01321327406913042, + 0.01985115557909012, + 0.019640540704131126, + 0.007614344824105501, + 0.008199156261980534, + 0.009555960074067116, + 0.014305016957223415, + 0.007500173058360815, + 0.004056414123624563, + 0.006140720099210739, + 0.007445552386343479, + 0.005922142416238785, + 0.005436677020043135, + 0.00480208033695817, + 0.004339602775871754, + 0.0031672543846070766, + 0.0037118245381861925, + 0.003464440582320094, + 0.004178618546575308, + 0.0028394118417054415, + 0.002061864361166954, + 0.003098537912592292, + 0.00229920563288033, + 0.002796600805595517, + 0.004002273082733154, + 0.0054229446686804295, + 0.008603655733168125, + 0.00273819244466722, + 0.0026792953722178936, + 0.0038694748654961586, + 0.004845916293561459, + 0.004174282308667898, + 0.0038357714656740427, + 0.007269757334142923, + 0.009759558364748955, + 0.003796517848968506, + 0.0038494097534567118, + 0.005772175267338753, + 0.007216096390038729, + 0.005313475150614977, + 0.006001171190291643, + 0.009587671607732773, + 0.007781775202602148, + 0.007969492115080357, + 0.005932102911174297, + 0.010883183218538761, + 0.007120621856302023, + 0.019233744591474533, + 0.016747241839766502, + 0.02091294899582863, + 0.019788729026913643, + 0.013425411656498909, + 0.013040118850767612, + 0.021846197545528412, + 0.020837754011154175, + 0.02012697421014309, + 0.030945084989070892, + 0.022348769009113312, + 0.03195634484291077, + 0.029614252969622612, + 0.0218014158308506, + 0.02366865612566471, + 0.020251421257853508, + 0.010788677260279655, + 0.012944774702191353, + 0.011760636232793331, + 0.010695217177271843, + 0.017101101577281952, + 0.01130056381225586, + 0.00948475394397974, + 0.008085686713457108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.004090989474207163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead tracking shot follows a massive icebreaker as it majestically navigates the vast expanse of Arctic ice. Painted a striking, weathered red, its sturdy steel prow bears the scratches and rust of past voyages, it pushes forward with an unstoppable, steady, and powerful rhythm. Under the vibrant light of Makoto Shinkai's signature aesthetic—a low-hanging sun paints the sky a soft hue of pink, purple, and golden orange—each time the prow collides with the meters-thick ice, the world seems to freeze. As the ice groans and cracks with a bang, countless ice crystals are instantly launched into a spectacular cloud of particles. Under the extreme slow-motion effects, these ice crystals are no longer chaotic fragments, but dance and spin gracefully in the air. Each tiny crystal accurately captures and refracts the dreamy light in the sky, like billions of suspended diamonds, shining with a rainbow spectrum, and drawing out slender and beautiful halo stripes in the lens, combining the tremendous power of machinery with the instantaneous beauty of nature into a visual picture that is both magnificent and full of poetry and emotion.", + "seed": 25438, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_13/index_9519_seed_25438_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": -0.023804173059761524, + "vlm_logprob": -0.023804173059761524, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.023804173059761524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013579140417277813 + ], + "opposite_logprob_tokens": [ + -4.888579368591309 + ], + "logprob_reward": -0.013579140417277813, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.013579140417277813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16176874935626984, + 0.05055266618728638, + 0.11015757918357849, + 0.07237138599157333, + 0.050881996750831604, + 0.06790459901094437, + 0.17933382093906403, + 0.04974015802145004, + 0.0704532042145729, + 0.18683584034442902 + ], + "credit_map": [ + 0.02091120183467865, + 0.029788227751851082, + 0.017323002219200134, + 0.02211775630712509, + 0.021893182769417763, + 0.016624266281723976, + 0.014827979728579521, + 0.018283123150467873, + 0.009396924637258053, + 0.008259420283138752, + 0.004724778700619936, + 0.006248001009225845, + 0.006015062797814608, + 0.006881853099912405, + 0.00476128188893199, + 0.004265344236046076, + 0.00842475239187479, + 0.007019866723567247, + 0.029299726709723473, + 0.006587218958884478, + 0.0075426786206662655, + 0.009747801348567009, + 0.007370150648057461, + 0.034165382385253906, + 0.008025255985558033, + 0.008116061799228191, + 0.015484688803553581, + 0.006778761278837919, + 0.007980706170201302, + 0.008638281375169754, + 0.008975359611213207, + 0.00837226677685976, + 0.012167951092123985, + 0.0052923220209777355, + 0.01188811007887125, + 0.005090184975415468, + 0.003985698334872723, + 0.006044803652912378, + 0.0036144263576716185, + 0.0027985002379864454, + 0.004916822537779808, + 0.008611643686890602, + 0.009262127801775932, + 0.0048691886477172375, + 0.012566144578158855, + 0.012435687705874443, + 0.007497268728911877, + 0.007745714392513037, + 0.024744728580117226, + 0.02995874360203743, + 0.015464114025235176, + 0.024122659116983414, + 0.026969825848937035, + 0.015650548040866852, + 0.03026353195309639, + 0.012159682810306549, + 0.00770984124392271, + 0.006117567420005798, + 0.006601717323064804, + 0.006020990666002035, + 0.006230633240193129, + 0.004623593762516975, + 0.0055394661612808704, + 0.006896349135786295, + 0.007039511110633612, + 0.008886933326721191, + 0.0075602103024721146, + 0.008156916126608849, + 0.010590535588562489, + 0.007634221110492945, + 0.010152214206755161, + 0.01043266337364912, + 0.021794414147734642, + 0.019938942044973373, + 0.02068837732076645, + 0.01627936214208603, + 0.0399644635617733, + 0.02870366908609867, + 0.02309669926762581, + 0.01636992208659649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.013579140417277813 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005760818719864 + ], + "opposite_logprob_tokens": [ + -2.6950576305389404 + ], + "logprob_reward": -0.07005760818719864, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005760818719864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12529754638671875, + 0.060417477041482925, + 0.07610821723937988, + 0.32295623421669006, + 0.045936450362205505, + 0.06494449079036713, + 0.06260855495929718, + 0.041424598544836044, + 0.10292436182498932, + 0.097382090985775 + ], + "credit_map": [ + 0.012980909086763859, + 0.025377199053764343, + 0.011042979545891285, + 0.01410352997481823, + 0.019111398607492447, + 0.015258912928402424, + 0.013902955688536167, + 0.013519643805921078, + 0.009261542931199074, + 0.007274312898516655, + 0.0077261957339942455, + 0.010353144258260727, + 0.005185382440686226, + 0.004886043723672628, + 0.00719034718349576, + 0.008540498092770576, + 0.006893343757838011, + 0.008353464305400848, + 0.015061036683619022, + 0.007996163330972195, + 0.0064347838051617146, + 0.008162193931639194, + 0.007021226454526186, + 0.016186000779271126, + 0.029319914057850838, + 0.01604783721268177, + 0.1074446439743042, + 0.018165262416005135, + 0.014679529704153538, + 0.01979587972164154, + 0.012722636573016644, + 0.10478050261735916, + 0.005990877281874418, + 0.006320038344711065, + 0.006278656423091888, + 0.003758776467293501, + 0.00598708214238286, + 0.009564662352204323, + 0.004133212845772505, + 0.0039031407795846462, + 0.006964215077459812, + 0.006786832585930824, + 0.01190466433763504, + 0.006788812577724457, + 0.012391121126711369, + 0.004591927397996187, + 0.009174030274152756, + 0.006342878565192223, + 0.010411478579044342, + 0.007667224388569593, + 0.005262637510895729, + 0.008336873725056648, + 0.007478937041014433, + 0.005205872934311628, + 0.009347074665129185, + 0.008898451924324036, + 0.006434905342757702, + 0.00530895683914423, + 0.004732798784971237, + 0.004480368923395872, + 0.005027080420404673, + 0.004786835517734289, + 0.004084383137524128, + 0.006569268181920052, + 0.010804926976561546, + 0.022740118205547333, + 0.009984331205487251, + 0.0115513252094388, + 0.016188938170671463, + 0.00852716900408268, + 0.01515673566609621, + 0.007970808073878288, + 0.01533019170165062, + 0.01282690279185772, + 0.011233832687139511, + 0.012974143959581852, + 0.014094706624746323, + 0.010568457655608654, + 0.009749894961714745, + 0.0106039484962821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.07005760818719864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207170344889164 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207170344889164, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207170344889164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11799804121255875, + 0.04301034286618233, + 0.02916298434138298, + 0.12137697637081146, + 0.05208303779363632, + 0.035388555377721786, + 0.09611837565898895, + 0.07532802224159241, + 0.3032468557357788, + 0.12628675997257233 + ], + "credit_map": [ + 0.016019156202673912, + 0.018837332725524902, + 0.014549799263477325, + 0.013277541846036911, + 0.012534012086689472, + 0.014199006371200085, + 0.012303507886826992, + 0.01627768948674202, + 0.007131094578653574, + 0.007647004909813404, + 0.004494490567594767, + 0.005389849655330181, + 0.004898771643638611, + 0.005257322918623686, + 0.004361879080533981, + 0.0038299330044537783, + 0.0032723122276365757, + 0.0030094392132014036, + 0.004438691306859255, + 0.0038948471192270517, + 0.0034446611534804106, + 0.0038895453326404095, + 0.003636315232142806, + 0.0035771748516708612, + 0.010277986526489258, + 0.007506133988499641, + 0.034726571291685104, + 0.006854542531073093, + 0.008012140169739723, + 0.00800586398690939, + 0.00698278471827507, + 0.03901095688343048, + 0.013680475763976574, + 0.005345265381038189, + 0.01183683518320322, + 0.005346756428480148, + 0.003501197788864374, + 0.0030203796923160553, + 0.005756082013249397, + 0.003596049267798662, + 0.003246049163863063, + 0.003989622462540865, + 0.003394906409084797, + 0.00316095189191401, + 0.007293334696441889, + 0.004957843571901321, + 0.004065162036567926, + 0.005280685145407915, + 0.011745492927730083, + 0.011403623037040234, + 0.010899358429014683, + 0.011040519922971725, + 0.013052885420620441, + 0.007632070686668158, + 0.013590035028755665, + 0.01675439067184925, + 0.009534093551337719, + 0.011999485082924366, + 0.005153551697731018, + 0.012598470784723759, + 0.011015319265425205, + 0.00726444972679019, + 0.012933631427586079, + 0.004829025361686945, + 0.03378754109144211, + 0.04133835434913635, + 0.04679933190345764, + 0.028559008613228798, + 0.03944563865661621, + 0.030632376670837402, + 0.04901827499270439, + 0.03366631641983986, + 0.01711178757250309, + 0.01320794876664877, + 0.010631556622684002, + 0.015777571126818657, + 0.02286461740732193, + 0.016923200339078903, + 0.017439251765608788, + 0.012330831959843636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.014207170344889164 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643358945846558 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643358945846558, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643358945846558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07131801545619965, + 0.020340485498309135, + 0.05727279931306839, + 0.1405433714389801, + 0.04629676789045334, + 0.027235684916377068, + 0.1438097208738327, + 0.07836569100618362, + 0.12185534834861755, + 0.29296213388442993 + ], + "credit_map": [ + 0.009306480176746845, + 0.014652736485004425, + 0.006765523925423622, + 0.00984262116253376, + 0.007924152538180351, + 0.009058138355612755, + 0.0076185367070138454, + 0.006149827037006617, + 0.0026978622190654278, + 0.004359545651823282, + 0.0018690862925723195, + 0.002480606082826853, + 0.0024013123475015163, + 0.002300196560099721, + 0.00268571893684566, + 0.0015461549628525972, + 0.004438743926584721, + 0.005831591784954071, + 0.009902688674628735, + 0.006894129328429699, + 0.0055559128522872925, + 0.006291442550718784, + 0.005875148810446262, + 0.012483134865760803, + 0.012453878298401833, + 0.008302094414830208, + 0.03750699758529663, + 0.009698094800114632, + 0.007798921782523394, + 0.009557328186929226, + 0.009788925759494305, + 0.04543713107705116, + 0.010076177306473255, + 0.005260393489152193, + 0.00825008749961853, + 0.003732160897925496, + 0.0064981477335095406, + 0.005869096145033836, + 0.003789963200688362, + 0.0028207353316247463, + 0.0026442166417837143, + 0.0033376466017216444, + 0.0021638930775225163, + 0.0023851271253079176, + 0.004989210516214371, + 0.004936811979860067, + 0.0029263729229569435, + 0.0038524058181792498, + 0.017739197239279747, + 0.017197439447045326, + 0.013096128590404987, + 0.019803034141659737, + 0.016689961776137352, + 0.009819637052714825, + 0.022368913516402245, + 0.027095390483736992, + 0.011281593702733517, + 0.01011163741350174, + 0.008112715557217598, + 0.009160386398434639, + 0.009337038733065128, + 0.007817596197128296, + 0.013192225247621536, + 0.009352493099868298, + 0.011231866665184498, + 0.014949443750083447, + 0.015190528705716133, + 0.014948145486414433, + 0.014953703619539738, + 0.013128952123224735, + 0.0228411965072155, + 0.014611500315368176, + 0.030285071581602097, + 0.033747877925634384, + 0.029327621683478355, + 0.02591816335916519, + 0.0679028183221817, + 0.04768224060535431, + 0.03310944139957428, + 0.024988876655697823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.008643358945846558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533587403595448 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533587403595448, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533587403595448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3006099760532379, + 0.11326362937688828, + 0.14561058580875397, + 0.03881870210170746, + 0.046811752021312714, + 0.055030882358551025, + 0.0718073919415474, + 0.06026748567819595, + 0.08957572281360626, + 0.07820385694503784 + ], + "credit_map": [ + 0.05162961408495903, + 0.037184108048677444, + 0.036430105566978455, + 0.032418571412563324, + 0.027047693729400635, + 0.04248763248324394, + 0.027754027396440506, + 0.04565819352865219, + 0.01918019913136959, + 0.022899476811289787, + 0.01051518227905035, + 0.013115201145410538, + 0.01464820932596922, + 0.015265889465808868, + 0.006588080432265997, + 0.011051391251385212, + 0.016680074855685234, + 0.014203399419784546, + 0.03584274649620056, + 0.011872698552906513, + 0.010412867181003094, + 0.012175223790109158, + 0.012325918301939964, + 0.03209765627980232, + 0.003996026702225208, + 0.0051054032519459724, + 0.004365424159914255, + 0.003574240254238248, + 0.0039325254037976265, + 0.005057862494140863, + 0.0058179800398647785, + 0.0069692437537014484, + 0.010923277586698532, + 0.004547102376818657, + 0.009739365428686142, + 0.0028475376311689615, + 0.005017013754695654, + 0.005383307114243507, + 0.00538727967068553, + 0.0029668703209608793, + 0.0044032493606209755, + 0.006525740493088961, + 0.004365573171526194, + 0.0038410136476159096, + 0.012098164297640324, + 0.009318395517766476, + 0.006018460728228092, + 0.00846028421074152, + 0.010184520855545998, + 0.012136041186749935, + 0.005517343059182167, + 0.010209386236965656, + 0.010710053145885468, + 0.00631298590451479, + 0.011297410354018211, + 0.005439647473394871, + 0.006474830210208893, + 0.006064547225832939, + 0.00491123553365469, + 0.005358878523111343, + 0.010158088058233261, + 0.007951371371746063, + 0.0063745612278580666, + 0.012973977252840996, + 0.008019520901143551, + 0.010742363519966602, + 0.011088753119111061, + 0.009721360169351101, + 0.012863672338426113, + 0.01129396352916956, + 0.017292629927396774, + 0.008553460240364075, + 0.009100939147174358, + 0.008226072415709496, + 0.009399150498211384, + 0.011739784851670265, + 0.012115871533751488, + 0.010385039262473583, + 0.010374109260737896, + 0.006862894631922245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012533587403595448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead tracking shot follows a massive icebreaker as it majestically navigates the vast expanse of Arctic ice. Painted a striking, weathered red, its sturdy steel prow bears the scratches and rust of past voyages, it pushes forward with an unstoppable, steady, and powerful rhythm. Under the vibrant light of Makoto Shinkai's signature aesthetic—a low-hanging sun paints the sky a soft hue of pink, purple, and golden orange—each time the prow collides with the meters-thick ice, the world seems to freeze. As the ice groans and cracks with a bang, countless ice crystals are instantly launched into a spectacular cloud of particles. Under the extreme slow-motion effects, these ice crystals are no longer chaotic fragments, but dance and spin gracefully in the air. Each tiny crystal accurately captures and refracts the dreamy light in the sky, like billions of suspended diamonds, shining with a rainbow spectrum, and drawing out slender and beautiful halo stripes in the lens, combining the tremendous power of machinery with the instantaneous beauty of nature into a visual picture that is both magnificent and full of poetry and emotion.", + "seed": 25439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_13/index_9519_seed_25439_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": -0.04403520300984383, + "vlm_logprob": -0.04403520300984383, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04403520300984383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01727217249572277 + ], + "opposite_logprob_tokens": [ + -4.767271995544434 + ], + "logprob_reward": -0.01727217249572277, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01727217249572277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08420466631650925, + 0.30403292179107666, + 0.05258307605981827, + 0.08006295561790466, + 0.02983124554157257, + 0.03180251643061638, + 0.0352165661752224, + 0.12965072691440582, + 0.16171976923942566, + 0.09089557826519012 + ], + "credit_map": [ + 0.009052695706486702, + 0.013340387493371964, + 0.011325778439640999, + 0.009220005013048649, + 0.006729354616254568, + 0.011947987601161003, + 0.00917123630642891, + 0.013417219743132591, + 0.06483370065689087, + 0.030389323830604553, + 0.03969835117459297, + 0.04545733705163002, + 0.027148563414812088, + 0.030577264726161957, + 0.035727038979530334, + 0.030201327055692673, + 0.007195896934717894, + 0.006504081655293703, + 0.0050981780514121056, + 0.008341213688254356, + 0.0069806273095309734, + 0.006148262415081263, + 0.0054021840915083885, + 0.006912633776664734, + 0.01028339471668005, + 0.009767098352313042, + 0.01105901412665844, + 0.011293021962046623, + 0.008720907382667065, + 0.00945567898452282, + 0.0073310998268425465, + 0.01215274166315794, + 0.004530066624283791, + 0.004697774071246386, + 0.003959872294217348, + 0.0036704347003251314, + 0.0048074061051011086, + 0.002941864775493741, + 0.0026898577343672514, + 0.002533970633521676, + 0.004208612721413374, + 0.004244188312441111, + 0.003982014022767544, + 0.00347574008628726, + 0.0037391739897429943, + 0.004390715155750513, + 0.0034664503764361143, + 0.004295621998608112, + 0.004513232968747616, + 0.007904774509370327, + 0.0047993967309594154, + 0.004627395886927843, + 0.0032310672104358673, + 0.003856668947264552, + 0.0026001064106822014, + 0.003683923976495862, + 0.015218494459986687, + 0.014828846789896488, + 0.020004864782094955, + 0.01808624342083931, + 0.013545125722885132, + 0.019401153549551964, + 0.012525467202067375, + 0.01604052074253559, + 0.011024820618331432, + 0.01206591259688139, + 0.022570662200450897, + 0.015793530270457268, + 0.010724477469921112, + 0.02763228490948677, + 0.040816836059093475, + 0.02109123207628727, + 0.00777010340243578, + 0.00931927002966404, + 0.008273533545434475, + 0.01026859600096941, + 0.008582343347370625, + 0.014397145248949528, + 0.016662584617733955, + 0.015622002072632313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.01727217249572277 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.100344717502594 + ], + "opposite_logprob_tokens": [ + -2.350344657897949 + ], + "logprob_reward": -0.100344717502594, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.100344717502594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10894886404275894, + 0.08655767887830734, + 0.05896148458123207, + 0.15568436682224274, + 0.14511771500110626, + 0.04336853697896004, + 0.13627667725086212, + 0.13640272617340088, + 0.0797300636768341, + 0.0489518940448761 + ], + "credit_map": [ + 0.014264799654483795, + 0.009199963882565498, + 0.017566459253430367, + 0.010124881751835346, + 0.008586321026086807, + 0.016956506296992302, + 0.01309696864336729, + 0.019152961671352386, + 0.02509196475148201, + 0.008895582519471645, + 0.010376700200140476, + 0.008285628631711006, + 0.008136491291224957, + 0.008857169188559055, + 0.009278909303247929, + 0.007635224144905806, + 0.013384675607085228, + 0.007749376818537712, + 0.0050421347841620445, + 0.008008663542568684, + 0.006354030687361956, + 0.00628232304006815, + 0.00598104065284133, + 0.00615923386067152, + 0.023467829450964928, + 0.021237684413790703, + 0.01986728608608246, + 0.018122762441635132, + 0.023896751925349236, + 0.016254017129540443, + 0.01482696458697319, + 0.01801106706261635, + 0.026654278859496117, + 0.02212037704885006, + 0.016845114529132843, + 0.015352942049503326, + 0.019178513437509537, + 0.015062820166349411, + 0.013421117328107357, + 0.01648254692554474, + 0.005943240597844124, + 0.007855367846786976, + 0.005438459571450949, + 0.006509913597255945, + 0.0038150970358401537, + 0.004133810289204121, + 0.005192117765545845, + 0.004480525851249695, + 0.019306719303131104, + 0.03444221243262291, + 0.013527470640838146, + 0.013682871125638485, + 0.011645828373730183, + 0.01694394275546074, + 0.011561977677047253, + 0.015165637247264385, + 0.017830369994044304, + 0.020546583458781242, + 0.015316378325223923, + 0.013265624642372131, + 0.023076750338077545, + 0.018797505646944046, + 0.013419677503407001, + 0.01414982695132494, + 0.00615201611071825, + 0.007845818996429443, + 0.01129337027668953, + 0.007275300566107035, + 0.005356729030609131, + 0.012615642510354519, + 0.018374372273683548, + 0.010816806927323341, + 0.006236005574464798, + 0.005674139130860567, + 0.005887999199330807, + 0.005409389268606901, + 0.005789684597402811, + 0.007128821220248938, + 0.005923640448600054, + 0.006902208551764488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.100344717502594 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811967745423317 + ], + "opposite_logprob_tokens": [ + -3.2881197929382324 + ], + "logprob_reward": -0.03811967745423317, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811967745423317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.064628966152668, + 0.0886659249663353, + 0.045879486948251724, + 0.09895317256450653, + 0.05628306418657303, + 0.018470250070095062, + 0.03094048611819744, + 0.3878985345363617, + 0.12771087884902954, + 0.08056925982236862 + ], + "credit_map": [ + 0.006946783512830734, + 0.00702917343005538, + 0.00846448726952076, + 0.005475266370922327, + 0.005034005735069513, + 0.011953831650316715, + 0.005867376923561096, + 0.01385804358869791, + 0.022690510377287865, + 0.01009367499500513, + 0.013553795404732227, + 0.008716030977666378, + 0.009664338082075119, + 0.00822723750025034, + 0.009151794016361237, + 0.006568535231053829, + 0.007886902429163456, + 0.006389905698597431, + 0.006056757178157568, + 0.005563511047512293, + 0.006387732457369566, + 0.005057307425886393, + 0.003973897080868483, + 0.004563474096357822, + 0.015941204503178596, + 0.014702755957841873, + 0.012375864200294018, + 0.011770302429795265, + 0.015804126858711243, + 0.00978641677647829, + 0.009912609122693539, + 0.008659892715513706, + 0.010035211220383644, + 0.010556092485785484, + 0.006838858127593994, + 0.007357658818364143, + 0.006333863828331232, + 0.005212730728089809, + 0.004092040006071329, + 0.005856610834598541, + 0.0029886148404330015, + 0.0038045335095375776, + 0.0017768172547221184, + 0.0021514762192964554, + 0.0018334421329200268, + 0.0021234385203570127, + 0.0014659740263596177, + 0.0023259532172232866, + 0.004003969021141529, + 0.0075107142329216, + 0.004300209227949381, + 0.003647855017334223, + 0.0019371806411072612, + 0.003320342395454645, + 0.0030373167246580124, + 0.003182899672538042, + 0.0236640814691782, + 0.027240945026278496, + 0.04847360774874687, + 0.04607328772544861, + 0.02944079600274563, + 0.06806893646717072, + 0.07657147198915482, + 0.06836540997028351, + 0.010712986811995506, + 0.011308178305625916, + 0.02105974219739437, + 0.017998497933149338, + 0.009260371327400208, + 0.02041384018957615, + 0.023906195536255836, + 0.013051068410277367, + 0.00979576911777258, + 0.007215050980448723, + 0.009827597066760063, + 0.007189556490629911, + 0.009429067373275757, + 0.012531927786767483, + 0.01187701616436243, + 0.012703274376690388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.03811967745423317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353394612669945 + ], + "opposite_logprob_tokens": [ + -3.651353359222412 + ], + "logprob_reward": -0.026353394612669945, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353394612669945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11041108518838882, + 0.03924846276640892, + 0.12056619673967361, + 0.03732900694012642, + 0.048862677067518234, + 0.02709323912858963, + 0.034983739256858826, + 0.13592912256717682, + 0.32077908515930176, + 0.12479738891124725 + ], + "credit_map": [ + 0.015442456118762493, + 0.015901917591691017, + 0.014028509147465229, + 0.006505743134766817, + 0.006683764513581991, + 0.020459802821278572, + 0.007900453172624111, + 0.023488439619541168, + 0.00470155430957675, + 0.00443495437502861, + 0.004093537572771311, + 0.005243748426437378, + 0.0055984314531087875, + 0.0049163056537508965, + 0.004468058701604605, + 0.005791869480162859, + 0.022178931161761284, + 0.01565718837082386, + 0.013865917921066284, + 0.014209217391908169, + 0.015385470353066921, + 0.013802680186927319, + 0.011792488396167755, + 0.01367430854588747, + 0.004746840801090002, + 0.004535443149507046, + 0.004374445881694555, + 0.005579924676567316, + 0.00445360504090786, + 0.004157805349677801, + 0.0042462521232664585, + 0.0052346899174153805, + 0.00673954701051116, + 0.008264039643108845, + 0.006267135497182608, + 0.006525739561766386, + 0.006932457443326712, + 0.0049772243946790695, + 0.004242611117660999, + 0.004913921467959881, + 0.004089194815605879, + 0.006065021269023418, + 0.003213699674233794, + 0.002890840172767639, + 0.0020851758308708668, + 0.003147310810163617, + 0.002302394947037101, + 0.0032995999790728092, + 0.00494487676769495, + 0.00882885605096817, + 0.0037157058250159025, + 0.004106706473976374, + 0.002571220276877284, + 0.0039021875709295273, + 0.0026804758235812187, + 0.004233711399137974, + 0.012283574789762497, + 0.01145833171904087, + 0.02440737374126911, + 0.017450107261538506, + 0.012541726231575012, + 0.02168029174208641, + 0.018508469685912132, + 0.01759924553334713, + 0.02875213325023651, + 0.027255956083536148, + 0.047250937670469284, + 0.03403161093592644, + 0.01905856467783451, + 0.050985150039196014, + 0.07263311743736267, + 0.040811628103256226, + 0.015232369303703308, + 0.01611373946070671, + 0.011894137598574162, + 0.012996958568692207, + 0.013974824920296669, + 0.01816987618803978, + 0.017710508778691292, + 0.01870497688651085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.026353394612669945 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808605298399925 + ], + "opposite_logprob_tokens": [ + -3.2880859375 + ], + "logprob_reward": -0.03808605298399925, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808605298399925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1090579554438591, + 0.1674019992351532, + 0.05515729635953903, + 0.06851984560489655, + 0.03342600166797638, + 0.02725185826420784, + 0.0227233674377203, + 0.13389401137828827, + 0.2877037525177002, + 0.09486397355794907 + ], + "credit_map": [ + 0.01317052822560072, + 0.014977734535932541, + 0.010724004358053207, + 0.01093095913529396, + 0.009143933653831482, + 0.017248019576072693, + 0.010341592133045197, + 0.022521188482642174, + 0.0356689989566803, + 0.01695678010582924, + 0.023554624989628792, + 0.022503890097141266, + 0.01544804684817791, + 0.0172494538128376, + 0.018827039748430252, + 0.01719316840171814, + 0.010565639473497868, + 0.007714823819696903, + 0.006424042861908674, + 0.0073859803378582, + 0.005876566749066114, + 0.0059755356051027775, + 0.0047918083146214485, + 0.006422901526093483, + 0.006946430075913668, + 0.00651940144598484, + 0.013844039291143417, + 0.011772913858294487, + 0.003908623941242695, + 0.008880412206053734, + 0.00462193600833416, + 0.012026090174913406, + 0.005104342009872198, + 0.004754873923957348, + 0.005074512213468552, + 0.004950135014951229, + 0.003922917414456606, + 0.003087451448664069, + 0.0032813295256346464, + 0.003250439651310444, + 0.002996087307110429, + 0.004336935933679342, + 0.0026838614139705896, + 0.004513158928602934, + 0.002293934812769294, + 0.003484289860352874, + 0.003666357137262821, + 0.003277233801782131, + 0.003042720491066575, + 0.005162119399756193, + 0.0023158148396760225, + 0.004168093204498291, + 0.0014094914076849818, + 0.002442746190354228, + 0.001263026730157435, + 0.0029193551745265722, + 0.011257538571953773, + 0.010926200076937675, + 0.01932511292397976, + 0.016699766740202904, + 0.009120267815887928, + 0.022508826106786728, + 0.02360466495156288, + 0.020451638847589493, + 0.03772144764661789, + 0.041000332683324814, + 0.04125206172466278, + 0.029708132147789, + 0.024575844407081604, + 0.0403086319565773, + 0.03976406902074814, + 0.03337324410676956, + 0.01169233676046133, + 0.010801469907164574, + 0.010929533280432224, + 0.009988437406718731, + 0.011323372833430767, + 0.01425443310290575, + 0.01192488707602024, + 0.013949502259492874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.03808605298399925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead tracking shot follows a massive icebreaker as it majestically navigates the vast expanse of Arctic ice. Painted a striking, weathered red, its sturdy steel prow bears the scratches and rust of past voyages, it pushes forward with an unstoppable, steady, and powerful rhythm. Under the vibrant light of Makoto Shinkai's signature aesthetic—a low-hanging sun paints the sky a soft hue of pink, purple, and golden orange—each time the prow collides with the meters-thick ice, the world seems to freeze. As the ice groans and cracks with a bang, countless ice crystals are instantly launched into a spectacular cloud of particles. Under the extreme slow-motion effects, these ice crystals are no longer chaotic fragments, but dance and spin gracefully in the air. Each tiny crystal accurately captures and refracts the dreamy light in the sky, like billions of suspended diamonds, shining with a rainbow spectrum, and drawing out slender and beautiful halo stripes in the lens, combining the tremendous power of machinery with the instantaneous beauty of nature into a visual picture that is both magnificent and full of poetry and emotion.", + "seed": 25440, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_13/index_9519_seed_25440_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": -0.07481052218936383, + "vlm_logprob": -0.07481052218936383, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.07481052218936383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011834177188575268 + ], + "opposite_logprob_tokens": [ + -5.761834144592285 + ], + "logprob_reward": -0.011834177188575268, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.011834177188575268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12308135628700256, + 0.03949366882443428, + 0.05182330310344696, + 0.15512703359127045, + 0.1226750835776329, + 0.14662539958953857, + 0.061658188700675964, + 0.06470629572868347, + 0.09794347733259201, + 0.1368662565946579 + ], + "credit_map": [ + 0.011564208194613457, + 0.013780317269265652, + 0.021901322528719902, + 0.015521341003477573, + 0.010264038108289242, + 0.018839463591575623, + 0.016101056709885597, + 0.015109612606465816, + 0.0036878855898976326, + 0.004368966445326805, + 0.003426767885684967, + 0.0036805327981710434, + 0.0035411310382187366, + 0.007494695950299501, + 0.003951249644160271, + 0.009342435747385025, + 0.005596427246928215, + 0.006235633976757526, + 0.007106727920472622, + 0.004715773276984692, + 0.004899803549051285, + 0.008998444303870201, + 0.006337631028145552, + 0.007932858541607857, + 0.007796763442456722, + 0.00781354308128357, + 0.010304645635187626, + 0.007785785477608442, + 0.007415353786200285, + 0.05525795742869377, + 0.0041224174201488495, + 0.05463056638836861, + 0.01975402422249317, + 0.012059492990374565, + 0.010766536928713322, + 0.006781646981835365, + 0.030523495748639107, + 0.0175599567592144, + 0.00771081680431962, + 0.017519110813736916, + 0.029252544045448303, + 0.0182482972741127, + 0.014787915162742138, + 0.012548401951789856, + 0.01653583161532879, + 0.023022187873721123, + 0.012258767150342464, + 0.019971445202827454, + 0.008706672117114067, + 0.007025774102658033, + 0.012212717905640602, + 0.007093088701367378, + 0.0072570666670799255, + 0.007855083793401718, + 0.005862551275640726, + 0.00564523646607995, + 0.010673296637833118, + 0.011750426143407822, + 0.007040685974061489, + 0.006495273672044277, + 0.00915779359638691, + 0.008936556987464428, + 0.005802415311336517, + 0.004849845543503761, + 0.009889287874102592, + 0.011280754581093788, + 0.00818437710404396, + 0.010702320374548435, + 0.01455144863575697, + 0.0203385166823864, + 0.013682018034160137, + 0.00931475404649973, + 0.0188911110162735, + 0.017997022718191147, + 0.010826022364199162, + 0.01885628141462803, + 0.020866798236966133, + 0.014209120534360409, + 0.024354320019483566, + 0.01086556725203991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.011834177188575268 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31346824765205383 + ], + "opposite_logprob_tokens": [ + -1.3134682178497314 + ], + "logprob_reward": -0.31346824765205383, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31346824765205383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10255422443151474, + 0.0551939457654953, + 0.07470858842134476, + 0.1861344277858734, + 0.14536239206790924, + 0.07664533704519272, + 0.11234604567289352, + 0.06108919531106949, + 0.1226649284362793, + 0.0633009746670723 + ], + "credit_map": [ + 0.010154446586966515, + 0.010497935116291046, + 0.016265083104372025, + 0.011660248972475529, + 0.007028450258076191, + 0.017674261704087257, + 0.00836322084069252, + 0.02091056853532791, + 0.009448438882827759, + 0.006575844716280699, + 0.006758973002433777, + 0.004997908603399992, + 0.004885819274932146, + 0.009379885159432888, + 0.0062027303501963615, + 0.006944345310330391, + 0.007996313273906708, + 0.010486548766493797, + 0.008778802119195461, + 0.006977901794016361, + 0.008760775439441204, + 0.012203233316540718, + 0.006701662205159664, + 0.012803352437913418, + 0.011823072098195553, + 0.0066570742055773735, + 0.010279153473675251, + 0.006012186408042908, + 0.006940504536032677, + 0.06774869561195374, + 0.005921029485762119, + 0.07075270265340805, + 0.016580520197749138, + 0.017495516687631607, + 0.011269153095781803, + 0.013600735925137997, + 0.03592763468623161, + 0.02031557261943817, + 0.009743092581629753, + 0.02043015882372856, + 0.007627654355019331, + 0.010671773925423622, + 0.010161833837628365, + 0.005674190819263458, + 0.013044433668255806, + 0.011668402701616287, + 0.007260486483573914, + 0.010536556132137775, + 0.014877064153552055, + 0.021825416013598442, + 0.009888914413750172, + 0.00926814042031765, + 0.017490969970822334, + 0.013698776252567768, + 0.009446972981095314, + 0.015849772840738297, + 0.00909634493291378, + 0.011504603549838066, + 0.0086413798853755, + 0.009050916880369186, + 0.006605743430554867, + 0.007027262821793556, + 0.006077372469007969, + 0.003085569478571415, + 0.011095110327005386, + 0.020287128165364265, + 0.011431778781116009, + 0.01805243082344532, + 0.013415489345788956, + 0.018534798175096512, + 0.019676964730024338, + 0.010171216912567616, + 0.006944982334971428, + 0.010544152930378914, + 0.007288631983101368, + 0.009785979986190796, + 0.009508408606052399, + 0.008370384573936462, + 0.005961782764643431, + 0.004896646365523338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.31346824765205383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337202288210392 + ], + "opposite_logprob_tokens": [ + -3.4087202548980713 + ], + "logprob_reward": -0.0337202288210392, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337202288210392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049365025013685226, + 0.032471999526023865, + 0.027580786496400833, + 0.3278122544288635, + 0.06611286848783493, + 0.08556432276964188, + 0.046861544251441956, + 0.11125040799379349, + 0.21377193927764893, + 0.03920890390872955 + ], + "credit_map": [ + 0.006111739668995142, + 0.005764521658420563, + 0.005034079309552908, + 0.005267764907330275, + 0.003802401712164283, + 0.009572182781994343, + 0.0049076322466135025, + 0.008904702961444855, + 0.0024703845847398043, + 0.00510797742754221, + 0.0030840111430734396, + 0.0029009042773395777, + 0.0028502601198852062, + 0.006112217903137207, + 0.002928996691480279, + 0.0070172459818422794, + 0.0018423941219225526, + 0.003904803190380335, + 0.0025723229628056288, + 0.003571090754121542, + 0.004061387851834297, + 0.004431657958775759, + 0.002830691635608673, + 0.004366439301520586, + 0.009661052376031876, + 0.006949884817004204, + 0.008714297786355019, + 0.007151150144636631, + 0.006055654492229223, + 0.14210566878318787, + 0.005660984665155411, + 0.14151354134082794, + 0.007937761954963207, + 0.007779195439070463, + 0.009681282564997673, + 0.006924925372004509, + 0.010995874181389809, + 0.008113425225019455, + 0.006779422517865896, + 0.007900978438556194, + 0.01749810017645359, + 0.0074719698168337345, + 0.01007121056318283, + 0.006000245455652475, + 0.01218190137296915, + 0.01585720106959343, + 0.00651762168854475, + 0.009966068901121616, + 0.006071166135370731, + 0.009072784334421158, + 0.005911984946578741, + 0.004649213515222073, + 0.006597265601158142, + 0.005355430766940117, + 0.003932268358767033, + 0.005271431524306536, + 0.02563653141260147, + 0.012727979570627213, + 0.011346818879246712, + 0.014403566718101501, + 0.011550004594027996, + 0.01282008457928896, + 0.015673937276005745, + 0.0070914835669100285, + 0.025972377508878708, + 0.033636461943387985, + 0.030247323215007782, + 0.027787426486611366, + 0.03328236564993858, + 0.024424534291028976, + 0.024001337587833405, + 0.014420097693800926, + 0.0037476744037121534, + 0.0061034769751131535, + 0.004741023760288954, + 0.005708300042897463, + 0.0055650644935667515, + 0.004461079370230436, + 0.004090495407581329, + 0.004791788756847382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.0337202288210392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254265196621418 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.01254265196621418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254265196621418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08894756436347961, + 0.029988544061779976, + 0.032774802297353745, + 0.26323896646499634, + 0.09888960421085358, + 0.06976000219583511, + 0.08576004952192307, + 0.04774821922183037, + 0.21826103329658508, + 0.0646311342716217 + ], + "credit_map": [ + 0.011687492951750755, + 0.007835879921913147, + 0.014771945774555206, + 0.00697705103084445, + 0.005505733657628298, + 0.01580394245684147, + 0.008278308436274529, + 0.018087204545736313, + 0.0028169923461973667, + 0.004848869051784277, + 0.0047327750362455845, + 0.004296422936022282, + 0.0030936049297451973, + 0.004166694357991219, + 0.0024163397029042244, + 0.003616845468059182, + 0.0036306267138570547, + 0.003399091772735119, + 0.004836271982640028, + 0.0030724203679710627, + 0.002372471848502755, + 0.005456947255879641, + 0.003358048154041171, + 0.0066489242017269135, + 0.008402041159570217, + 0.005738560575991869, + 0.006198232527822256, + 0.0042688860557973385, + 0.005066294688731432, + 0.11473988741636276, + 0.004182810895144939, + 0.11464225500822067, + 0.01139380969107151, + 0.011078749783337116, + 0.013319194316864014, + 0.009528147988021374, + 0.019059723243117332, + 0.012319824658334255, + 0.010367700830101967, + 0.01182244997471571, + 0.009804432280361652, + 0.005983774084597826, + 0.006064253393560648, + 0.004493845161050558, + 0.009870870970189571, + 0.016711249947547913, + 0.0058398800902068615, + 0.010991697199642658, + 0.010518154129385948, + 0.015527348965406418, + 0.009602589532732964, + 0.010029674507677555, + 0.013297973200678825, + 0.010203521698713303, + 0.007436801213771105, + 0.009143983945250511, + 0.004977807868272066, + 0.010809837840497494, + 0.005294688977301121, + 0.003681792877614498, + 0.006376136094331741, + 0.007576265372335911, + 0.004887958988547325, + 0.004143735393881798, + 0.024444906041026115, + 0.039607785642147064, + 0.028426973149180412, + 0.03043919801712036, + 0.029352689161896706, + 0.026141416281461716, + 0.024804646149277687, + 0.015043433755636215, + 0.007848942652344704, + 0.01314961165189743, + 0.00654997443780303, + 0.008884821087121964, + 0.008722318336367607, + 0.007828369736671448, + 0.006398861296474934, + 0.00524823646992445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.01254265196621418 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 icebreaker ship painted a striking, weathered red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned overhead and tracking the ship as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship collide with meters-thick ice, causing it to crack and launch ice crystals into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ice crystals shown in extreme slow motion, dancing and spinning gracefully while refracting light into rainbow spectra and halo stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky lit by a low-hanging sun with soft hues of pink, purple, and golden orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024873053189367056 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024873053189367056, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024873053189367056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23582513630390167, + 0.09574884921312332, + 0.04251463711261749, + 0.23306867480278015, + 0.09682880342006683, + 0.05055253952741623, + 0.03836829960346222, + 0.05695372074842453, + 0.10841502249240875, + 0.04172428324818611 + ], + "credit_map": [ + 0.019735857844352722, + 0.016856029629707336, + 0.02558615244925022, + 0.01308123767375946, + 0.012996709905564785, + 0.06394847482442856, + 0.01618173159658909, + 0.06743896007537842, + 0.007598757743835449, + 0.00611094431951642, + 0.009742950089275837, + 0.004445082042366266, + 0.004275376908481121, + 0.028761666268110275, + 0.004666829016059637, + 0.030147254467010498, + 0.003929785452783108, + 0.00421468960121274, + 0.005296140909194946, + 0.0030629774555563927, + 0.0030505696777254343, + 0.009431758895516396, + 0.0034487235825508833, + 0.010079995729029179, + 0.007064776495099068, + 0.00730125792324543, + 0.004832119680941105, + 0.005071169696748257, + 0.006144278217107058, + 0.09913559257984161, + 0.0041978200897574425, + 0.09932167828083038, + 0.01126896869391203, + 0.009077340364456177, + 0.007498434279114008, + 0.007539256475865841, + 0.023955058306455612, + 0.015037625096738338, + 0.006530303042382002, + 0.015921825543045998, + 0.0057610198855400085, + 0.009860657155513763, + 0.005362099967896938, + 0.003576932242140174, + 0.007955015636980534, + 0.00861760601401329, + 0.004447681363672018, + 0.004971529822796583, + 0.0055868858471512794, + 0.006574371829628944, + 0.005160716362297535, + 0.004236204549670219, + 0.005077352747321129, + 0.004130556248128414, + 0.004066575784236193, + 0.003535640425980091, + 0.008017086423933506, + 0.006729898974299431, + 0.007758740801364183, + 0.00579857500270009, + 0.005761890206485987, + 0.011862574145197868, + 0.005521383602172136, + 0.005503569729626179, + 0.01298998948186636, + 0.020200051367282867, + 0.01240378525108099, + 0.013305443339049816, + 0.015588548965752125, + 0.01427557971328497, + 0.01135400589555502, + 0.008297628723084927, + 0.005290604196488857, + 0.006747346371412277, + 0.003499212209135294, + 0.006066946778446436, + 0.006815630942583084, + 0.005480557214468718, + 0.003919435199350119, + 0.0039045503363013268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0024873053189367056 + } + ], + "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/0000128/group_14/index_tensor([4148, 4148])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_14/index_tensor([4148, 4148])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..141e286cec04989c65ef37a8581657143bc7dc40 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_14/index_tensor([4148, 4148])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic overhead shot reveals a massive lava insect rising from the utter darkness of the void. The creature's massive body is composed of a cracked obsidian carapace, a glow of incandescent magma seeping from deep within the cracks, as if the fury of the earth's core has been given life. Its translucent wings, like superheated amber, are covered in an intricate web of fiery veins and vibrate explosively at a rate too fast to be perceived by the naked eye. These frenzied movements instantly unleash a chaotic storm of countless glowing embers and delicate, swirling ash, resembling an inverted volcanic eruption against the pitch-black backdrop. With each violent flap of its wings, its massive form is propelled vertically upward by an unparalleled force, soaring into the sky. A strong and dramatic backlight originating from behind it completely engulfs the side of it facing the camera in deep shadows, but the sharp flames like a knife edge outline the contours of each of its jagged limbs and hideous head, forming the iconic and tense contrast of light and shadow of the Caravaggio school. The whole scene is filled with a primitive, violent and awe-inspiring doomsday aesthetic atmosphere.", + "seed": 24706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_14/index_4148_seed_24706_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": -0.008485016506165266, + "vlm_logprob": -0.008485016506165266, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.008485016506165266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015033048577606678 + ], + "opposite_logprob_tokens": [ + -4.39003324508667 + ], + "logprob_reward": -0.015033048577606678, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015033048577606678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10817717760801315, + 0.04754849150776863, + 0.322054922580719, + 0.05297892168164253, + 0.038065388798713684, + 0.17318543791770935, + 0.07657721638679504, + 0.05848235264420509, + 0.06341502070426941, + 0.05951511487364769 + ], + "credit_map": [ + 0.010739408433437347, + 0.012510393746197224, + 0.015561901032924652, + 0.02128896676003933, + 0.012558580376207829, + 0.013411074876785278, + 0.010046782903373241, + 0.012060065753757954, + 0.005519839469343424, + 0.005703422706574202, + 0.005358485039323568, + 0.0058864448219537735, + 0.004919284489005804, + 0.005581778008490801, + 0.008339596912264824, + 0.00623963912948966, + 0.060929443687200546, + 0.039128560572862625, + 0.04176632687449455, + 0.03327673301100731, + 0.024599384516477585, + 0.047397155314683914, + 0.024844884872436523, + 0.050112441182136536, + 0.007061605807393789, + 0.006382954306900501, + 0.007073706481605768, + 0.003894916968420148, + 0.0033626907970756292, + 0.009420688264071941, + 0.0061293430626392365, + 0.00965301226824522, + 0.005008331499993801, + 0.005154402460902929, + 0.00843630451709032, + 0.00391105143353343, + 0.0036007205490022898, + 0.005078705959022045, + 0.0029790750704705715, + 0.0038967998698353767, + 0.019952449947595596, + 0.019952110946178436, + 0.03344598412513733, + 0.024447713047266006, + 0.012938088737428188, + 0.013352066278457642, + 0.021121937781572342, + 0.027975084260106087, + 0.008344336412847042, + 0.009379593655467033, + 0.010180821642279625, + 0.010601703077554703, + 0.010381733067333698, + 0.009705371223390102, + 0.008615206927061081, + 0.009368451312184334, + 0.006780638825148344, + 0.008534769527614117, + 0.006631333380937576, + 0.005081691313534975, + 0.00870452169328928, + 0.009603665210306644, + 0.005921943113207817, + 0.007223790977150202, + 0.007770576514303684, + 0.00873236171901226, + 0.006859488785266876, + 0.006760048680007458, + 0.010215743444859982, + 0.009277270175516605, + 0.006693905685096979, + 0.007105627562850714, + 0.007761887274682522, + 0.007087167352437973, + 0.011361904442310333, + 0.00726416427642107, + 0.005127614364027977, + 0.008039843291044235, + 0.0074886237271130085, + 0.0053839124739170074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.015033048577606678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036406919825822115 + ], + "opposite_logprob_tokens": [ + -5.628640651702881 + ], + "logprob_reward": -0.0036406919825822115, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036406919825822115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09114321321249008, + 0.049008406698703766, + 0.20990443229675293, + 0.06708280742168427, + 0.028919387608766556, + 0.1044163778424263, + 0.15683196485042572, + 0.06466975808143616, + 0.16872987151145935, + 0.05929376557469368 + ], + "credit_map": [ + 0.012224097736179829, + 0.014323195442557335, + 0.01354389637708664, + 0.013759681023657322, + 0.007850860245525837, + 0.00987408496439457, + 0.008817516267299652, + 0.010749883949756622, + 0.00554188946262002, + 0.010349475778639317, + 0.004846520721912384, + 0.005315050948411226, + 0.004707620944827795, + 0.006237144581973553, + 0.0044579836539924145, + 0.007552719675004482, + 0.042211953550577164, + 0.02612162195146084, + 0.027012662962079048, + 0.021537160500884056, + 0.017375532537698746, + 0.029887927696108818, + 0.015457804314792156, + 0.03029976598918438, + 0.009876294992864132, + 0.007105628028512001, + 0.007378553040325642, + 0.004290545824915171, + 0.003057725727558136, + 0.012564534321427345, + 0.00602942006662488, + 0.016780102625489235, + 0.0037401120644062757, + 0.0036323308013379574, + 0.00414099870249629, + 0.002405013656243682, + 0.004672706127166748, + 0.003678882261738181, + 0.003232963616028428, + 0.0034163817763328552, + 0.008775866590440273, + 0.013616982847452164, + 0.035898227244615555, + 0.007530604023486376, + 0.004320037551224232, + 0.005641327705234289, + 0.005369189195334911, + 0.023264145478606224, + 0.022847160696983337, + 0.01402746420353651, + 0.023754579946398735, + 0.017921291291713715, + 0.015765834599733353, + 0.022905655205249786, + 0.017595022916793823, + 0.022014951333403587, + 0.008096697740256786, + 0.007118447683751583, + 0.008427889086306095, + 0.007103930227458477, + 0.006903704255819321, + 0.00900338776409626, + 0.008054266683757305, + 0.009961435571312904, + 0.021166309714317322, + 0.017577653750777245, + 0.024733709171414375, + 0.022012319415807724, + 0.015606021508574486, + 0.02205294370651245, + 0.01896354928612709, + 0.026617364957928658, + 0.008397000841796398, + 0.0062566339038312435, + 0.013255255296826363, + 0.007192990742623806, + 0.0052507915534079075, + 0.008128015324473381, + 0.005981610622256994, + 0.004831467755138874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.0036406919825822115 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195065014064312 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195065014064312, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195065014064312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17809531092643738, + 0.10334865748882294, + 0.15136955678462982, + 0.0508541502058506, + 0.04674911126494408, + 0.0585908368229866, + 0.13494761288166046, + 0.11396409571170807, + 0.10187339782714844, + 0.06020735204219818 + ], + "credit_map": [ + 0.024608522653579712, + 0.01850588619709015, + 0.026282690465450287, + 0.020830899477005005, + 0.01824946328997612, + 0.021883303299546242, + 0.020221730694174767, + 0.027512796223163605, + 0.021527942270040512, + 0.016343101859092712, + 0.013949986547231674, + 0.008101335726678371, + 0.006472869776189327, + 0.013014323078095913, + 0.01280797179788351, + 0.011131116189062595, + 0.023437824100255966, + 0.0160865169018507, + 0.014244187623262405, + 0.019989222288131714, + 0.012851695530116558, + 0.024302782490849495, + 0.012080004438757896, + 0.02837730571627617, + 0.005287564359605312, + 0.006268290802836418, + 0.005793483927845955, + 0.004447222221642733, + 0.004192692693322897, + 0.007864058949053288, + 0.006683036684989929, + 0.010317797772586346, + 0.005651263054460287, + 0.006563273724168539, + 0.004771027248352766, + 0.004459788091480732, + 0.008326147682964802, + 0.0055092922411859035, + 0.005520120728760958, + 0.005948199424892664, + 0.006335277575999498, + 0.00849646981805563, + 0.00854585226625204, + 0.007948965765535831, + 0.004838707856833935, + 0.006359995808452368, + 0.008275332860648632, + 0.007790225557982922, + 0.014401120133697987, + 0.015018338337540627, + 0.018642274662852287, + 0.017537135630846024, + 0.014823177829384804, + 0.021780598908662796, + 0.015309231355786324, + 0.01743573695421219, + 0.013499535620212555, + 0.013187352567911148, + 0.01228329911828041, + 0.011187871918082237, + 0.01953194849193096, + 0.016798539087176323, + 0.015034142881631851, + 0.012441398575901985, + 0.010915204882621765, + 0.013243884779512882, + 0.014506992883980274, + 0.01064329594373703, + 0.011400649324059486, + 0.014277270063757896, + 0.009707793593406677, + 0.017178291454911232, + 0.00859653577208519, + 0.007236805744469166, + 0.01578211784362793, + 0.005956582259386778, + 0.005158704239875078, + 0.0086178844794631, + 0.005109868478029966, + 0.0037488483358174562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.014195065014064312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736189126968384 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736189126968384, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736189126968384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11051187664270401, + 0.035641636699438095, + 0.5611914992332458, + 0.03190186247229576, + 0.027360640466213226, + 0.05015164613723755, + 0.03460825979709625, + 0.05206027254462242, + 0.04701311141252518, + 0.04955919459462166 + ], + "credit_map": [ + 0.010501044802367687, + 0.011283708736300468, + 0.016186237335205078, + 0.020014502108097076, + 0.010918250307440758, + 0.013072186149656773, + 0.014706340618431568, + 0.013829605653882027, + 0.004473860375583172, + 0.004020883236080408, + 0.004503131378442049, + 0.003433585399761796, + 0.0032670481596142054, + 0.0049797785468399525, + 0.0044797975569963455, + 0.006483553443104029, + 0.12671944499015808, + 0.07614897191524506, + 0.07282952219247818, + 0.05583081766963005, + 0.041967011988162994, + 0.07485384494066238, + 0.04210905358195305, + 0.07073285430669785, + 0.004569787532091141, + 0.0033232325222343206, + 0.004087869543582201, + 0.002752356231212616, + 0.0024501807056367397, + 0.004727710038423538, + 0.003639791626483202, + 0.006350934971123934, + 0.004176948685199022, + 0.004378989804536104, + 0.00527327973395586, + 0.002593760844320059, + 0.0019590023439377546, + 0.003878176910802722, + 0.0024207376409322023, + 0.0026797442696988583, + 0.0038241338916122913, + 0.0053566936403512955, + 0.012716301716864109, + 0.00582158612087369, + 0.00298033538274467, + 0.0032285116612911224, + 0.00489375414326787, + 0.011330333538353443, + 0.005455937702208757, + 0.002907161833718419, + 0.006549706682562828, + 0.0038710038643330336, + 0.0030916461255401373, + 0.005878165829926729, + 0.00302147283218801, + 0.0038331628311425447, + 0.006179865915328264, + 0.005245383828878403, + 0.007551402784883976, + 0.004799538291990757, + 0.007225589826703072, + 0.007383017800748348, + 0.006333775818347931, + 0.007341699209064245, + 0.0059454492293298244, + 0.005869074258953333, + 0.005837008822709322, + 0.005112777464091778, + 0.007074689492583275, + 0.006486805155873299, + 0.005621335003525019, + 0.005065974313765764, + 0.008057218044996262, + 0.007011081092059612, + 0.007835480384528637, + 0.004312498029321432, + 0.003665811847895384, + 0.008014507591724396, + 0.005719722248613834, + 0.004942870698869228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.006736189126968384 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820087829604745 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820087829604745, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820087829604745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12022753059864044, + 0.05622287467122078, + 0.37318113446235657, + 0.03402554988861084, + 0.06935141980648041, + 0.07317913323640823, + 0.09761494398117065, + 0.058427128940820694, + 0.07087607681751251, + 0.046894222497940063 + ], + "credit_map": [ + 0.023194672539830208, + 0.021753814071416855, + 0.012117546983063221, + 0.013401036150753498, + 0.012049290351569653, + 0.015441881492733955, + 0.0108767906203866, + 0.011392496526241302, + 0.00849459134042263, + 0.009577762335538864, + 0.007474545855075121, + 0.004594509955495596, + 0.005037661176174879, + 0.007337040267884731, + 0.005630718544125557, + 0.008076043799519539, + 0.08682345598936081, + 0.049839314073324203, + 0.04672062769532204, + 0.040477532893419266, + 0.027485879138112068, + 0.048837192356586456, + 0.028568657115101814, + 0.04442849010229111, + 0.005056628491729498, + 0.004545707255601883, + 0.005017880350351334, + 0.003365685697644949, + 0.0024142363108694553, + 0.0057591586373746395, + 0.0035067005082964897, + 0.004359554499387741, + 0.010607371106743813, + 0.009403439238667488, + 0.011572729796171188, + 0.007270054891705513, + 0.007096588145941496, + 0.01065314281731844, + 0.006708654575049877, + 0.006039436440914869, + 0.00740424171090126, + 0.009379922412335873, + 0.012674636207520962, + 0.007935668341815472, + 0.006114139687269926, + 0.005798102356493473, + 0.011052742600440979, + 0.012819680385291576, + 0.010601414367556572, + 0.008778193965554237, + 0.014424223452806473, + 0.012988322414457798, + 0.009522194042801857, + 0.015433049760758877, + 0.012254899367690086, + 0.013612651266157627, + 0.007378616835922003, + 0.00845448486506939, + 0.0061041428707540035, + 0.005963004659861326, + 0.006952608469873667, + 0.008862408809363842, + 0.007397076115012169, + 0.007314786780625582, + 0.010751510038971901, + 0.009590777568519115, + 0.008236173540353775, + 0.009467331692576408, + 0.008135608397424221, + 0.009506558999419212, + 0.006812774110585451, + 0.008375338278710842, + 0.006908326409757137, + 0.006410373840481043, + 0.009222457185387611, + 0.004255358595401049, + 0.004560778383165598, + 0.007676777429878712, + 0.003985878545790911, + 0.003874274203553796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.002820087829604745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic overhead shot reveals a massive lava insect rising from the utter darkness of the void. The creature's massive body is composed of a cracked obsidian carapace, a glow of incandescent magma seeping from deep within the cracks, as if the fury of the earth's core has been given life. Its translucent wings, like superheated amber, are covered in an intricate web of fiery veins and vibrate explosively at a rate too fast to be perceived by the naked eye. These frenzied movements instantly unleash a chaotic storm of countless glowing embers and delicate, swirling ash, resembling an inverted volcanic eruption against the pitch-black backdrop. With each violent flap of its wings, its massive form is propelled vertically upward by an unparalleled force, soaring into the sky. A strong and dramatic backlight originating from behind it completely engulfs the side of it facing the camera in deep shadows, but the sharp flames like a knife edge outline the contours of each of its jagged limbs and hideous head, forming the iconic and tense contrast of light and shadow of the Caravaggio school. The whole scene is filled with a primitive, violent and awe-inspiring doomsday aesthetic atmosphere.", + "seed": 24707, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_14/index_4148_seed_24707_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": -0.008968492317944766, + "vlm_logprob": -0.008968492317944766, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.008968492317944766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024623852223157883 + ], + "opposite_logprob_tokens": [ + -3.7746238708496094 + ], + "logprob_reward": -0.024623852223157883, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024623852223157883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08053820580244064, + 0.04122726991772652, + 0.03963170945644379, + 0.09015268087387085, + 0.07449150830507278, + 0.10077903419733047, + 0.06273363530635834, + 0.33266592025756836, + 0.14354108273983002, + 0.03423898667097092 + ], + "credit_map": [ + 0.00875056628137827, + 0.014320926740765572, + 0.008866822347044945, + 0.0072646173648536205, + 0.008947648108005524, + 0.013142179697751999, + 0.010234889574348927, + 0.009010554291307926, + 0.004123109858483076, + 0.004663726780563593, + 0.004937959834933281, + 0.0046667251735925674, + 0.006627664901316166, + 0.004335732199251652, + 0.006290281191468239, + 0.0055820695124566555, + 0.007753524463623762, + 0.003674953943118453, + 0.006217739079147577, + 0.003355140797793865, + 0.004316198639571667, + 0.005116153974086046, + 0.003948285710066557, + 0.005249710753560066, + 0.007968718186020851, + 0.010632263496518135, + 0.009483234025537968, + 0.009360691532492638, + 0.010270548053085804, + 0.009656217880547047, + 0.017852863296866417, + 0.014928148128092289, + 0.008139129728078842, + 0.008347072638571262, + 0.013376421295106411, + 0.009023603051900864, + 0.0069198403507471085, + 0.009097960777580738, + 0.00999277550727129, + 0.009594705887138844, + 0.01237098965793848, + 0.011164252646267414, + 0.01907363161444664, + 0.016316456720232964, + 0.00747342174872756, + 0.006300901994109154, + 0.013723711483180523, + 0.014355666004121304, + 0.006122806575149298, + 0.00691650016233325, + 0.012111513875424862, + 0.007928390055894852, + 0.006070683244615793, + 0.005062911659479141, + 0.009091056883335114, + 0.009429769590497017, + 0.025409765541553497, + 0.017860999330878258, + 0.10497626662254333, + 0.035045940428972244, + 0.018587859347462654, + 0.014592648483812809, + 0.02599654719233513, + 0.0901959165930748, + 0.01805385760962963, + 0.015138709917664528, + 0.03026358038187027, + 0.014598585665225983, + 0.017693009227514267, + 0.008054216392338276, + 0.02209993079304695, + 0.01763918809592724, + 0.00477165263146162, + 0.002847802359610796, + 0.006729540880769491, + 0.005317113362252712, + 0.0023835606407374144, + 0.0017890046583488584, + 0.004583222791552544, + 0.005817087832838297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.024623852223157883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005276799667626619 + ], + "opposite_logprob_tokens": [ + -5.255276679992676 + ], + "logprob_reward": -0.005276799667626619, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005276799667626619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049654506146907806, + 0.023326506838202477, + 0.04692617803812027, + 0.09181670844554901, + 0.04722927138209343, + 0.12596188485622406, + 0.13574925065040588, + 0.3363202214241028, + 0.08413641154766083, + 0.058879051357507706 + ], + "credit_map": [ + 0.004705443978309631, + 0.007074305322021246, + 0.005865360144525766, + 0.004593148361891508, + 0.007735047489404678, + 0.007497698999941349, + 0.006655033212155104, + 0.005528469104319811, + 0.002746553160250187, + 0.0026927937287837267, + 0.0023231375962495804, + 0.0028491774573922157, + 0.003937256056815386, + 0.002282005501911044, + 0.0028192016761749983, + 0.003676382591947913, + 0.006298650987446308, + 0.003456042381003499, + 0.005612632725387812, + 0.0048400345258414745, + 0.0038841792847961187, + 0.006799932103604078, + 0.0073548248037695885, + 0.008679880760610104, + 0.008854713290929794, + 0.011627672240138054, + 0.012951299548149109, + 0.007395986467599869, + 0.012303249910473824, + 0.011787567287683487, + 0.013617929071187973, + 0.013278290629386902, + 0.006048886105418205, + 0.006375190801918507, + 0.008322709240019321, + 0.00457131490111351, + 0.004959804005920887, + 0.005014605354517698, + 0.005952049978077412, + 0.0059847142547369, + 0.010888609103858471, + 0.014992749318480492, + 0.014667869545519352, + 0.017376530915498734, + 0.01591138355433941, + 0.014716087840497494, + 0.021546335890889168, + 0.01586231403052807, + 0.014538551680743694, + 0.01736728474497795, + 0.020211735740303993, + 0.0161778274923563, + 0.020453117787837982, + 0.013834293931722641, + 0.01658741384744644, + 0.01657901704311371, + 0.025875261053442955, + 0.01927024871110916, + 0.1026851236820221, + 0.03648734092712402, + 0.019124796614050865, + 0.015836702659726143, + 0.02213127538561821, + 0.09490946680307388, + 0.010753074660897255, + 0.005978308618068695, + 0.019666584208607674, + 0.0122294332832098, + 0.005764481611549854, + 0.003237710101529956, + 0.015412205830216408, + 0.011094611138105392, + 0.007637058384716511, + 0.005686589982360601, + 0.011021477170288563, + 0.008518901653587818, + 0.0036074775271117687, + 0.0047517819330096245, + 0.008446180261671543, + 0.009209586307406425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.005276799667626619 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098350182175636 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098350182175636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098350182175636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05337030813097954, + 0.06550367921590805, + 0.038715068250894547, + 0.04846356436610222, + 0.032561153173446655, + 0.035717085003852844, + 0.05813850834965706, + 0.1611953228712082, + 0.3752630650997162, + 0.1310722380876541 + ], + "credit_map": [ + 0.006571952253580093, + 0.007336431182920933, + 0.005410575307905674, + 0.004457093775272369, + 0.005398342851549387, + 0.012170668691396713, + 0.0072058723308146, + 0.0048193736001849174, + 0.007647226564586163, + 0.006816802080720663, + 0.00978095829486847, + 0.006618930026888847, + 0.01106496062129736, + 0.007971448823809624, + 0.0074838921427726746, + 0.008119461126625538, + 0.00625262875109911, + 0.002505748998373747, + 0.004170364234596491, + 0.002935216762125492, + 0.0031112751457840204, + 0.004295777529478073, + 0.008386174216866493, + 0.007057883311063051, + 0.005329343024641275, + 0.0037375965621322393, + 0.009577200748026371, + 0.006050528958439827, + 0.005506318528205156, + 0.0041453102603554726, + 0.006217112764716148, + 0.007900155149400234, + 0.00531837297603488, + 0.0029101979453116655, + 0.0050087571144104, + 0.004442981444299221, + 0.0030104732140898705, + 0.0028885742649435997, + 0.0044334204867482185, + 0.004548375960439444, + 0.0036571102682501078, + 0.0038540305104106665, + 0.00489465706050396, + 0.0045698112808167934, + 0.0037556153256446123, + 0.0029481854289770126, + 0.00706904474645853, + 0.004968630149960518, + 0.005669728387147188, + 0.004754109773784876, + 0.012517676688730717, + 0.008253301493823528, + 0.004041007719933987, + 0.0029835232999175787, + 0.011225052177906036, + 0.008694110438227654, + 0.021997079253196716, + 0.01017373614013195, + 0.04026971757411957, + 0.022175421938300133, + 0.008329503238201141, + 0.009474579244852066, + 0.011088315397500992, + 0.03768698126077652, + 0.04324023053050041, + 0.029318654909729958, + 0.08399546891450882, + 0.05458487942814827, + 0.028896668925881386, + 0.016293969005346298, + 0.07543909549713135, + 0.04349411651492119, + 0.02007412537932396, + 0.01102486066520214, + 0.026507394388318062, + 0.02190076932311058, + 0.007374939508736134, + 0.007111907470971346, + 0.018900981172919273, + 0.018177246674895287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.004098350182175636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641467429697514 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.008641467429697514, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641467429697514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16407272219657898, + 0.012236633338034153, + 0.02720917947590351, + 0.04497245326638222, + 0.04698681831359863, + 0.057066578418016434, + 0.050999872386455536, + 0.19637902081012726, + 0.2957736551761627, + 0.10430310666561127 + ], + "credit_map": [ + 0.01016164105385542, + 0.047810692340135574, + 0.011110978201031685, + 0.008504529483616352, + 0.014844842255115509, + 0.045210737735033035, + 0.01471043098717928, + 0.0117188710719347, + 0.0017501641996204853, + 0.0011928988387808204, + 0.001971513731405139, + 0.0016483762301504612, + 0.0019005322828888893, + 0.0011631334200501442, + 0.0012971187243238091, + 0.0013128957943990827, + 0.005103333853185177, + 0.002690057037398219, + 0.0037854567635804415, + 0.0018313751788809896, + 0.0028598022181540728, + 0.004758685361593962, + 0.0029539770912379026, + 0.0032264923211187124, + 0.004121801815927029, + 0.004603675100952387, + 0.007031414192169905, + 0.004582195542752743, + 0.004830708261579275, + 0.004362574312835932, + 0.008136896416544914, + 0.00730318995192647, + 0.0068998392671346664, + 0.005098051391541958, + 0.008491713553667068, + 0.005730446428060532, + 0.0046779378317296505, + 0.003955870401114225, + 0.0063951085321605206, + 0.005737850908190012, + 0.007786089554429054, + 0.004762416705489159, + 0.014795872382819653, + 0.0074262167327106, + 0.003463726956397295, + 0.0028774659149348736, + 0.00931637641042471, + 0.006638415157794952, + 0.0052238390780985355, + 0.004940049722790718, + 0.00843067653477192, + 0.008252530358731747, + 0.005469376686960459, + 0.0031531178392469883, + 0.009321678429841995, + 0.006208603270351887, + 0.03684809058904648, + 0.01626632548868656, + 0.027437638491392136, + 0.038503292948007584, + 0.013407030142843723, + 0.013534479774534702, + 0.020142167806625366, + 0.030240003019571304, + 0.044714685529470444, + 0.022072816267609596, + 0.07771138846874237, + 0.03702929988503456, + 0.014887683093547821, + 0.012942321598529816, + 0.04810468479990959, + 0.03831076622009277, + 0.01336879376322031, + 0.007943348027765751, + 0.02422183007001877, + 0.013520337641239166, + 0.005479650106281042, + 0.006037652026861906, + 0.019117597490549088, + 0.014613897539675236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.008641467429697514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022019920870661736 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.0022019920870661736, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022019920870661736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25650277733802795, + 0.056724511086940765, + 0.017166757956147194, + 0.029161542654037476, + 0.04984241724014282, + 0.20337307453155518, + 0.05156773328781128, + 0.12793654203414917, + 0.10521162301301956, + 0.10251300781965256 + ], + "credit_map": [ + 0.013824948109686375, + 0.0731377899646759, + 0.014444184489548206, + 0.012128673493862152, + 0.028901202604174614, + 0.07349859923124313, + 0.017199838533997536, + 0.02336755208671093, + 0.006309753283858299, + 0.006678673904389143, + 0.007262673228979111, + 0.005304192192852497, + 0.011092061176896095, + 0.006647316738963127, + 0.006669159047305584, + 0.006760682910680771, + 0.003422596724703908, + 0.0018027430633082986, + 0.002246766583994031, + 0.001813325914554298, + 0.0019844064954668283, + 0.0018311968306079507, + 0.0020651572849601507, + 0.0020005647093057632, + 0.004662582650780678, + 0.0026577492244541645, + 0.004628928378224373, + 0.0036612371914088726, + 0.0029568523168563843, + 0.003007252234965563, + 0.003742528147995472, + 0.003844413673505187, + 0.004877045284956694, + 0.005054708104580641, + 0.006620889995247126, + 0.006996133830398321, + 0.005865195300430059, + 0.007440340239554644, + 0.00682370038703084, + 0.006164408288896084, + 0.021063826978206635, + 0.017130864784121513, + 0.04205076023936272, + 0.027074849233031273, + 0.01456752885133028, + 0.013594222255051136, + 0.04236077517271042, + 0.025530247017741203, + 0.004728328436613083, + 0.005418852902948856, + 0.009462549351155758, + 0.006689006928354502, + 0.006563414353877306, + 0.004169411025941372, + 0.007570320274680853, + 0.0069658481515944, + 0.0177269596606493, + 0.007305868901312351, + 0.03656862676143646, + 0.016561971977353096, + 0.004706264473497868, + 0.003966209478676319, + 0.008602154441177845, + 0.03249849006533623, + 0.01530724298208952, + 0.00810740515589714, + 0.029116977006196976, + 0.012574802152812481, + 0.006329213734716177, + 0.003880572272464633, + 0.017944764345884323, + 0.011950640007853508, + 0.009797920472919941, + 0.011059467680752277, + 0.019630344584584236, + 0.015582921914756298, + 0.006696250755339861, + 0.0066249920055270195, + 0.016106439754366875, + 0.017014671117067337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0022019920870661736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic overhead shot reveals a massive lava insect rising from the utter darkness of the void. The creature's massive body is composed of a cracked obsidian carapace, a glow of incandescent magma seeping from deep within the cracks, as if the fury of the earth's core has been given life. Its translucent wings, like superheated amber, are covered in an intricate web of fiery veins and vibrate explosively at a rate too fast to be perceived by the naked eye. These frenzied movements instantly unleash a chaotic storm of countless glowing embers and delicate, swirling ash, resembling an inverted volcanic eruption against the pitch-black backdrop. With each violent flap of its wings, its massive form is propelled vertically upward by an unparalleled force, soaring into the sky. A strong and dramatic backlight originating from behind it completely engulfs the side of it facing the camera in deep shadows, but the sharp flames like a knife edge outline the contours of each of its jagged limbs and hideous head, forming the iconic and tense contrast of light and shadow of the Caravaggio school. The whole scene is filled with a primitive, violent and awe-inspiring doomsday aesthetic atmosphere.", + "seed": 24708, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_14/index_4148_seed_24708_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": -0.08820764189586043, + "vlm_logprob": -0.08820764189586043, + "vlm_yes_no_margin": 3.675000011920929, + "vlm_token_credit": -0.08820764189586043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021622398868203163 + ], + "opposite_logprob_tokens": [ + -3.8966224193573 + ], + "logprob_reward": -0.021622398868203163, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021622398868203163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17757317423820496, + 0.08249405771493912, + 0.22128434479236603, + 0.08315560966730118, + 0.11095120012760162, + 0.06697355955839157, + 0.07682278752326965, + 0.09876514226198196, + 0.039844658225774765, + 0.042135484516620636 + ], + "credit_map": [ + 0.02358461543917656, + 0.01796610653400421, + 0.019320936873555183, + 0.02408229000866413, + 0.026639945805072784, + 0.02434399165213108, + 0.016650905832648277, + 0.024984395131468773, + 0.010966419242322445, + 0.011772065423429012, + 0.0098066795617342, + 0.01190111879259348, + 0.010833864100277424, + 0.009369405917823315, + 0.009933932684361935, + 0.007910572923719883, + 0.043374937027692795, + 0.0240559633821249, + 0.027659859508275986, + 0.029149988666176796, + 0.01998448744416237, + 0.025041501969099045, + 0.019344523549079895, + 0.032673075795173645, + 0.01868293806910515, + 0.007180755492299795, + 0.009466188959777355, + 0.007046149577945471, + 0.006861966568976641, + 0.013587896712124348, + 0.006356390193104744, + 0.013973316177725792, + 0.008137180469930172, + 0.01239699125289917, + 0.015588738024234772, + 0.008442558348178864, + 0.018845371901988983, + 0.01871289126574993, + 0.011607079766690731, + 0.017220381647348404, + 0.005682106129825115, + 0.007334998808801174, + 0.010493086650967598, + 0.009806588292121887, + 0.004876646213233471, + 0.012999444268643856, + 0.005683604162186384, + 0.010097081772983074, + 0.008333807811141014, + 0.010098333470523357, + 0.010486838407814503, + 0.015760201960802078, + 0.006858911830931902, + 0.011433327570557594, + 0.006612993776798248, + 0.007238367106765509, + 0.008865753188729286, + 0.011019346304237843, + 0.011565301567316055, + 0.017364423722028732, + 0.007302462588995695, + 0.01383535098284483, + 0.014582447707653046, + 0.01423005759716034, + 0.007079815957695246, + 0.005178083665668964, + 0.0042318375781178474, + 0.0045446185395121574, + 0.003764706663787365, + 0.00553583400323987, + 0.0034979835618287325, + 0.006011779885739088, + 0.0066218809224665165, + 0.0051885973662137985, + 0.005582787562161684, + 0.00444859080016613, + 0.004418214317411184, + 0.00596984475851059, + 0.0026104021817445755, + 0.007295167539268732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.021622398868203163 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007648942526429892 + ], + "opposite_logprob_tokens": [ + -4.882648944854736 + ], + "logprob_reward": -0.007648942526429892, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007648942526429892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1666487157344818, + 0.12062480300664902, + 0.23787681758403778, + 0.04776529595255852, + 0.06596007943153381, + 0.09468474984169006, + 0.053464893251657486, + 0.13135921955108643, + 0.05675773322582245, + 0.024857696145772934 + ], + "credit_map": [ + 0.024650735780596733, + 0.026307428255677223, + 0.0174091886729002, + 0.030739733949303627, + 0.018077675253152847, + 0.01822356879711151, + 0.015583557076752186, + 0.015656806528568268, + 0.013680309057235718, + 0.020386090502142906, + 0.01703953742980957, + 0.01618952676653862, + 0.016017895191907883, + 0.01051195990294218, + 0.016474012285470963, + 0.010325455106794834, + 0.03957640007138252, + 0.039289046078920364, + 0.019464503973722458, + 0.02420783042907715, + 0.02512221783399582, + 0.02527008205652237, + 0.029485968872904778, + 0.03546074405312538, + 0.004963323473930359, + 0.0061112348921597, + 0.0052540237084031105, + 0.004564068745821714, + 0.005472214892506599, + 0.007834888994693756, + 0.0064979437738657, + 0.0070675951428711414, + 0.005056070163846016, + 0.005874818656593561, + 0.01140174362808466, + 0.005761368200182915, + 0.0063469079323112965, + 0.012356867082417011, + 0.007052768487483263, + 0.012109529227018356, + 0.009179364889860153, + 0.00820192415267229, + 0.018401069566607475, + 0.013498106971383095, + 0.009966091252863407, + 0.015937115997076035, + 0.005905453581362963, + 0.01359561737626791, + 0.004431018605828285, + 0.004012562800198793, + 0.008749970234930515, + 0.013068147003650665, + 0.004757287912070751, + 0.0075518288649618626, + 0.005219083745032549, + 0.005674988031387329, + 0.01219949685037136, + 0.013521254062652588, + 0.012995230033993721, + 0.016650108620524406, + 0.017579736188054085, + 0.014027786441147327, + 0.031319912523031235, + 0.013065686449408531, + 0.00679357722401619, + 0.007415924686938524, + 0.009415642358362675, + 0.005024708341807127, + 0.005795060656964779, + 0.007439369801431894, + 0.006412843707948923, + 0.008460601791739464, + 0.0036819034721702337, + 0.001941211405210197, + 0.003174801589921117, + 0.0021685101091861725, + 0.0028119932394474745, + 0.003928590100258589, + 0.0022384615149348974, + 0.004912223666906357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.007648942526429892 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606719382107258 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.01606719382107258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606719382107258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1487649828195572, + 0.0985679030418396, + 0.12435850501060486, + 0.05375958979129791, + 0.07502825558185577, + 0.26555442810058594, + 0.064704529941082, + 0.08709924668073654, + 0.03390491008758545, + 0.04825761914253235 + ], + "credit_map": [ + 0.024944938719272614, + 0.014332120306789875, + 0.016643892973661423, + 0.020650872960686684, + 0.022234749048948288, + 0.02057351917028427, + 0.014177992939949036, + 0.015206903219223022, + 0.020885957404971123, + 0.012112857773900032, + 0.024326499551534653, + 0.01207665540277958, + 0.006033735815435648, + 0.007097794208675623, + 0.006603498477488756, + 0.009430903010070324, + 0.023797348141670227, + 0.020719287917017937, + 0.013939548283815384, + 0.00986102968454361, + 0.01131463423371315, + 0.012156290002167225, + 0.012603572569787502, + 0.01996680535376072, + 0.010059764608740807, + 0.006468606647104025, + 0.004850730765610933, + 0.005016005132347345, + 0.004813775885850191, + 0.006759892217814922, + 0.006777005735784769, + 0.009013812057673931, + 0.007064740639179945, + 0.00862734206020832, + 0.013825510628521442, + 0.007101680152118206, + 0.008701794780790806, + 0.011862593702971935, + 0.007217572536319494, + 0.010627026669681072, + 0.02372434176504612, + 0.02145649679005146, + 0.04791292920708656, + 0.030977357178926468, + 0.026906590908765793, + 0.0439336895942688, + 0.029064634814858437, + 0.041578397154808044, + 0.005461693275719881, + 0.004537417087703943, + 0.012532101944088936, + 0.008647513575851917, + 0.005492433439940214, + 0.010482335463166237, + 0.006206895224750042, + 0.011344144120812416, + 0.007431276608258486, + 0.010479232296347618, + 0.008852464146912098, + 0.013222298584878445, + 0.01121152937412262, + 0.01171288825571537, + 0.015296733938157558, + 0.008892831392586231, + 0.0064956871792674065, + 0.0027685523964464664, + 0.0072227115742862225, + 0.0039568254724144936, + 0.0028860585298389196, + 0.0041412608698010445, + 0.0030282801017165184, + 0.003405533265322447, + 0.005742812994867563, + 0.006752456538379192, + 0.006474472116678953, + 0.005796222016215324, + 0.0061067393980920315, + 0.006013123784214258, + 0.004423430189490318, + 0.006948363967239857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.01606719382107258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38705915212631226 + ], + "opposite_logprob_tokens": [ + -1.137059211730957 + ], + "logprob_reward": -0.38705915212631226, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38705915212631226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4173056185245514, + 0.12177274376153946, + 0.07191707193851471, + 0.0449853278696537, + 0.06756138801574707, + 0.06413247436285019, + 0.05317099764943123, + 0.06526404619216919, + 0.05893712118268013, + 0.03495318815112114 + ], + "credit_map": [ + 0.06423252075910568, + 0.050881486386060715, + 0.04643835127353668, + 0.05669998750090599, + 0.05947618931531906, + 0.05758477374911308, + 0.03296162188053131, + 0.049030717462301254, + 0.033080026507377625, + 0.011738882400095463, + 0.029845699667930603, + 0.008552728220820427, + 0.008732105605304241, + 0.007549665402621031, + 0.006675638724118471, + 0.015598009340465069, + 0.012449119240045547, + 0.008909507654607296, + 0.009172416292130947, + 0.008266358636319637, + 0.005908262450248003, + 0.010219876654446125, + 0.007946113124489784, + 0.009045426733791828, + 0.011169140227138996, + 0.004064507782459259, + 0.0047536310739815235, + 0.003163988934829831, + 0.00353394728153944, + 0.005941536743193865, + 0.004855460487306118, + 0.007503121625632048, + 0.00600765785202384, + 0.0072108604945242405, + 0.010359212756156921, + 0.006528458558022976, + 0.010482166893780231, + 0.010342992842197418, + 0.006436266470700502, + 0.010193785652518272, + 0.006912661250680685, + 0.007251407019793987, + 0.01114091370254755, + 0.009839379228651524, + 0.005567797925323248, + 0.011891478672623634, + 0.005111327860504389, + 0.006417518947273493, + 0.009386667050421238, + 0.005748427007347345, + 0.008423808962106705, + 0.00630910973995924, + 0.0040657599456608295, + 0.007210473995655775, + 0.0045378017239272594, + 0.007488952949643135, + 0.008679374121129513, + 0.008706485852599144, + 0.009723605588078499, + 0.010828842408955097, + 0.006036641541868448, + 0.008903945796191692, + 0.007391058839857578, + 0.004994092974811792, + 0.010169943794608116, + 0.005612908862531185, + 0.009162032045423985, + 0.0070372992195189, + 0.005020685959607363, + 0.007764322217553854, + 0.005212009884417057, + 0.008957924321293831, + 0.004794994834810495, + 0.004635767079889774, + 0.003775140969082713, + 0.003783751279115677, + 0.007264447398483753, + 0.003958519082516432, + 0.002903142012655735, + 0.0038374289870262146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.38705915212631226 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640522137284279 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640522137284279, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640522137284279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2581300735473633, + 0.1361169070005417, + 0.0651090070605278, + 0.06677073240280151, + 0.07467099279165268, + 0.10036981105804443, + 0.12455109506845474, + 0.08793444186449051, + 0.0547439344227314, + 0.03160298615694046 + ], + "credit_map": [ + 0.04062080755829811, + 0.024878237396478653, + 0.026008253917098045, + 0.028070397675037384, + 0.039445698261260986, + 0.04389864206314087, + 0.025620492175221443, + 0.029587525874376297, + 0.03322099149227142, + 0.008622630499303341, + 0.03814331442117691, + 0.015375914052128792, + 0.006721963174641132, + 0.010190857574343681, + 0.008695375174283981, + 0.015145853161811829, + 0.011666364036500454, + 0.007929451763629913, + 0.008573695085942745, + 0.00630468875169754, + 0.006873456761240959, + 0.007779083214700222, + 0.006755081936717033, + 0.009227185510098934, + 0.009648915380239487, + 0.007940254174172878, + 0.008598923683166504, + 0.006014550570398569, + 0.00631945813074708, + 0.009849167428910732, + 0.00822379533201456, + 0.010175665840506554, + 0.005307240877300501, + 0.00604627700522542, + 0.01280066929757595, + 0.006879372056573629, + 0.008196813054382801, + 0.012655708007514477, + 0.007966611534357071, + 0.014818300493061543, + 0.008771468885242939, + 0.00804541353136301, + 0.022246958687901497, + 0.010747738182544708, + 0.007282764185220003, + 0.018096309155225754, + 0.011208870448172092, + 0.013970281928777695, + 0.010261994786560535, + 0.008528675884008408, + 0.022168897092342377, + 0.02883034572005272, + 0.010690581053495407, + 0.019959047436714172, + 0.008742500096559525, + 0.015369058586657047, + 0.01163124293088913, + 0.01137636974453926, + 0.009668789803981781, + 0.012090172618627548, + 0.009395371191203594, + 0.010952329263091087, + 0.014885080978274345, + 0.007935088127851486, + 0.008472444489598274, + 0.0061415620148181915, + 0.011812183074653149, + 0.006262189708650112, + 0.004524394404143095, + 0.006933039985597134, + 0.0049070813693106174, + 0.005691036581993103, + 0.00396697549149394, + 0.0026137162931263447, + 0.004583188332617283, + 0.0034187338314950466, + 0.0035089319571852684, + 0.004287790507078171, + 0.0032472219318151474, + 0.005976428743451834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.008640522137284279 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic overhead shot reveals a massive lava insect rising from the utter darkness of the void. The creature's massive body is composed of a cracked obsidian carapace, a glow of incandescent magma seeping from deep within the cracks, as if the fury of the earth's core has been given life. Its translucent wings, like superheated amber, are covered in an intricate web of fiery veins and vibrate explosively at a rate too fast to be perceived by the naked eye. These frenzied movements instantly unleash a chaotic storm of countless glowing embers and delicate, swirling ash, resembling an inverted volcanic eruption against the pitch-black backdrop. With each violent flap of its wings, its massive form is propelled vertically upward by an unparalleled force, soaring into the sky. A strong and dramatic backlight originating from behind it completely engulfs the side of it facing the camera in deep shadows, but the sharp flames like a knife edge outline the contours of each of its jagged limbs and hideous head, forming the iconic and tense contrast of light and shadow of the Caravaggio school. The whole scene is filled with a primitive, violent and awe-inspiring doomsday aesthetic atmosphere.", + "seed": 24709, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_14/index_4148_seed_24709_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": -0.036866307631134985, + "vlm_logprob": -0.036866307631134985, + "vlm_yes_no_margin": 4.149999976158142, + "vlm_token_credit": -0.036866307631134985, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014060367830097675 + ], + "opposite_logprob_tokens": [ + -4.3890604972839355 + ], + "logprob_reward": -0.014060367830097675, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014060367830097675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020991818979382515, + 0.025845156982541084, + 0.09779911488294601, + 0.018798213452100754, + 0.06455224752426147, + 0.05296929180622101, + 0.013538708910346031, + 0.01717359386384487, + 0.019456177949905396, + 0.6688756942749023 + ], + "credit_map": [ + 0.002260883804410696, + 0.0026749896351248026, + 0.0022544702515006065, + 0.0026329513639211655, + 0.003353454638272524, + 0.002353283576667309, + 0.0031100125052034855, + 0.002351772505789995, + 0.002401885576546192, + 0.002831678604707122, + 0.0030841960106045008, + 0.0032430507708340883, + 0.003586546052247286, + 0.004438992124050856, + 0.003517524339258671, + 0.002741283504292369, + 0.00877358764410019, + 0.015437675639986992, + 0.02342226542532444, + 0.009042825549840927, + 0.006423516198992729, + 0.006780728232115507, + 0.015470948070287704, + 0.012447568587958813, + 0.0020881604868918657, + 0.001457098638638854, + 0.001805326552130282, + 0.0018968437798321247, + 0.0019346659537404776, + 0.0017536379164084792, + 0.004906435031443834, + 0.002956046722829342, + 0.006818258203566074, + 0.0109708858653903, + 0.007574940100312233, + 0.008401094004511833, + 0.006903297733515501, + 0.00624696072191, + 0.008730745874345303, + 0.008906067349016666, + 0.0063001420348882675, + 0.004569394048303366, + 0.00521090067923069, + 0.004053291864693165, + 0.005820378195494413, + 0.008715628646314144, + 0.009611117653548717, + 0.008688440546393394, + 0.0018495135009288788, + 0.0011118004331365228, + 0.0018298577051609755, + 0.0015952661633491516, + 0.0011732053244486451, + 0.0018412010977044702, + 0.0030168197117745876, + 0.0011210446245968342, + 0.0012606671079993248, + 0.0017512054182589054, + 0.0027857525274157524, + 0.0030303476378321648, + 0.0023190034553408623, + 0.0026689006481319666, + 0.0019780672155320644, + 0.0013796486891806126, + 0.0018295897170901299, + 0.001560900593176484, + 0.0028947722166776657, + 0.0014300047187134624, + 0.0035323298070579767, + 0.00340062053874135, + 0.0022549722343683243, + 0.0025529898703098297, + 0.027327148243784904, + 0.0071365246549248695, + 0.2863163948059082, + 0.019710224121809006, + 0.014190668240189552, + 0.2954418957233429, + 0.008929284289479256, + 0.009823566302657127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.014060367830097675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006760344374924898 + ], + "opposite_logprob_tokens": [ + -5.006760120391846 + ], + "logprob_reward": -0.006760344374924898, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006760344374924898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04760563373565674, + 0.02624134160578251, + 0.054989565163850784, + 0.04159516096115112, + 0.028316786512732506, + 0.020400889217853546, + 0.016837457194924355, + 0.021324966102838516, + 0.05066307261586189, + 0.6920250654220581 + ], + "credit_map": [ + 0.004165214486420155, + 0.005313468631356955, + 0.005168503616005182, + 0.0060499622486531734, + 0.0062282877042889595, + 0.008211789652705193, + 0.007637752685695887, + 0.004830653313547373, + 0.00216468027792871, + 0.0026592803187668324, + 0.00288023566827178, + 0.0022596397902816534, + 0.0035766151268035173, + 0.004456078168004751, + 0.004022929351776838, + 0.004221879877150059, + 0.005417959298938513, + 0.00461565749719739, + 0.00853719562292099, + 0.0039059158880263567, + 0.005331889726221561, + 0.00439411262050271, + 0.010562445968389511, + 0.01222438644617796, + 0.005234945099800825, + 0.003397530410438776, + 0.004359133075922728, + 0.0046234289184212685, + 0.005355129949748516, + 0.0036507081240415573, + 0.008663476444780827, + 0.0063108098693192005, + 0.0030026191379874945, + 0.0051637738943099976, + 0.00298503041267395, + 0.0032179364934563637, + 0.003091193735599518, + 0.0033153435215353966, + 0.003716328414157033, + 0.0038245581090450287, + 0.0017329708207398653, + 0.0014996521640568972, + 0.004364574793726206, + 0.001555148628540337, + 0.0018137795850634575, + 0.0022137199994176626, + 0.0041884807869791985, + 0.003032562555745244, + 0.002216651802882552, + 0.0013050802517682314, + 0.002388471271842718, + 0.0018042135052382946, + 0.0015221344074234366, + 0.003216541837900877, + 0.002403504913672805, + 0.0019808595534414053, + 0.0021097520366311073, + 0.0019160741940140724, + 0.0036955580580979586, + 0.00327099347487092, + 0.0024910597130656242, + 0.0028228051960468292, + 0.002426487859338522, + 0.002592236502096057, + 0.0032223821617662907, + 0.005676900967955589, + 0.00769719248637557, + 0.004431955516338348, + 0.008181472308933735, + 0.008703887462615967, + 0.006359222810715437, + 0.006390062160789967, + 0.02841927483677864, + 0.009663236327469349, + 0.29164087772369385, + 0.022721722722053528, + 0.015160604380071163, + 0.30004504323005676, + 0.012090397998690605, + 0.012283939868211746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.006760344374924898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072005033493042 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016072005033493042, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072005033493042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1705787181854248, + 0.08276966959238052, + 0.1677660495042801, + 0.06144297122955322, + 0.16078463196754456, + 0.12528355419635773, + 0.04168812558054924, + 0.032349102199077606, + 0.05458012968301773, + 0.1027570515871048 + ], + "credit_map": [ + 0.017538247630000114, + 0.015860915184020996, + 0.02038084901869297, + 0.020422225818037987, + 0.01903439126908779, + 0.031332213431596756, + 0.026056068018078804, + 0.019953791052103043, + 0.007052375003695488, + 0.00991392694413662, + 0.008980095386505127, + 0.01164910476654768, + 0.007828096859157085, + 0.011450838297605515, + 0.012532815337181091, + 0.013362418860197067, + 0.01748022437095642, + 0.02026430331170559, + 0.017351265996694565, + 0.012901181355118752, + 0.017053384333848953, + 0.015291528776288033, + 0.03032231703400612, + 0.03710184246301651, + 0.009186873212456703, + 0.00571875786408782, + 0.005978568457067013, + 0.0051220739260315895, + 0.00736947963014245, + 0.007207344751805067, + 0.011418760754168034, + 0.009441114030778408, + 0.01512837689369917, + 0.02473105676472187, + 0.01930522732436657, + 0.01904677413403988, + 0.018484722822904587, + 0.015474290587008, + 0.027481762692332268, + 0.02113242633640766, + 0.011974340304732323, + 0.0114855682477355, + 0.018471023067831993, + 0.011897198855876923, + 0.015429843217134476, + 0.013084081001579762, + 0.025279153138399124, + 0.017662350088357925, + 0.004685726948082447, + 0.0034343518782407045, + 0.005988176446408033, + 0.004771227017045021, + 0.004687158390879631, + 0.007224568631500006, + 0.006796123459935188, + 0.004100791644304991, + 0.002600882202386856, + 0.0023799871560186148, + 0.005870071239769459, + 0.003905082121491432, + 0.0033364570699632168, + 0.006691075395792723, + 0.002614927012473345, + 0.004950619302690029, + 0.0038146390579640865, + 0.005778416525572538, + 0.007308317348361015, + 0.004087867680937052, + 0.01080712303519249, + 0.008131607435643673, + 0.007469254545867443, + 0.007182904053479433, + 0.006207514088600874, + 0.008168661035597324, + 0.024103950709104538, + 0.010222524404525757, + 0.008003460243344307, + 0.02487110160291195, + 0.014841761440038681, + 0.006338079459965229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.016072005033493042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279663562774658 + ], + "opposite_logprob_tokens": [ + -2.017796516418457 + ], + "logprob_reward": -0.14279663562774658, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14279663562774658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11702308058738708, + 0.04061112552881241, + 0.049288347363471985, + 0.06846783310174942, + 0.1286475956439972, + 0.16298730671405792, + 0.012553432025015354, + 0.01925353892147541, + 0.0270970668643713, + 0.3740707039833069 + ], + "credit_map": [ + 0.012532740831375122, + 0.01571628265082836, + 0.01207571942359209, + 0.01395169273018837, + 0.011297367513179779, + 0.021590132266283035, + 0.017560720443725586, + 0.01229842472821474, + 0.0040186429396271706, + 0.004456501919776201, + 0.003099302528426051, + 0.0038413500878959894, + 0.005488282535225153, + 0.007688101381063461, + 0.006574530154466629, + 0.005444411188364029, + 0.005672135390341282, + 0.005945862270891666, + 0.003955680411309004, + 0.0039636255241930485, + 0.0067351218312978745, + 0.0058922721073031425, + 0.0074411071836948395, + 0.009682544507086277, + 0.008479399606585503, + 0.005693665239959955, + 0.0059667062014341354, + 0.006239046808332205, + 0.007961404509842396, + 0.005390281323343515, + 0.017519140616059303, + 0.011218181811273098, + 0.011316156946122646, + 0.012624582275748253, + 0.016540804877877235, + 0.009883832186460495, + 0.015874691307544708, + 0.017808588221669197, + 0.02727346308529377, + 0.017325466498732567, + 0.015907229855656624, + 0.015297768637537956, + 0.02071654610335827, + 0.018805591389536858, + 0.02229376696050167, + 0.01564265415072441, + 0.0316145233809948, + 0.022709231823682785, + 0.0012723889667540789, + 0.0008352978620678186, + 0.00193305523134768, + 0.0011079238029196858, + 0.0011945079313591123, + 0.0023322496563196182, + 0.0022919680923223495, + 0.001586040947586298, + 0.0013606647262349725, + 0.001383992494083941, + 0.003237163880839944, + 0.0029950125608593225, + 0.0022757232654839754, + 0.0034268905874341726, + 0.0024084202013909817, + 0.002165669808164239, + 0.0023432085290551186, + 0.0024177334271371365, + 0.003987875301390886, + 0.0028094027657061815, + 0.004041917156428099, + 0.004319016356021166, + 0.0036538185086101294, + 0.0035240952856838703, + 0.015573042444884777, + 0.00484283734112978, + 0.15888611972332, + 0.011461971327662468, + 0.008129476569592953, + 0.16398543119430542, + 0.005667588673532009, + 0.005524241365492344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.14279663562774658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 insect-like creature with a cracked obsidian carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have translucent amber wings with fiery veins that vibrate explosively?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature propelled vertically upward, unleashing a storm of glowing embers and swirling ash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a dramatic overhead angle against a pitch-black void backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight creating deep shadows on the creature while outlining its jagged limbs with sharp flames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642185289412737 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004642185289412737, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642185289412737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052265193313360214, + 0.02981475554406643, + 0.03127725422382355, + 0.017319681122899055, + 0.06482909619808197, + 0.06098093092441559, + 0.014289354905486107, + 0.017863720655441284, + 0.03681159391999245, + 0.6745483875274658 + ], + "credit_map": [ + 0.0052276370115578175, + 0.008023136295378208, + 0.005096237640827894, + 0.0056963819079101086, + 0.007982966490089893, + 0.00839819386601448, + 0.0067499312572181225, + 0.005090710241347551, + 0.0023007639683783054, + 0.002783349249511957, + 0.003740291576832533, + 0.0032927561551332474, + 0.004253734368830919, + 0.005401605740189552, + 0.004500600043684244, + 0.003541655372828245, + 0.0036750687286257744, + 0.0030452178325504065, + 0.004553269594907761, + 0.0033777705393731594, + 0.0034863492473959923, + 0.0035165795125067234, + 0.005327129270881414, + 0.004295869264751673, + 0.0019873210694640875, + 0.00143669662065804, + 0.0014509271131828427, + 0.0019959162455052137, + 0.0016615679487586021, + 0.00163859105668962, + 0.004275408107787371, + 0.0028732535429298878, + 0.005620334763079882, + 0.009530560113489628, + 0.007629703730344772, + 0.0071160318329930305, + 0.008047780953347683, + 0.00794270634651184, + 0.010518921539187431, + 0.008423058316111565, + 0.006616655737161636, + 0.005557679571211338, + 0.00705915130674839, + 0.0070229568518698215, + 0.008240869268774986, + 0.00791608914732933, + 0.009833650663495064, + 0.008733878843486309, + 0.001480304286815226, + 0.0013548200950026512, + 0.001954209990799427, + 0.0016333158127963543, + 0.0025030001997947693, + 0.0018464571330696344, + 0.0020479527302086353, + 0.0014692951226606965, + 0.0012591226259246469, + 0.001452415599487722, + 0.0032856299076229334, + 0.0025819765869528055, + 0.002604435430839658, + 0.0027841529808938503, + 0.0018066349439322948, + 0.0020893521141260862, + 0.002029027557000518, + 0.0025692901108413935, + 0.006033387966454029, + 0.0034360832069069147, + 0.006937748286873102, + 0.006644661072641611, + 0.005602905061095953, + 0.003558490192517638, + 0.027990994974970818, + 0.007530330680310726, + 0.28794461488723755, + 0.01970653049647808, + 0.014191558584570885, + 0.2971993386745453, + 0.009173174388706684, + 0.010811849497258663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.004642185289412737 + } + ], + "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/0000128/group_15/index_tensor([5825, 5825])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_15/index_tensor([5825, 5825])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e1eb02a523dbd33d150b7739078c5e43ee004fcb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_15/index_tensor([5825, 5825])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera captures a steady, first-person perspective from the cockpit of a deep-sea submersible, drifting silently and deliberately forward through the darkness of the abyss. The only interior illumination comes from the soft, cool glow of the digital instrument panel, its blue and green readouts and blinking indicator lights reflected faintly on the thickly reinforced circular observation area. Outside, the submersible's powerful cone of cool white light cuts through the inky water, revealing a vast alien landscape of the seafloor below, covered in dense, shimmering, circular metallic nodules that gleam under the artificial light. Ethereal bioluminescence, resembling ghostly jellyfish and swarms of tiny, glowing specks, drifts gracefully across the observation area, like a slow-motion starfield, their soft cyan and white glow gently pulsing through the oppressive depths. In the distance, a massive, chimney-like hydrothermal vent looms, continuously spewing thick, opaque black smoke that billows upward into the shattered void. The entire scene is rendered in a distinctly dark anime style, characterized by sharp, hard-edged shadows, high contrast, and a moody blue-black color palette, creating an atmosphere of isolated wonder and eerie, solemn silence.", + "seed": 21365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_15/index_5825_seed_21365_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": -0.10878133177757263, + "vlm_logprob": -0.10878133177757263, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.10878133177757263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20707155764102936 + ], + "opposite_logprob_tokens": [ + -1.7070715427398682 + ], + "logprob_reward": -0.20707155764102936, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20707155764102936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2098700851202011, + 0.07664064317941666, + 0.012428284622728825, + 0.31638720631599426, + 0.20820006728172302, + 0.11670370399951935, + 0.012120986357331276, + 0.019257275387644768, + 0.011260582134127617, + 0.017131194472312927 + ], + "credit_map": [ + 0.015581854619085789, + 0.00897952914237976, + 0.04458151385188103, + 0.02882956713438034, + 0.0117030693218112, + 0.0567183643579483, + 0.02102949097752571, + 0.022446680814027786, + 0.005110582802444696, + 0.008374233730137348, + 0.007651223801076412, + 0.012758596800267696, + 0.011446555145084858, + 0.006893751677125692, + 0.01831541396677494, + 0.0060902866534888744, + 0.001309619634412229, + 0.0022068500984460115, + 0.0012623023940250278, + 0.0017217963468283415, + 0.0017771522980183363, + 0.0016554944450035691, + 0.0009471793891862035, + 0.001547889318317175, + 0.03756418451666832, + 0.022511430084705353, + 0.020986376330256462, + 0.01541480515152216, + 0.04231078550219536, + 0.12465064972639084, + 0.023913653567433357, + 0.029035314917564392, + 0.020902691408991814, + 0.02188018336892128, + 0.011712295934557915, + 0.012697556987404823, + 0.033541809767484665, + 0.061192914843559265, + 0.025368932634592056, + 0.02090367116034031, + 0.011056198738515377, + 0.019278554245829582, + 0.008541839197278023, + 0.009364661760628223, + 0.021623775362968445, + 0.017765386030077934, + 0.01440972089767456, + 0.014663568697869778, + 0.0010124968830496073, + 0.0018517961725592613, + 0.0009933513356372714, + 0.0012487585190683603, + 0.0020899996161460876, + 0.002390461741015315, + 0.0013101999647915363, + 0.001223921775817871, + 0.0020730257965624332, + 0.0020500256214290857, + 0.001354419975541532, + 0.0012506676139310002, + 0.004141110926866531, + 0.00450202077627182, + 0.0016274783993139863, + 0.002258527558296919, + 0.0008226491627283394, + 0.0011910421308130026, + 0.001198334968648851, + 0.0012279751244932413, + 0.0021008714102208614, + 0.0026047644205391407, + 0.0012338791275396943, + 0.0008810656727291644, + 0.0014115750091150403, + 0.0015958399744704366, + 0.0021244233939796686, + 0.002247583121061325, + 0.003098276676610112, + 0.0021316721104085445, + 0.0027027372270822525, + 0.0018190875416621566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.20707155764102936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372933343052864 + ], + "opposite_logprob_tokens": [ + -3.408729314804077 + ], + "logprob_reward": -0.03372933343052864, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372933343052864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10302568227052689, + 0.21899661421775818, + 0.024732301011681557, + 0.29378464818000793, + 0.13643404841423035, + 0.15607783198356628, + 0.02699059247970581, + 0.016566362231969833, + 0.012182348407804966, + 0.011209583841264248 + ], + "credit_map": [ + 0.010226082988083363, + 0.006333987694233656, + 0.021588262170553207, + 0.01063084602355957, + 0.00776336994022131, + 0.027472157031297684, + 0.009298987686634064, + 0.009711985476315022, + 0.015371516346931458, + 0.020962359383702278, + 0.023409994319081306, + 0.03979726880788803, + 0.029679007828235626, + 0.019639836624264717, + 0.04939467832446098, + 0.020741961896419525, + 0.00276273675262928, + 0.0033070845529437065, + 0.0026471896562725306, + 0.003270884044468403, + 0.0026145586743950844, + 0.0026509067974984646, + 0.002356264740228653, + 0.00512267742305994, + 0.03431185334920883, + 0.022699441760778427, + 0.019350741058588028, + 0.014344719238579273, + 0.03867807239294052, + 0.11468556523323059, + 0.022309590131044388, + 0.027404645457863808, + 0.013143311254680157, + 0.014996211044490337, + 0.0080726845189929, + 0.008624167181551456, + 0.02268882468342781, + 0.038062479346990585, + 0.017090585082769394, + 0.01375578437000513, + 0.016358382999897003, + 0.024078914895653725, + 0.011061517521739006, + 0.01188065204769373, + 0.03033366985619068, + 0.025688933208584785, + 0.019857462495565414, + 0.01681830734014511, + 0.002490806393325329, + 0.005435593891888857, + 0.0018965578638017178, + 0.0021766687277704477, + 0.004477243404835463, + 0.004893037956207991, + 0.002294740406796336, + 0.0033259433694183826, + 0.001473685959354043, + 0.0023173070512712, + 0.0015661874786019325, + 0.0018934676190838218, + 0.002904994646087289, + 0.0031180749647319317, + 0.0017595967510715127, + 0.0015330484602600336, + 0.0012223002268001437, + 0.001379060442559421, + 0.0011748762335628271, + 0.0011229842202737927, + 0.002322571584954858, + 0.002955725183710456, + 0.0010511184809729457, + 0.0009537127916701138, + 0.0010429329704493284, + 0.0013883705250918865, + 0.0012424129527062178, + 0.001391747733578086, + 0.0018097006250172853, + 0.0017800208879634738, + 0.0014880168018862605, + 0.001066381111741066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.03372933343052864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541946023702621 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541946023702621, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541946023702621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04481106996536255, + 0.02785012125968933, + 0.029837608337402344, + 0.3579625189304352, + 0.27039122581481934, + 0.05008068308234215, + 0.023418648168444633, + 0.1012473925948143, + 0.048644036054611206, + 0.04575667902827263 + ], + "credit_map": [ + 0.004361303523182869, + 0.0051596821285784245, + 0.005258653312921524, + 0.012783975340425968, + 0.0036077462136745453, + 0.004164723213762045, + 0.006746700033545494, + 0.002728287596255541, + 0.002409669803455472, + 0.0037130999844521284, + 0.0036740584764629602, + 0.004655883647501469, + 0.003665188793092966, + 0.00258033093996346, + 0.00369667517952621, + 0.003455214435234666, + 0.002421292942017317, + 0.0045780460350215435, + 0.0033166261855512857, + 0.004349445924162865, + 0.004537165630608797, + 0.003492790972813964, + 0.0031616350170224905, + 0.003980605397373438, + 0.04239500313997269, + 0.025966394692659378, + 0.024232657626271248, + 0.017127998173236847, + 0.048335108906030655, + 0.1404918134212494, + 0.02700822614133358, + 0.032405342906713486, + 0.028380345553159714, + 0.027137381955981255, + 0.014765352010726929, + 0.015873169526457787, + 0.042027149349451065, + 0.08393799513578415, + 0.030415162444114685, + 0.027854688465595245, + 0.0047640264965593815, + 0.0055157835595309734, + 0.002936022821813822, + 0.004032950382679701, + 0.009354076348245144, + 0.010078015737235546, + 0.0045056939125061035, + 0.008894111961126328, + 0.0017919030506163836, + 0.0028538082260638475, + 0.002206706441938877, + 0.0022093511652201414, + 0.003994700498878956, + 0.0042708199471235275, + 0.0022493808064609766, + 0.003841976635158062, + 0.007655061315745115, + 0.01640779711306095, + 0.008670120500028133, + 0.011714925989508629, + 0.01807219535112381, + 0.02108229510486126, + 0.008327883668243885, + 0.009317120537161827, + 0.00289068091660738, + 0.006236258428543806, + 0.0061622136272490025, + 0.008938325569033623, + 0.007343713194131851, + 0.0063148862682282925, + 0.005274047609418631, + 0.005483912769705057, + 0.002787365112453699, + 0.00525215920060873, + 0.004303416237235069, + 0.00793993379920721, + 0.007620455697178841, + 0.006912687327712774, + 0.0061807213351130486, + 0.004759938921779394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.012541946023702621 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153886079788208 + ], + "opposite_logprob_tokens": [ + -1.7015388011932373 + ], + "logprob_reward": -0.20153886079788208, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153886079788208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07279656827449799, + 0.03372480347752571, + 0.03890995308756828, + 0.06091231107711792, + 0.5397917032241821, + 0.08286357671022415, + 0.04663897678256035, + 0.060581956058740616, + 0.029294952750205994, + 0.03448524326086044 + ], + "credit_map": [ + 0.006302161607891321, + 0.005048045422881842, + 0.015010305680334568, + 0.005539365112781525, + 0.0036465690936893225, + 0.021870162338018417, + 0.006574444007128477, + 0.008805505000054836, + 0.0025015727151185274, + 0.003741219872608781, + 0.003692048601806164, + 0.006193930748850107, + 0.005281451623886824, + 0.0031124441884458065, + 0.006434783339500427, + 0.0027673514559865, + 0.0030637227464467287, + 0.006507375743240118, + 0.004148445557802916, + 0.007782268803566694, + 0.00450182193890214, + 0.0036219325847923756, + 0.0036564460024237633, + 0.005627936217933893, + 0.006347088608890772, + 0.010169213637709618, + 0.005267794243991375, + 0.005318229552358389, + 0.008537483401596546, + 0.015819283202290535, + 0.005159399472177029, + 0.004293814301490784, + 0.04812774434685707, + 0.06947958469390869, + 0.03303765133023262, + 0.04002279415726662, + 0.09710655361413956, + 0.13049986958503723, + 0.06734101474285126, + 0.05417648330330849, + 0.006449204869568348, + 0.01803426258265972, + 0.004695102572441101, + 0.007520936895161867, + 0.012285908684134483, + 0.011934635229408741, + 0.006664312444627285, + 0.015279201790690422, + 0.004312905482947826, + 0.009484762325882912, + 0.0028711454942822456, + 0.004111553076654673, + 0.009211422875523567, + 0.007102808449417353, + 0.004374088253825903, + 0.005170290358364582, + 0.006244069896638393, + 0.006910966709256172, + 0.004978587850928307, + 0.006078658625483513, + 0.011303049512207508, + 0.012016107328236103, + 0.005716164596378803, + 0.007334343623369932, + 0.0024040855932980776, + 0.0035303288605064154, + 0.0026588861364871264, + 0.003502308391034603, + 0.006084507331252098, + 0.005688606761395931, + 0.003073767526075244, + 0.002352460753172636, + 0.0022768229246139526, + 0.0038834617007523775, + 0.0030129540245980024, + 0.005130114499479532, + 0.006527951452881098, + 0.007950604893267155, + 0.0029492396861314774, + 0.0027540880255401134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.20153886079788208 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902496099472046 + ], + "opposite_logprob_tokens": [ + -2.4640250205993652 + ], + "logprob_reward": -0.08902496099472046, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902496099472046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13027353584766388, + 0.03189825639128685, + 0.04268182814121246, + 0.5034262537956238, + 0.13109587132930756, + 0.07583211362361908, + 0.020316321402788162, + 0.030903873965144157, + 0.01878722943365574, + 0.014784742146730423 + ], + "credit_map": [ + 0.009851864539086819, + 0.009599362500011921, + 0.025762030854821205, + 0.022740621119737625, + 0.006575231906026602, + 0.029303856194019318, + 0.012284553609788418, + 0.014156014658510685, + 0.0026589627377688885, + 0.00360956322401762, + 0.003680153051391244, + 0.004302332643419504, + 0.0051368046551942825, + 0.003331109881401062, + 0.006280042231082916, + 0.00289928843267262, + 0.004637979436665773, + 0.006189218256622553, + 0.003751372918486595, + 0.007015974726527929, + 0.00625503109768033, + 0.005071808584034443, + 0.003968897741287947, + 0.005791546776890755, + 0.05958123505115509, + 0.038416627794504166, + 0.03429591655731201, + 0.02477213181555271, + 0.06711069494485855, + 0.19631268084049225, + 0.03743036091327667, + 0.045506589114665985, + 0.011246143840253353, + 0.017307143658399582, + 0.008026397787034512, + 0.0092834597453475, + 0.023146890103816986, + 0.03227425366640091, + 0.017715655267238617, + 0.012095929123461246, + 0.004299127031117678, + 0.014769488014280796, + 0.005634639412164688, + 0.006462492048740387, + 0.013914096169173717, + 0.009702169336378574, + 0.010192951187491417, + 0.010857143439352512, + 0.00141730101313442, + 0.0031917113810777664, + 0.0016175330383703113, + 0.002001379616558552, + 0.0031859276350587606, + 0.003070094622671604, + 0.0022505198139697313, + 0.0035818538162857294, + 0.0028340257704257965, + 0.0033464108128100634, + 0.0021961552556604147, + 0.002419872209429741, + 0.006201086100190878, + 0.00682483147829771, + 0.0025233146734535694, + 0.004558175336569548, + 0.0014278951566666365, + 0.0036135322879999876, + 0.0018086953787133098, + 0.0021746132988482714, + 0.0035449215210974216, + 0.0030738769564777613, + 0.001768306246958673, + 0.0013753884704783559, + 0.0011227113427594304, + 0.0021941098384559155, + 0.0014856386696919799, + 0.0020178621634840965, + 0.002960707526654005, + 0.002075864700600505, + 0.0016990401782095432, + 0.0012288079597055912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.08902496099472046 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera captures a steady, first-person perspective from the cockpit of a deep-sea submersible, drifting silently and deliberately forward through the darkness of the abyss. The only interior illumination comes from the soft, cool glow of the digital instrument panel, its blue and green readouts and blinking indicator lights reflected faintly on the thickly reinforced circular observation area. Outside, the submersible's powerful cone of cool white light cuts through the inky water, revealing a vast alien landscape of the seafloor below, covered in dense, shimmering, circular metallic nodules that gleam under the artificial light. Ethereal bioluminescence, resembling ghostly jellyfish and swarms of tiny, glowing specks, drifts gracefully across the observation area, like a slow-motion starfield, their soft cyan and white glow gently pulsing through the oppressive depths. In the distance, a massive, chimney-like hydrothermal vent looms, continuously spewing thick, opaque black smoke that billows upward into the shattered void. The entire scene is rendered in a distinctly dark anime style, characterized by sharp, hard-edged shadows, high contrast, and a moody blue-black color palette, creating an atmosphere of isolated wonder and eerie, solemn silence.", + "seed": 21366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_15/index_5825_seed_21366_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": -0.0967989151366055, + "vlm_logprob": -0.0967989151366055, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.0967989151366055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06521403789520264 + ], + "opposite_logprob_tokens": [ + -2.815214157104492 + ], + "logprob_reward": -0.06521403789520264, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06521403789520264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07495328783988953, + 0.11599335074424744, + 0.18680162727832794, + 0.02640298381447792, + 0.11170496791601181, + 0.06082829087972641, + 0.1999099999666214, + 0.09423508495092392, + 0.053405534476041794, + 0.07576491683721542 + ], + "credit_map": [ + 0.008156737312674522, + 0.00903802365064621, + 0.008355490863323212, + 0.010680995881557465, + 0.006324352230876684, + 0.009538467973470688, + 0.012819635681807995, + 0.010039575397968292, + 0.020733343437314034, + 0.016705375164747238, + 0.010304996743798256, + 0.014405583031475544, + 0.007778297644108534, + 0.01186616811901331, + 0.023597804829478264, + 0.010601772926747799, + 0.016386395320296288, + 0.025124479085206985, + 0.016170402988791466, + 0.046723902225494385, + 0.00927545316517353, + 0.014514869078993797, + 0.04335634037852287, + 0.015249773859977722, + 0.001940011279657483, + 0.0032939035445451736, + 0.0023756411392241716, + 0.004346892237663269, + 0.003103245748206973, + 0.004696234595030546, + 0.003909279126673937, + 0.002737773349508643, + 0.008535221219062805, + 0.01448800228536129, + 0.008128033019602299, + 0.01717413403093815, + 0.01567106880247593, + 0.01198606751859188, + 0.023450594395399094, + 0.012271841056644917, + 0.0050805616192519665, + 0.00890736561268568, + 0.006663541775196791, + 0.01117679476737976, + 0.006519213318824768, + 0.008998056873679161, + 0.00817844457924366, + 0.005304303485900164, + 0.008053501136600971, + 0.08375439047813416, + 0.0038190509658306837, + 0.007541857659816742, + 0.007178858853876591, + 0.0813693106174469, + 0.0034419260919094086, + 0.004751086235046387, + 0.007287775631994009, + 0.022196590900421143, + 0.011471730656921864, + 0.010630535893142223, + 0.006486440077424049, + 0.016811013221740723, + 0.010684369131922722, + 0.008666625246405602, + 0.004462174605578184, + 0.012996701523661613, + 0.003951486200094223, + 0.004693058785051107, + 0.004348183982074261, + 0.013432232663035393, + 0.0060214269906282425, + 0.0035002639051526785, + 0.004460635129362345, + 0.01718345656991005, + 0.006787884049117565, + 0.008526906371116638, + 0.005594201385974884, + 0.01633208990097046, + 0.009182896465063095, + 0.007696842774748802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.06521403789520264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268380511552095 + ], + "opposite_logprob_tokens": [ + -5.255268573760986 + ], + "logprob_reward": -0.005268380511552095, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268380511552095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06507478654384613, + 0.05854782462120056, + 0.06823765486478806, + 0.01181560754776001, + 0.030168399214744568, + 0.5589335560798645, + 0.06606493890285492, + 0.06531159579753876, + 0.03311734274029732, + 0.04272834584116936 + ], + "credit_map": [ + 0.008060680702328682, + 0.006814438384026289, + 0.008260670118033886, + 0.008452534675598145, + 0.003719493979588151, + 0.009979377500712872, + 0.011159805580973625, + 0.00862779002636671, + 0.010459363460540771, + 0.008119417354464531, + 0.005957828368991613, + 0.008287128992378712, + 0.0031872871331870556, + 0.005365418270230293, + 0.013387549668550491, + 0.0037838283460587263, + 0.005718247964978218, + 0.009348259307444096, + 0.0054467967711389065, + 0.012694722972810268, + 0.005564295686781406, + 0.0071772364899516106, + 0.015610991045832634, + 0.006677105091512203, + 0.0009548557572998106, + 0.0017357501201331615, + 0.0009142435155808926, + 0.002117056166753173, + 0.0013685444137081504, + 0.0017978998366743326, + 0.0018405007431283593, + 0.0010867567034438252, + 0.003130852710455656, + 0.00467201741412282, + 0.002398064127191901, + 0.005192747805267572, + 0.0030225098598748446, + 0.0027986897621303797, + 0.005959213245660067, + 0.002994303125888109, + 0.022460153326392174, + 0.24805903434753418, + 0.007779635023325682, + 0.008542208932340145, + 0.016950780525803566, + 0.23746824264526367, + 0.007060816045850515, + 0.010612692683935165, + 0.0025331920478492975, + 0.026684612035751343, + 0.0015519045991823077, + 0.0023020815569907427, + 0.0034308466129004955, + 0.026013623923063278, + 0.0015501382295042276, + 0.0019985316321253777, + 0.006280978210270405, + 0.016713684424757957, + 0.006054547615349293, + 0.006640206091105938, + 0.006622668355703354, + 0.012987610884010792, + 0.005184171721339226, + 0.004827726166695356, + 0.001693099271506071, + 0.010159202851355076, + 0.0012163477949798107, + 0.0017243900801986456, + 0.0030296084005385637, + 0.011054418981075287, + 0.002694897586479783, + 0.0015453774249181151, + 0.0025384430773556232, + 0.01157306507229805, + 0.0030532353557646275, + 0.0034038543235510588, + 0.0031467208173125982, + 0.01201984565705061, + 0.004063876811414957, + 0.0029293035622686148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.005268380511552095 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107768900692463 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107768900692463, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107768900692463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03815943002700806, + 0.03650152310729027, + 0.024790456518530846, + 0.017181305214762688, + 0.05953585356473923, + 0.5283128619194031, + 0.17798317968845367, + 0.06080710142850876, + 0.02016417682170868, + 0.036564115434885025 + ], + "credit_map": [ + 0.004541642963886261, + 0.007163359317928553, + 0.004019356332719326, + 0.004759502597153187, + 0.002902700798586011, + 0.005064427852630615, + 0.006104332860559225, + 0.0036041054408997297, + 0.009985366836190224, + 0.003981894347816706, + 0.0018752127652987838, + 0.002515968633815646, + 0.0017308746464550495, + 0.0037183696404099464, + 0.010733239352703094, + 0.001960598398000002, + 0.0022934088483452797, + 0.0037412610836327076, + 0.002083311788737774, + 0.0055461120791733265, + 0.002079008612781763, + 0.0026715612038969994, + 0.004747024271637201, + 0.0016287695616483688, + 0.0015030927024781704, + 0.0026334526482969522, + 0.0013409293023869395, + 0.00405519874766469, + 0.002254173159599304, + 0.0016608345322310925, + 0.0023738646414130926, + 0.001359758898615837, + 0.004798409063369036, + 0.008367007598280907, + 0.0033511335495859385, + 0.00959295965731144, + 0.008851679973304272, + 0.007289699278771877, + 0.012269966304302216, + 0.00501500116661191, + 0.02133365161716938, + 0.23558309674263, + 0.007062581833451986, + 0.008490603417158127, + 0.01583479344844818, + 0.2253800928592682, + 0.0046334778890013695, + 0.009994573891162872, + 0.006343785207718611, + 0.0753205269575119, + 0.003655001288279891, + 0.004605260211974382, + 0.006776934023946524, + 0.07478408515453339, + 0.002863504458218813, + 0.0036340884398669004, + 0.005686704069375992, + 0.011715809814631939, + 0.006151316687464714, + 0.0062384032644331455, + 0.00839977990835905, + 0.00987580232322216, + 0.006402252707630396, + 0.006337035447359085, + 0.0018572479020804167, + 0.003230130532756448, + 0.0018295804038643837, + 0.0026344566140323877, + 0.002507612807676196, + 0.0035001009237021208, + 0.0027709968853741884, + 0.0018340508686378598, + 0.0031207113061100245, + 0.008663535118103027, + 0.002902443753555417, + 0.0037019355222582817, + 0.004017634317278862, + 0.00581392552703619, + 0.004776898771524429, + 0.0035670287907123566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.01107768900692463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341660022735596 + ], + "opposite_logprob_tokens": [ + -1.313416600227356 + ], + "logprob_reward": -0.31341660022735596, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341660022735596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07597687095403671, + 0.04549510404467583, + 0.0529121458530426, + 0.04106374457478523, + 0.04140518233180046, + 0.3297554552555084, + 0.1731962412595749, + 0.14835847914218903, + 0.03643235191702843, + 0.05540449544787407 + ], + "credit_map": [ + 0.0076950183138251305, + 0.010221605189144611, + 0.008276856504380703, + 0.009149909019470215, + 0.008425268344581127, + 0.009809330105781555, + 0.0121757872402668, + 0.010223094373941422, + 0.010152113623917103, + 0.005759255029261112, + 0.0026765328366309404, + 0.004314987920224667, + 0.002166962483897805, + 0.003984388895332813, + 0.013963875360786915, + 0.0024769820738583803, + 0.005618663504719734, + 0.006931915879249573, + 0.003585369559004903, + 0.006056512705981731, + 0.00990214478224516, + 0.010553177446126938, + 0.006775550078600645, + 0.0034888091031461954, + 0.004833674058318138, + 0.0057867104187607765, + 0.0026142229326069355, + 0.007081905845552683, + 0.006208107341080904, + 0.005111957434564829, + 0.005894605536013842, + 0.0035325614735484123, + 0.0035788370296359062, + 0.005242225248366594, + 0.0027171762194484472, + 0.0070371502079069614, + 0.006747601553797722, + 0.0054198941215872765, + 0.0075511145405471325, + 0.0031111801508814096, + 0.012933931313455105, + 0.14045602083206177, + 0.006077518220990896, + 0.010690752416849136, + 0.0103659862652421, + 0.13515280187129974, + 0.007243751548230648, + 0.006834668572992086, + 0.0059314812533557415, + 0.07433004677295685, + 0.002386001404374838, + 0.003947847057133913, + 0.006099574733525515, + 0.07451224327087402, + 0.0026128769386559725, + 0.0033761647064238787, + 0.014212976209819317, + 0.032899439334869385, + 0.014893289655447006, + 0.014010327868163586, + 0.020027436316013336, + 0.027076294645667076, + 0.011720235459506512, + 0.013518458232283592, + 0.004546714946627617, + 0.005313577130436897, + 0.0035083291586488485, + 0.003286707680672407, + 0.004896772559732199, + 0.005716149695217609, + 0.005054505076259375, + 0.004109593108296394, + 0.004286249168217182, + 0.010037939995527267, + 0.00462691904976964, + 0.007292770314961672, + 0.00872244406491518, + 0.010259859263896942, + 0.0052083637565374374, + 0.0049699442461133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.31341660022735596 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901786804199219 + ], + "opposite_logprob_tokens": [ + -2.464017868041992 + ], + "logprob_reward": -0.08901786804199219, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901786804199219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04316584765911102, + 0.03023555688560009, + 0.017307037487626076, + 0.021907156333327293, + 0.03658607602119446, + 0.2581017017364502, + 0.5133364200592041, + 0.03174364194273949, + 0.0160964485257864, + 0.031520113348960876 + ], + "credit_map": [ + 0.0060121952556073666, + 0.0042405459098517895, + 0.00512636499479413, + 0.006045239046216011, + 0.003632340347394347, + 0.0061277602799236774, + 0.0054929377511143684, + 0.006488468963652849, + 0.004018350504338741, + 0.006451811641454697, + 0.002470139181241393, + 0.002827485790476203, + 0.002767950063571334, + 0.0027105468325316906, + 0.005309460684657097, + 0.0036798142828047276, + 0.0018085191259160638, + 0.002734161214902997, + 0.0011258753947913647, + 0.0033169968519359827, + 0.0027900065761059523, + 0.00210884609259665, + 0.0023960741236805916, + 0.0010265595046803355, + 0.0027007185854017735, + 0.0030883965082466602, + 0.0011982967844232917, + 0.0032612907234579325, + 0.0033734848257154226, + 0.0033786252606660128, + 0.0030231003183871508, + 0.0018832454225048423, + 0.002726528560742736, + 0.005049860570579767, + 0.002388161141425371, + 0.00701395096257329, + 0.005755098536610603, + 0.0046131182461977005, + 0.005936385598033667, + 0.0031029758974909782, + 0.010098487138748169, + 0.11445802450180054, + 0.0035890841390937567, + 0.0044583422131836414, + 0.00800288561731577, + 0.10956019163131714, + 0.002740770112723112, + 0.005193961784243584, + 0.017621418461203575, + 0.22850845754146576, + 0.006091793533414602, + 0.008870136924088001, + 0.014456421136856079, + 0.22526004910469055, + 0.004043312277644873, + 0.008484840393066406, + 0.0026121006812900305, + 0.0069104330614209175, + 0.004053091164678335, + 0.0038731805980205536, + 0.002271967474371195, + 0.0050315894186496735, + 0.003545681480318308, + 0.003445601789280772, + 0.000894107564818114, + 0.0038169703911989927, + 0.0009335235226899385, + 0.0018393739592283964, + 0.0017696594586595893, + 0.004348196089267731, + 0.0015713085886090994, + 0.0009233106975443661, + 0.002694954164326191, + 0.00748884491622448, + 0.00311249028891325, + 0.003023965284228325, + 0.0027229932602494955, + 0.00590415857732296, + 0.0035665605682879686, + 0.0030061486177146435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.08901786804199219 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera captures a steady, first-person perspective from the cockpit of a deep-sea submersible, drifting silently and deliberately forward through the darkness of the abyss. The only interior illumination comes from the soft, cool glow of the digital instrument panel, its blue and green readouts and blinking indicator lights reflected faintly on the thickly reinforced circular observation area. Outside, the submersible's powerful cone of cool white light cuts through the inky water, revealing a vast alien landscape of the seafloor below, covered in dense, shimmering, circular metallic nodules that gleam under the artificial light. Ethereal bioluminescence, resembling ghostly jellyfish and swarms of tiny, glowing specks, drifts gracefully across the observation area, like a slow-motion starfield, their soft cyan and white glow gently pulsing through the oppressive depths. In the distance, a massive, chimney-like hydrothermal vent looms, continuously spewing thick, opaque black smoke that billows upward into the shattered void. The entire scene is rendered in a distinctly dark anime style, characterized by sharp, hard-edged shadows, high contrast, and a moody blue-black color palette, creating an atmosphere of isolated wonder and eerie, solemn silence.", + "seed": 21367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_15/index_5825_seed_21367_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": -0.07934554740786552, + "vlm_logprob": -0.07934554740786552, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.07934554740786552, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08251874893903732 + ], + "opposite_logprob_tokens": [ + -2.5825188159942627 + ], + "logprob_reward": -0.08251874893903732, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08251874893903732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03913252055644989, + 0.02606632187962532, + 0.0394829586148262, + 0.034525200724601746, + 0.022376177832484245, + 0.019241757690906525, + 0.028447803109884262, + 0.04014384374022484, + 0.04306073114275932, + 0.7075226902961731 + ], + "credit_map": [ + 0.0024872461799532175, + 0.0067808120511472225, + 0.005231800023466349, + 0.003591930726543069, + 0.008248223923146725, + 0.002999349497258663, + 0.006402201484888792, + 0.0033909555058926344, + 0.0023161338176578283, + 0.004318103194236755, + 0.0019688766915351152, + 0.003255415242165327, + 0.004445655271410942, + 0.0030738848727196455, + 0.0044320933520793915, + 0.002256162231788039, + 0.00267154723405838, + 0.004677715711295605, + 0.006694629788398743, + 0.00698451604694128, + 0.0038645383901894093, + 0.0038504325784742832, + 0.004524224903434515, + 0.0062153534963727, + 0.0022049748804420233, + 0.004284787457436323, + 0.004072288982570171, + 0.004867586772888899, + 0.004635070450603962, + 0.003951231483370066, + 0.004860831890255213, + 0.00564842764288187, + 0.0014896830543875694, + 0.0035613987129181623, + 0.001656102016568184, + 0.0017572786891832948, + 0.0030411959160119295, + 0.005575238261371851, + 0.003878036281093955, + 0.0014172445517033339, + 0.0012786607258021832, + 0.0024185157380998135, + 0.0028076081071048975, + 0.0024655002634972334, + 0.003069054801017046, + 0.0026306859217584133, + 0.002781679853796959, + 0.0017900547245517373, + 0.0019497978501021862, + 0.0033642000053077936, + 0.003186378628015518, + 0.004140255972743034, + 0.0031206777784973383, + 0.0038662301376461983, + 0.004688342567533255, + 0.004131918773055077, + 0.0024037363473325968, + 0.006523374933749437, + 0.004694231320172548, + 0.005161667242646217, + 0.0056564765982329845, + 0.00663726544007659, + 0.0049802218563854694, + 0.004086867906153202, + 0.002796861808747053, + 0.004399023950099945, + 0.0069803036749362946, + 0.005868691951036453, + 0.0056715416721999645, + 0.006574851460754871, + 0.004601738881319761, + 0.006167718209326267, + 0.05064716190099716, + 0.09439210593700409, + 0.08784954249858856, + 0.09116847068071365, + 0.07544064521789551, + 0.11836621165275574, + 0.1231599822640419, + 0.06649856269359589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.08251874893903732 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018208375200629234 + ], + "opposite_logprob_tokens": [ + -4.0182085037231445 + ], + "logprob_reward": -0.018208375200629234, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018208375200629234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11105837672948837, + 0.035492341965436935, + 0.04711069166660309, + 0.05572575330734253, + 0.06596619635820389, + 0.01830226182937622, + 0.022238345816731453, + 0.048748504370450974, + 0.07030176371335983, + 0.5250557065010071 + ], + "credit_map": [ + 0.010265767574310303, + 0.01915954425930977, + 0.01840040273964405, + 0.008772468194365501, + 0.01873767003417015, + 0.010495400056242943, + 0.014669753611087799, + 0.010557382367551327, + 0.002558972919359803, + 0.0051482068374753, + 0.0025546529795974493, + 0.00446257833391428, + 0.0062794811092317104, + 0.0041486420668661594, + 0.007228384725749493, + 0.003111425554379821, + 0.002880426589399576, + 0.005809172987937927, + 0.007135181222110987, + 0.008331641554832458, + 0.006808931939303875, + 0.0048719896003603935, + 0.007701048627495766, + 0.0035723024047911167, + 0.00355969718657434, + 0.00724749593064189, + 0.007021518889814615, + 0.007470265496522188, + 0.006821474526077509, + 0.0061498526483774185, + 0.01015022024512291, + 0.0073052337393164635, + 0.003716956125572324, + 0.011529173702001572, + 0.005719912238419056, + 0.005945175886154175, + 0.00603216839954257, + 0.0176140945404768, + 0.007890762761235237, + 0.007517961319535971, + 0.0014770370908081532, + 0.0026486392598599195, + 0.0032627061009407043, + 0.0023891928140074015, + 0.001668440643697977, + 0.0034474304411560297, + 0.0017411696026101708, + 0.0016676471568644047, + 0.0015407850733026862, + 0.002866119844838977, + 0.0018591646803542972, + 0.00273987278342247, + 0.003610866842791438, + 0.0037394671235233545, + 0.003490074537694454, + 0.002391997491940856, + 0.0030919930431991816, + 0.007315750699490309, + 0.005294738337397575, + 0.005429836921393871, + 0.00768553139641881, + 0.008608476258814335, + 0.006026911549270153, + 0.005295267328619957, + 0.006299455184489489, + 0.008606641553342342, + 0.009370866231620312, + 0.007403083611279726, + 0.008656277321279049, + 0.010630588978528976, + 0.006814968306571245, + 0.012519889511168003, + 0.03624514490365982, + 0.0749722495675087, + 0.0631556287407875, + 0.06495845317840576, + 0.05838197097182274, + 0.08863069862127304, + 0.0896514356136322, + 0.04906017333269119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.018208375200629234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641822263598442 + ], + "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": [ + 0.04554468020796776, + 0.05613962560892105, + 0.09068219363689423, + 0.33851638436317444, + 0.04999479651451111, + 0.0939117893576622, + 0.02613726630806923, + 0.02951536327600479, + 0.04123707115650177, + 0.22832076251506805 + ], + "credit_map": [ + 0.006743282545357943, + 0.007789953611791134, + 0.007835609838366508, + 0.0030431768391281366, + 0.007291609887033701, + 0.005114836618304253, + 0.00462113693356514, + 0.0031050746329128742, + 0.004151771776378155, + 0.010260054841637611, + 0.0033734182361513376, + 0.006046661175787449, + 0.012106407433748245, + 0.005841747857630253, + 0.010485775768756866, + 0.0038737838622182608, + 0.005394402891397476, + 0.009748565964400768, + 0.010387260466814041, + 0.013659200631082058, + 0.012785274535417557, + 0.009221023879945278, + 0.015337744727730751, + 0.014148720540106297, + 0.01829349622130394, + 0.05877193063497543, + 0.05097358301281929, + 0.048771679401397705, + 0.03223240002989769, + 0.045266930013895035, + 0.05357284098863602, + 0.030633511021733284, + 0.0038427021354436874, + 0.006315345410257578, + 0.006285934243351221, + 0.005237222649157047, + 0.0073281461372971535, + 0.0077445427887141705, + 0.007546047680079937, + 0.005694854538887739, + 0.005471060983836651, + 0.01351054199039936, + 0.011279637925326824, + 0.011811464093625546, + 0.013120926916599274, + 0.016320889815688133, + 0.011822511442005634, + 0.010574761778116226, + 0.001833234797231853, + 0.003796529956161976, + 0.0017657416174188256, + 0.003219755133613944, + 0.003096693893894553, + 0.0034917080774903297, + 0.005517994984984398, + 0.0034156094770878553, + 0.0020052262116223574, + 0.0029145856387913227, + 0.0032665454782545567, + 0.004683354869484901, + 0.00572882080450654, + 0.0044032554142177105, + 0.0036608604714274406, + 0.0028527143876999617, + 0.0031173708848655224, + 0.006121237762272358, + 0.005291977897286415, + 0.004420140292495489, + 0.004962598904967308, + 0.00760657200589776, + 0.005349332932382822, + 0.004367837682366371, + 0.015224700793623924, + 0.03232995420694351, + 0.0283114705234766, + 0.027442390099167824, + 0.026873042806982994, + 0.036990731954574585, + 0.03945395350456238, + 0.02169450744986534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "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 perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270202398300171 + ], + "opposite_logprob_tokens": [ + -2.1270203590393066 + ], + "logprob_reward": -0.1270202398300171, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270202398300171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08094455301761627, + 0.04763542488217354, + 0.17067672312259674, + 0.09999201446771622, + 0.2074439525604248, + 0.06420020014047623, + 0.04283677041530609, + 0.1060989573597908, + 0.06687246263027191, + 0.1132989302277565 + ], + "credit_map": [ + 0.009865579195320606, + 0.015098636969923973, + 0.009077785536646843, + 0.0065869176760315895, + 0.01645692065358162, + 0.004494932945817709, + 0.013447331264615059, + 0.005916450172662735, + 0.002532804850488901, + 0.009370213374495506, + 0.002666001208126545, + 0.0045296368189156055, + 0.012079576030373573, + 0.004168274346739054, + 0.008789381943643093, + 0.0034995393361896276, + 0.010017219930887222, + 0.022435834631323814, + 0.012568868696689606, + 0.023119177669286728, + 0.02695392072200775, + 0.022870976477861404, + 0.03594089299440384, + 0.016769833862781525, + 0.008011337369680405, + 0.02022046037018299, + 0.011402654461562634, + 0.012164146639406681, + 0.011134956032037735, + 0.012689224444329739, + 0.00951453298330307, + 0.01485469564795494, + 0.011659704148769379, + 0.02892731875181198, + 0.019741347059607506, + 0.01921342872083187, + 0.030583079904317856, + 0.04516581818461418, + 0.025951026007533073, + 0.02620224468410015, + 0.003988552838563919, + 0.006624056492000818, + 0.011196091771125793, + 0.011136758141219616, + 0.007446017116308212, + 0.008846674114465714, + 0.008404286578297615, + 0.006557765882462263, + 0.0022498401813209057, + 0.005745931528508663, + 0.0024266825057566166, + 0.004224920645356178, + 0.008513343520462513, + 0.005866040010005236, + 0.007462118286639452, + 0.00634789140895009, + 0.0068495012819767, + 0.0142317870631814, + 0.009893670678138733, + 0.011572062037885189, + 0.023156486451625824, + 0.014705076813697815, + 0.011220299638807774, + 0.014470072463154793, + 0.004565501119941473, + 0.012438863515853882, + 0.006753270979970694, + 0.006136775016784668, + 0.013018413446843624, + 0.011140208691358566, + 0.007984906435012817, + 0.004834518767893314, + 0.00726736756041646, + 0.014897623099386692, + 0.013621496967971325, + 0.013444027863442898, + 0.014142216183245182, + 0.018298659473657608, + 0.020272424444556236, + 0.011355103924870491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.1270202398300171 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033855080604553 + ], + "opposite_logprob_tokens": [ + -1.9103385210037231 + ], + "logprob_reward": -0.16033855080604553, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033855080604553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07583542913198471, + 0.03470354154706001, + 0.11628589779138565, + 0.11789122223854065, + 0.13369672000408173, + 0.12262601405382156, + 0.051548950374126434, + 0.0717601627111435, + 0.11431770026683807, + 0.16133427619934082 + ], + "credit_map": [ + 0.012999823316931725, + 0.013681404292583466, + 0.011522813700139523, + 0.005558227188885212, + 0.015172303654253483, + 0.005321590229868889, + 0.0061388020403683186, + 0.005440463311970234, + 0.0028627202846109867, + 0.00597243569791317, + 0.0027899974957108498, + 0.0032060144003480673, + 0.006500939838588238, + 0.004954701755195856, + 0.005358791444450617, + 0.003057940164580941, + 0.00687036756426096, + 0.016822434961795807, + 0.013005398213863373, + 0.016444029286503792, + 0.02056395821273327, + 0.014837540686130524, + 0.01976952701807022, + 0.007972651161253452, + 0.00967754889279604, + 0.019011560827493668, + 0.018844066187739372, + 0.01574203372001648, + 0.01220300979912281, + 0.01522065233439207, + 0.015814632177352905, + 0.011377719230949879, + 0.005739447195082903, + 0.029695842415094376, + 0.007783680222928524, + 0.009456745348870754, + 0.017588669434189796, + 0.03539697453379631, + 0.01723506674170494, + 0.010800301097333431, + 0.008035003207623959, + 0.01926826499402523, + 0.01249427068978548, + 0.011671011336147785, + 0.014748858287930489, + 0.027620162814855576, + 0.013986109755933285, + 0.014802340418100357, + 0.003528250614181161, + 0.008311011828482151, + 0.0039115226827561855, + 0.005526411812752485, + 0.008394250646233559, + 0.007843640632927418, + 0.00814705342054367, + 0.005886807572096586, + 0.004822594579309225, + 0.010345213115215302, + 0.00785841979086399, + 0.008605440147221088, + 0.012453312054276466, + 0.010482474230229855, + 0.007250675931572914, + 0.009942041710019112, + 0.008581588976085186, + 0.015077506192028522, + 0.017225375398993492, + 0.01107115764170885, + 0.013060484081506729, + 0.020023217424750328, + 0.013973848894238472, + 0.015304522588849068, + 0.010913943871855736, + 0.022169839590787888, + 0.02058505080640316, + 0.02013831026852131, + 0.01848030462861061, + 0.026555119082331657, + 0.02783103473484516, + 0.014660666696727276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.16033855080604553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera captures a steady, first-person perspective from the cockpit of a deep-sea submersible, drifting silently and deliberately forward through the darkness of the abyss. The only interior illumination comes from the soft, cool glow of the digital instrument panel, its blue and green readouts and blinking indicator lights reflected faintly on the thickly reinforced circular observation area. Outside, the submersible's powerful cone of cool white light cuts through the inky water, revealing a vast alien landscape of the seafloor below, covered in dense, shimmering, circular metallic nodules that gleam under the artificial light. Ethereal bioluminescence, resembling ghostly jellyfish and swarms of tiny, glowing specks, drifts gracefully across the observation area, like a slow-motion starfield, their soft cyan and white glow gently pulsing through the oppressive depths. In the distance, a massive, chimney-like hydrothermal vent looms, continuously spewing thick, opaque black smoke that billows upward into the shattered void. The entire scene is rendered in a distinctly dark anime style, characterized by sharp, hard-edged shadows, high contrast, and a moody blue-black color palette, creating an atmosphere of isolated wonder and eerie, solemn silence.", + "seed": 21368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_15/index_5825_seed_21368_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": -0.03881449461914599, + "vlm_logprob": -0.03881449461914599, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03881449461914599, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04830896481871605 + ], + "opposite_logprob_tokens": [ + -3.173308849334717 + ], + "logprob_reward": -0.04830896481871605, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04830896481871605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20070579648017883, + 0.06210845708847046, + 0.042122066020965576, + 0.0424099899828434, + 0.1129162460565567, + 0.06122554838657379, + 0.07783741503953934, + 0.21708816289901733, + 0.11771632730960846, + 0.06586998701095581 + ], + "credit_map": [ + 0.014446720480918884, + 0.03318718075752258, + 0.02194073796272278, + 0.032564036548137665, + 0.030587196350097656, + 0.023305922746658325, + 0.02405823953449726, + 0.020615767687559128, + 0.005916579160839319, + 0.009007946588099003, + 0.0058319768868386745, + 0.0076074679382145405, + 0.009090169332921505, + 0.007074290886521339, + 0.0061134882271289825, + 0.011466537602245808, + 0.004075157456099987, + 0.006196908187121153, + 0.004254800733178854, + 0.004835113417357206, + 0.0051424866542220116, + 0.00788537785410881, + 0.00488481018692255, + 0.004847410600632429, + 0.003247976303100586, + 0.004998224787414074, + 0.0040402780286967754, + 0.007550700567662716, + 0.008667287416756153, + 0.00547595601528883, + 0.004548007156699896, + 0.003881558310240507, + 0.011055577546358109, + 0.01807883195579052, + 0.011048073880374432, + 0.00978011917322874, + 0.02258450537919998, + 0.023961564525961876, + 0.008295247331261635, + 0.008112328127026558, + 0.005579275544732809, + 0.006589837837964296, + 0.011587923392653465, + 0.009220961481332779, + 0.008070379495620728, + 0.005981497932225466, + 0.007130282465368509, + 0.0070653920993208885, + 0.006779995281249285, + 0.010688979178667068, + 0.013105803169310093, + 0.011114660650491714, + 0.008801177144050598, + 0.008786453865468502, + 0.0064426567405462265, + 0.012117687612771988, + 0.02308955043554306, + 0.028153369203209877, + 0.08485987782478333, + 0.02885192260146141, + 0.014220460318028927, + 0.016846664249897003, + 0.008836203254759312, + 0.012230110354721546, + 0.011847643181681633, + 0.014056248590350151, + 0.024147560819983482, + 0.015856128185987473, + 0.012694799341261387, + 0.0104930829256773, + 0.014085243456065655, + 0.014535622671246529, + 0.009183429181575775, + 0.007602727506309748, + 0.021982068195939064, + 0.007605533581227064, + 0.004226360935717821, + 0.008233278058469296, + 0.0034622475504875183, + 0.0035743373446166515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.04830896481871605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028389885555952787 + ], + "opposite_logprob_tokens": [ + -5.877839088439941 + ], + "logprob_reward": -0.0028389885555952787, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028389885555952787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09947020560503006, + 0.08690600097179413, + 0.06891060620546341, + 0.05062096565961838, + 0.09836243838071823, + 0.06590774655342102, + 0.24582692980766296, + 0.16775164008140564, + 0.050837382674217224, + 0.06540614366531372 + ], + "credit_map": [ + 0.007722300477325916, + 0.012286650948226452, + 0.013023055158555508, + 0.025327758863568306, + 0.014113632962107658, + 0.012263676151633263, + 0.00927477516233921, + 0.005458350293338299, + 0.009417890571057796, + 0.012578706257045269, + 0.009326628409326077, + 0.015829190611839294, + 0.013424842618405819, + 0.008883493021130562, + 0.006361133884638548, + 0.011084110476076603, + 0.005113057792186737, + 0.010141085833311081, + 0.006547878962010145, + 0.009041235782206059, + 0.009532921947538853, + 0.013835050165653229, + 0.0064504980109632015, + 0.008248879574239254, + 0.004248886369168758, + 0.007197747938334942, + 0.004091643262654543, + 0.007328574080020189, + 0.01190855260938406, + 0.007311750203371048, + 0.003397357417270541, + 0.005136456806212664, + 0.008986685425043106, + 0.013473929837346077, + 0.012560684233903885, + 0.01091768965125084, + 0.02189120650291443, + 0.014984277077019215, + 0.007520735263824463, + 0.008027228526771069, + 0.006403924431651831, + 0.012559185735881329, + 0.006498430855572224, + 0.006018851418048143, + 0.016819395124912262, + 0.0066245198249816895, + 0.005965700373053551, + 0.005017738789319992, + 0.021898144856095314, + 0.03951728716492653, + 0.08932943642139435, + 0.023602008819580078, + 0.02156517282128334, + 0.023180773481726646, + 0.01335838995873928, + 0.013375716283917427, + 0.011637899093329906, + 0.02105802856385708, + 0.04631305858492851, + 0.02609970048069954, + 0.0178823284804821, + 0.01796894706785679, + 0.016092387959361076, + 0.010699287056922913, + 0.004277767613530159, + 0.004207078367471695, + 0.0073661524802446365, + 0.007348412182182074, + 0.005496927071362734, + 0.0051956684328615665, + 0.006770012900233269, + 0.010175363160669804, + 0.008107819594442844, + 0.008700661361217499, + 0.01663694716989994, + 0.00742825772613287, + 0.007397301495075226, + 0.006467478349804878, + 0.0058607179671525955, + 0.004806956276297569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0028389885555952787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738320458680391 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738320458680391, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738320458680391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0537213496863842, + 0.050709933042526245, + 0.04809586703777313, + 0.03974834457039833, + 0.201108917593956, + 0.09414288401603699, + 0.25290215015411377, + 0.14230912923812866, + 0.06350705772638321, + 0.0537542849779129 + ], + "credit_map": [ + 0.004357887897640467, + 0.005911714397370815, + 0.00600598007440567, + 0.008272242732346058, + 0.006839980371296406, + 0.011224552989006042, + 0.0050633735954761505, + 0.006045622751116753, + 0.004986534360796213, + 0.006262251641601324, + 0.005335070192813873, + 0.007244792766869068, + 0.009425606578588486, + 0.005678970366716385, + 0.004492334555834532, + 0.00728437677025795, + 0.003906422294676304, + 0.007876884192228317, + 0.00291532464325428, + 0.004118009004741907, + 0.006123632192611694, + 0.013445761054754257, + 0.0042465622536838055, + 0.005463276989758015, + 0.003740874817594886, + 0.004447267856448889, + 0.0046343011781573296, + 0.004922149237245321, + 0.008085431531071663, + 0.005852379836142063, + 0.003354874672368169, + 0.0047110687009990215, + 0.018326742574572563, + 0.03205244988203049, + 0.02189680188894272, + 0.021716218441724777, + 0.04833676293492317, + 0.026711206883192062, + 0.01716703549027443, + 0.014901715330779552, + 0.008687693625688553, + 0.012146567925810814, + 0.011032904498279095, + 0.010902934707701206, + 0.02235579304397106, + 0.01117298100143671, + 0.010089338757097721, + 0.007754684425890446, + 0.025435661897063255, + 0.03595210239291191, + 0.05937505140900612, + 0.02854222245514393, + 0.028197001665830612, + 0.036205969750881195, + 0.018742674961686134, + 0.020451484248042107, + 0.014028181321918964, + 0.021074369549751282, + 0.05798778310418129, + 0.017118502408266068, + 0.009783689863979816, + 0.010971941985189915, + 0.005800515878945589, + 0.005544146057218313, + 0.00554257957264781, + 0.006339969579130411, + 0.013564277440309525, + 0.007920772768557072, + 0.00536632165312767, + 0.006562447175383568, + 0.007737981621176004, + 0.01047271117568016, + 0.005190103314816952, + 0.009423456154763699, + 0.01377174537628889, + 0.004667883738875389, + 0.004785333294421434, + 0.0063611650839447975, + 0.005443268921226263, + 0.004111333284527063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.006738320458680391 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023275146260857582 + ], + "opposite_logprob_tokens": [ + -3.773275136947632 + ], + "logprob_reward": -0.023275146260857582, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023275146260857582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03865080326795578, + 0.05378435552120209, + 0.1960870772600174, + 0.20463283360004425, + 0.15109725296497345, + 0.08422799408435822, + 0.09803783893585205, + 0.05112912133336067, + 0.04232189804315567, + 0.08003084361553192 + ], + "credit_map": [ + 0.0033684917725622654, + 0.005753066390752792, + 0.0053300997242331505, + 0.0061481911689043045, + 0.006377902813255787, + 0.003791910596191883, + 0.0037334109656512737, + 0.004147728439420462, + 0.005260712932795286, + 0.00852350052446127, + 0.0031894368585199118, + 0.005318158771842718, + 0.015349341556429863, + 0.004918971564620733, + 0.0046607013791799545, + 0.006563536822795868, + 0.01745564304292202, + 0.03622226417064667, + 0.008700116537511349, + 0.015350854024291039, + 0.03560370206832886, + 0.048472415655851364, + 0.017187004908919334, + 0.01709507219493389, + 0.015322037972509861, + 0.023992013186216354, + 0.013473490253090858, + 0.02700909972190857, + 0.05149661377072334, + 0.025416851043701172, + 0.022587550804018974, + 0.025335168465971947, + 0.014301019720733166, + 0.025721421465277672, + 0.017259934917092323, + 0.013218708336353302, + 0.027934018522500992, + 0.021988604217767715, + 0.014548838138580322, + 0.016124702990055084, + 0.008176695555448532, + 0.01244365330785513, + 0.010715880431234837, + 0.008918296545743942, + 0.018092602491378784, + 0.008540211245417595, + 0.00929610151797533, + 0.00804455392062664, + 0.009523236192762852, + 0.01185278408229351, + 0.011099082417786121, + 0.012479950673878193, + 0.015987707301974297, + 0.016489144414663315, + 0.010711885057389736, + 0.009894055314362049, + 0.0039054437074810266, + 0.007444823160767555, + 0.017068563029170036, + 0.005517621524631977, + 0.006044199224561453, + 0.004174984525889158, + 0.004251635633409023, + 0.002721851458773017, + 0.003367654513567686, + 0.005371349398046732, + 0.006857023108750582, + 0.0056839571334421635, + 0.005431132856756449, + 0.0037584027741104364, + 0.006024958565831184, + 0.005827419925481081, + 0.010889924131333828, + 0.011983026750385761, + 0.017332853749394417, + 0.00933038629591465, + 0.009325582534074783, + 0.007724943570792675, + 0.006194218061864376, + 0.007249908987432718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.023275146260857582 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a steady first-person view from inside a submersible cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior illuminated by a soft cool glow from a digital instrument panel with blue and green readouts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cone of cool white light cut through the darkness to reveal a seafloor covered in dense, shimmering, circular metallic nodules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive chimney-like hydrothermal vent visible in the distance spewing thick opaque black smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a dark anime style with sharp hard-edged shadows, high contrast, and a moody blue-black color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291105300188065 + ], + "opposite_logprob_tokens": [ + -2.2379109859466553 + ], + "logprob_reward": -0.11291105300188065, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291105300188065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11478104442358017, + 0.151187002658844, + 0.0511472225189209, + 0.07737645506858826, + 0.13662491738796234, + 0.11962030082941055, + 0.1826598048210144, + 0.05149601399898529, + 0.03293256834149361, + 0.08217458426952362 + ], + "credit_map": [ + 0.01164326909929514, + 0.015834014862775803, + 0.012677617371082306, + 0.013176867738366127, + 0.019040241837501526, + 0.020792923867702484, + 0.011803871020674706, + 0.009812243282794952, + 0.017098302021622658, + 0.017169691622257233, + 0.010387971065938473, + 0.016978926956653595, + 0.02178330533206463, + 0.019599968567490578, + 0.013028185814619064, + 0.03514067828655243, + 0.004028335679322481, + 0.006635194178670645, + 0.004280782770365477, + 0.007321221753954887, + 0.009599551558494568, + 0.007523335982114077, + 0.005135101266205311, + 0.0066237058490514755, + 0.005471392534673214, + 0.00847718771547079, + 0.004374411888420582, + 0.011433335021138191, + 0.0172612052410841, + 0.009980282746255398, + 0.008355560712516308, + 0.012023087590932846, + 0.010392827913165092, + 0.02490716427564621, + 0.015840796753764153, + 0.010749844834208488, + 0.020164277404546738, + 0.03177564963698387, + 0.006828604266047478, + 0.01596575416624546, + 0.011417551897466183, + 0.01586821675300598, + 0.020016441121697426, + 0.017439214512705803, + 0.017669983208179474, + 0.011485513299703598, + 0.013117765076458454, + 0.01260562427341938, + 0.018398219719529152, + 0.026352008804678917, + 0.04679310321807861, + 0.01989169418811798, + 0.020301319658756256, + 0.024054523557424545, + 0.012495850212872028, + 0.014373104088008404, + 0.004688155837357044, + 0.00798928551375866, + 0.02077195793390274, + 0.006171736866235733, + 0.0038293979596346617, + 0.0033794830087572336, + 0.0019275472732260823, + 0.0027384585700929165, + 0.0034119412302970886, + 0.0036730682477355003, + 0.0063856798224151134, + 0.00495366333052516, + 0.0031268021557480097, + 0.0035766714718192816, + 0.0034494129940867424, + 0.004355332348495722, + 0.010068976320326328, + 0.010162044316530228, + 0.012209432199597359, + 0.00911546591669321, + 0.008539406582713127, + 0.008509603329002857, + 0.009533829987049103, + 0.014035828411579132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.11291105300188065 + } + ], + "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/0000128/group_2/index_tensor([2245, 2245])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_2/index_tensor([2245, 2245])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..58e068aef53a8cc9e29483984d3fecf8de390f62 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_2/index_tensor([2245, 2245])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static, full-body photograph captures the intense courtship ritual between two serrated river crabs, their bodies rendered in hyperrealistic CGI, highlighting every minute texture and damp sheen. They lie amidst moss-covered rocks in a clear, shallow stream, dappled sunlight filtering through the unseen forest canopy, casting a dancing, corroded pattern across the scene. One freshwater crab, its carapace a complex mosaic of dark brown and pale ochre, cautiously approaches the other with slow, deliberate sideways swoops. Its legs carefully rest on the velvety green moss clinging to the damp river rocks. Steadily and deliberately, it extends its large right claw, its claw slightly open, in a subtle, non-threatening display. The second crab remains motionless, like a cautious, observing statue, its eyestalks intently tracking the suitor's advances, its own claws firmly tucked against its body. Clear water flows gently around them, revealing the sandy, pebbly riverbed beneath, creating an atmosphere of palpable tranquility, a natural tension where the raw drama unfolds with lifelike precision.", + "seed": 17318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_2/index_2245_seed_17318_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": -0.07675537783652545, + "vlm_logprob": -0.07675537783652545, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07675537783652545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3003801703453064 + ], + "opposite_logprob_tokens": [ + -1.4253802299499512 + ], + "logprob_reward": -0.3003801703453064, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3003801703453064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19160442054271698, + 0.26172882318496704, + 0.24334022402763367, + 0.11964984238147736, + 0.03340572491288185, + 0.011445696465671062, + 0.012913001701235771, + 0.047641824930906296, + 0.05875905603170395, + 0.0195113904774189 + ], + "credit_map": [ + 0.018699949607253075, + 0.04810323938727379, + 0.014936951920390129, + 0.014843703247606754, + 0.01205681823194027, + 0.05236409232020378, + 0.01118846982717514, + 0.01941119134426117, + 0.026835603639483452, + 0.09362798184156418, + 0.008505629375576973, + 0.007969917729496956, + 0.01540451217442751, + 0.0841328427195549, + 0.00793714914470911, + 0.017315194010734558, + 0.04973667487502098, + 0.0963205024600029, + 0.01576937921345234, + 0.01051622536033392, + 0.027479803189635277, + 0.020610978826880455, + 0.010519657284021378, + 0.012387016788125038, + 0.014320701360702515, + 0.010476529598236084, + 0.007799424231052399, + 0.011567625217139721, + 0.01789977215230465, + 0.01559096947312355, + 0.028242992237210274, + 0.013751824386417866, + 0.006304984446614981, + 0.0046077570877969265, + 0.004453885834664106, + 0.002755246590822935, + 0.0037009865045547485, + 0.005063097923994064, + 0.00323731591925025, + 0.003282449208199978, + 0.001681082765571773, + 0.0014113114448264241, + 0.0016291072824969888, + 0.0012979685561731458, + 0.0017756024608388543, + 0.001696167979389429, + 0.0008985651656985283, + 0.0010558912763372064, + 0.002307180780917406, + 0.0015936573036015034, + 0.0014672547113150358, + 0.0012736557982861996, + 0.0015275542391464114, + 0.0016517965123057365, + 0.0012951408280059695, + 0.0017967616440728307, + 0.008719632402062416, + 0.00485091982409358, + 0.0034131682477891445, + 0.00426502525806427, + 0.005451932083815336, + 0.013051080517470837, + 0.002948604291304946, + 0.004941465798765421, + 0.016000479459762573, + 0.003062879666686058, + 0.004257700871676207, + 0.003577382303774357, + 0.0025814117398113012, + 0.02292451076209545, + 0.002540603745728731, + 0.0038140846882015467, + 0.0022396931890398264, + 0.004382049664855003, + 0.0020198512356728315, + 0.002763911848887801, + 0.002198257949203253, + 0.0020282319746911526, + 0.001898063812404871, + 0.0019813317339867353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.3003801703453064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014222684316337109 + ], + "opposite_logprob_tokens": [ + -4.264222621917725 + ], + "logprob_reward": -0.014222684316337109, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014222684316337109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07829845696687698, + 0.09058917313814163, + 0.2323487401008606, + 0.058291591703891754, + 0.020367790013551712, + 0.013873626478016376, + 0.02792501635849476, + 0.1702805459499359, + 0.02261776104569435, + 0.28540730476379395 + ], + "credit_map": [ + 0.009088246151804924, + 0.010031876154243946, + 0.006143528502434492, + 0.006726814433932304, + 0.013185490854084492, + 0.013151229359209538, + 0.007615790702402592, + 0.012355481274425983, + 0.008945348672568798, + 0.02467978186905384, + 0.005522739142179489, + 0.0041719949804246426, + 0.007161268033087254, + 0.02703099139034748, + 0.0060907830484211445, + 0.006986256688833237, + 0.044721826910972595, + 0.08566468209028244, + 0.015325596556067467, + 0.011929652653634548, + 0.0265243761241436, + 0.019034411758184433, + 0.016692379489541054, + 0.012455799616873264, + 0.007062325719743967, + 0.0056438324972987175, + 0.004397745709866285, + 0.005181852728128433, + 0.008353456854820251, + 0.008290654979646206, + 0.012116282247006893, + 0.007245440036058426, + 0.0037354317028075457, + 0.0022185214329510927, + 0.0020880522206425667, + 0.0017692295368760824, + 0.0018185994122177362, + 0.004823633935302496, + 0.002036286285147071, + 0.0018780352547764778, + 0.003406173549592495, + 0.001741041080094874, + 0.0010818636510521173, + 0.0009735009516589344, + 0.0012417167890816927, + 0.003037733491510153, + 0.0010301097063347697, + 0.0013614865019917488, + 0.004442557692527771, + 0.0036236972082406282, + 0.002455778419971466, + 0.0028335212264209986, + 0.0033047860488295555, + 0.003949125297367573, + 0.0024194270372390747, + 0.004896123427897692, + 0.04255198314785957, + 0.011506911367177963, + 0.01481142733246088, + 0.006825105287134647, + 0.011812565848231316, + 0.0610855408012867, + 0.008320272900164127, + 0.013366731815040112, + 0.0036195965949445963, + 0.0022596789058297873, + 0.001902933232486248, + 0.0015786818694323301, + 0.0036685923114418983, + 0.005140692461282015, + 0.002074020216241479, + 0.0023735626600682735, + 0.05830332264304161, + 0.06013157218694687, + 0.022549103945493698, + 0.028021177276968956, + 0.03555871918797493, + 0.029719790443778038, + 0.022841088473796844, + 0.028282513841986656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.014222684316337109 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607564091682434 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.01607564091682434, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607564091682434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20602989196777344, + 0.2219892144203186, + 0.17879050970077515, + 0.1595665067434311, + 0.042275674641132355, + 0.0345066599547863, + 0.019773295149207115, + 0.0588887557387352, + 0.025215866044163704, + 0.05296357348561287 + ], + "credit_map": [ + 0.024156738072633743, + 0.05373310297727585, + 0.015995092689990997, + 0.01785259321331978, + 0.014690587297081947, + 0.04897503927350044, + 0.012788158841431141, + 0.017838584259152412, + 0.02760489098727703, + 0.08163267374038696, + 0.00907963141798973, + 0.007754306774586439, + 0.016276510432362556, + 0.05719209834933281, + 0.008504655212163925, + 0.013944455422461033, + 0.035746973007917404, + 0.07002359628677368, + 0.011578776873648167, + 0.007999534718692303, + 0.02059757523238659, + 0.014869265258312225, + 0.008714493364095688, + 0.009260292164981365, + 0.02058526501059532, + 0.013776581734418869, + 0.010373515076935291, + 0.014226286672055721, + 0.022097919136285782, + 0.019506752490997314, + 0.035590723156929016, + 0.023409461602568626, + 0.00612918846309185, + 0.006988043896853924, + 0.0053840321488678455, + 0.004023803863674402, + 0.004942549858242273, + 0.004865607246756554, + 0.0037554092705249786, + 0.0061870417557656765, + 0.007772865705192089, + 0.0031779541168361902, + 0.0032925570849329233, + 0.003135210834443569, + 0.004168028011918068, + 0.007486890535801649, + 0.002359498990699649, + 0.0031136563047766685, + 0.0026872751768678427, + 0.0019919280894100666, + 0.0021523027680814266, + 0.0019111400470137596, + 0.0023387472610920668, + 0.0021481653675436974, + 0.0024027840700000525, + 0.004140953067690134, + 0.010154273360967636, + 0.0057484908029437065, + 0.010197101160883904, + 0.004218081012368202, + 0.008390349335968494, + 0.009146720170974731, + 0.004445912782102823, + 0.006587827578186989, + 0.006038106977939606, + 0.0020553655922412872, + 0.001987589057534933, + 0.0014125428861007094, + 0.002067209454253316, + 0.0083352355286479, + 0.001654378604143858, + 0.0016654393402859569, + 0.006344163324683905, + 0.006720051169395447, + 0.00600010622292757, + 0.010239873081445694, + 0.0070643010549247265, + 0.005052565596997738, + 0.005249023903161287, + 0.006293491460382938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01607564091682434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816025868058205 + ], + "opposite_logprob_tokens": [ + -3.529815912246704 + ], + "logprob_reward": -0.029816025868058205, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816025868058205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15712928771972656, + 0.08625596016645432, + 0.4068724513053894, + 0.08535948395729065, + 0.022711003199219704, + 0.024462029337882996, + 0.01072516106069088, + 0.05175114795565605, + 0.12496455758810043, + 0.02976883202791214 + ], + "credit_map": [ + 0.012621430680155754, + 0.0343167670071125, + 0.00984723586589098, + 0.009853717871010303, + 0.012874742969870567, + 0.050300564616918564, + 0.009130067192018032, + 0.018184777349233627, + 0.008738412521779537, + 0.025959743186831474, + 0.004154636058956385, + 0.00346656097099185, + 0.00477589899674058, + 0.029409129172563553, + 0.003317100927233696, + 0.00643448531627655, + 0.0797077864408493, + 0.15303528308868408, + 0.026714641600847244, + 0.020253369584679604, + 0.045807939022779465, + 0.033740364015102386, + 0.02633160538971424, + 0.02128150686621666, + 0.011330028995871544, + 0.007127842865884304, + 0.005826187320053577, + 0.007862872444093227, + 0.011475183069705963, + 0.010417012497782707, + 0.018760554492473602, + 0.012559815309941769, + 0.0037764301523566246, + 0.004094852600246668, + 0.0015659532509744167, + 0.00205448386259377, + 0.002227315679192543, + 0.0030144418124109507, + 0.0018740883097052574, + 0.0041034407913684845, + 0.005618528462946415, + 0.0026279001031070948, + 0.001835790928453207, + 0.0019423332996666431, + 0.0034181098453700542, + 0.005250944755971432, + 0.0016291362699121237, + 0.002139289164915681, + 0.0011545773595571518, + 0.0010203939164057374, + 0.0012360004475340247, + 0.0015178403118625283, + 0.0012520482996478677, + 0.001199426595121622, + 0.0014084929134696722, + 0.0019363820319995284, + 0.010163418017327785, + 0.0045030442997813225, + 0.006735522765666246, + 0.0041994210332632065, + 0.005619571078568697, + 0.013211561366915703, + 0.0031198933720588684, + 0.004198723938316107, + 0.03797527402639389, + 0.005784010048955679, + 0.00526479072868824, + 0.004052742384374142, + 0.006332878489047289, + 0.05279812961816788, + 0.004808336030691862, + 0.007948406040668488, + 0.0047282264567911625, + 0.004695332143455744, + 0.0031365593895316124, + 0.0037395565304905176, + 0.004566066898405552, + 0.0031078506726771593, + 0.0029756175354123116, + 0.0028196251951158047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.029816025868058205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328236773610115 + ], + "opposite_logprob_tokens": [ + -3.773282289505005 + ], + "logprob_reward": -0.02328236773610115, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328236773610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1912543922662735, + 0.15307500958442688, + 0.20741015672683716, + 0.024208806455135345, + 0.04346327856183052, + 0.08253786712884903, + 0.03392135351896286, + 0.14919860661029816, + 0.049325983971357346, + 0.0656045600771904 + ], + "credit_map": [ + 0.019773831591010094, + 0.0451248437166214, + 0.014679248444736004, + 0.01502182986587286, + 0.011225401423871517, + 0.05002480372786522, + 0.013486206531524658, + 0.021918224170804024, + 0.02001776732504368, + 0.030775535851716995, + 0.011137201450765133, + 0.005148453172296286, + 0.008880759589374065, + 0.059745822101831436, + 0.005737157538533211, + 0.011632307432591915, + 0.040140677243471146, + 0.07739139348268509, + 0.01362354401499033, + 0.010564063675701618, + 0.023711953312158585, + 0.016901783645153046, + 0.014055688865482807, + 0.011021058075129986, + 0.00429189670830965, + 0.0019695463124662638, + 0.0019064630614593625, + 0.002495200140401721, + 0.0021249777637422085, + 0.00278598559089005, + 0.002883662935346365, + 0.005751072894781828, + 0.007906193844974041, + 0.00569721357896924, + 0.002662152051925659, + 0.0034041800536215305, + 0.0067471847869455814, + 0.007987012155354023, + 0.0045364536345005035, + 0.004522890318185091, + 0.01864813268184662, + 0.006025748793035746, + 0.005082226358354092, + 0.007270270027220249, + 0.014630057848989964, + 0.01648748107254505, + 0.006584216374903917, + 0.007809736765921116, + 0.004247717559337616, + 0.003655419684946537, + 0.00349462334997952, + 0.00425723334774375, + 0.005746081937104464, + 0.0037143523804843426, + 0.004200449213385582, + 0.004605478141456842, + 0.03790660947561264, + 0.009435133077204227, + 0.00973441917449236, + 0.008237223140895367, + 0.006514363922178745, + 0.0569516196846962, + 0.00789723265916109, + 0.012521998025476933, + 0.006943638436496258, + 0.003994670696556568, + 0.008761581033468246, + 0.0059442780911922455, + 0.005878035444766283, + 0.008140466175973415, + 0.005138042848557234, + 0.004525269381701946, + 0.015195096842944622, + 0.0151688102632761, + 0.005046878941357136, + 0.005629429593682289, + 0.008294408209621906, + 0.006527372635900974, + 0.004633925389498472, + 0.005108639132231474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02328236773610115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, full-body photograph captures the intense courtship ritual between two serrated river crabs, their bodies rendered in hyperrealistic CGI, highlighting every minute texture and damp sheen. They lie amidst moss-covered rocks in a clear, shallow stream, dappled sunlight filtering through the unseen forest canopy, casting a dancing, corroded pattern across the scene. One freshwater crab, its carapace a complex mosaic of dark brown and pale ochre, cautiously approaches the other with slow, deliberate sideways swoops. Its legs carefully rest on the velvety green moss clinging to the damp river rocks. Steadily and deliberately, it extends its large right claw, its claw slightly open, in a subtle, non-threatening display. The second crab remains motionless, like a cautious, observing statue, its eyestalks intently tracking the suitor's advances, its own claws firmly tucked against its body. Clear water flows gently around them, revealing the sandy, pebbly riverbed beneath, creating an atmosphere of palpable tranquility, a natural tension where the raw drama unfolds with lifelike precision.", + "seed": 17319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_2/index_2245_seed_17319_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": -0.1540520556271076, + "vlm_logprob": -0.1540520556271076, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.1540520556271076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6603694558143616 + ], + "opposite_logprob_tokens": [ + -0.7853694558143616 + ], + "logprob_reward": -0.6603694558143616, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6603694558143616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1505099982023239, + 0.036562174558639526, + 0.03741448000073433, + 0.12115735560655594, + 0.060876768082380295, + 0.11220218241214752, + 0.09718112647533417, + 0.026821117848157883, + 0.06620416790246964, + 0.2910706102848053 + ], + "credit_map": [ + 0.02034064568579197, + 0.016905412077903748, + 0.01477294135838747, + 0.018667947500944138, + 0.020978864282369614, + 0.01670733280479908, + 0.02300349250435829, + 0.01913335546851158, + 0.004087695386260748, + 0.0041108326986432076, + 0.004195377696305513, + 0.004431184381246567, + 0.004592055454850197, + 0.0042307027615606785, + 0.006529791280627251, + 0.004384537693113089, + 0.003356655826792121, + 0.0035729450173676014, + 0.0038422176148742437, + 0.003767941379919648, + 0.00487630907446146, + 0.007226127665489912, + 0.004848438780754805, + 0.005923843942582607, + 0.008082852698862553, + 0.008857413195073605, + 0.006970892660319805, + 0.00900217704474926, + 0.036808717995882034, + 0.012608003802597523, + 0.02627122960984707, + 0.012556069530546665, + 0.0065732174552977085, + 0.0075679318979382515, + 0.006782666780054569, + 0.004139278084039688, + 0.011836009100079536, + 0.011447332799434662, + 0.007293103728443384, + 0.005237228702753782, + 0.01158212497830391, + 0.01844109781086445, + 0.01289985328912735, + 0.012548946775496006, + 0.01475189346820116, + 0.01642991602420807, + 0.0139838308095932, + 0.011564522050321102, + 0.011818330734968185, + 0.014869078062474728, + 0.010781880468130112, + 0.018692288547754288, + 0.010544340126216412, + 0.00816468708217144, + 0.01420440711081028, + 0.008106113411486149, + 0.002195351989939809, + 0.003687084186822176, + 0.002274999860674143, + 0.002458422677591443, + 0.0034344582818448544, + 0.0038761005271226168, + 0.004109558649361134, + 0.004785140976309776, + 0.005381835624575615, + 0.00881333090364933, + 0.007920881733298302, + 0.006548731587827206, + 0.006953386589884758, + 0.010865932330489159, + 0.01191958412528038, + 0.007800478488206863, + 0.01373294461518526, + 0.014020213857293129, + 0.00748438062146306, + 0.009105074219405651, + 0.10705576092004776, + 0.023867741227149963, + 0.024981383234262466, + 0.09082310646772385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.6603694558143616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026388922706246376 + ], + "opposite_logprob_tokens": [ + -3.6513888835906982 + ], + "logprob_reward": -0.026388922706246376, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026388922706246376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11654074490070343, + 0.034131791442632675, + 0.055215898901224136, + 0.11196039617061615, + 0.148639976978302, + 0.07502105087041855, + 0.04730239510536194, + 0.07956579327583313, + 0.05865824967622757, + 0.27296367287635803 + ], + "credit_map": [ + 0.015480230562388897, + 0.012177609838545322, + 0.010770324617624283, + 0.015960421413183212, + 0.015860743820667267, + 0.011823886074125767, + 0.020255515351891518, + 0.014212020672857761, + 0.004530038218945265, + 0.0032915635965764523, + 0.002974143484607339, + 0.0028203576803207397, + 0.005744329188019037, + 0.005287657026201487, + 0.0053976369090378284, + 0.004086069297045469, + 0.005947846919298172, + 0.005722799338400364, + 0.005265811458230019, + 0.00441052345559001, + 0.013974284753203392, + 0.007639612536877394, + 0.00607242900878191, + 0.0061825974844396114, + 0.007329422980546951, + 0.009064054116606712, + 0.005963376257568598, + 0.00681665726006031, + 0.0374574139714241, + 0.008780987933278084, + 0.021185938268899918, + 0.015362548641860485, + 0.01176911685615778, + 0.017211943864822388, + 0.01885536126792431, + 0.013761663809418678, + 0.024846643209457397, + 0.02709731087088585, + 0.021382182836532593, + 0.013715765438973904, + 0.007171576842665672, + 0.00930952001363039, + 0.009092210792005062, + 0.008192467503249645, + 0.013337057083845139, + 0.009136192500591278, + 0.010942592285573483, + 0.007839434780180454, + 0.0044731865637004375, + 0.007061631418764591, + 0.005377870053052902, + 0.006971215829253197, + 0.00508894631639123, + 0.007035920862108469, + 0.006441102828830481, + 0.004852522164583206, + 0.007578976918011904, + 0.009563989005982876, + 0.009626904502511024, + 0.009192368015646935, + 0.008876427076756954, + 0.012478155083954334, + 0.010000862181186676, + 0.012248114682734013, + 0.006388792768120766, + 0.007386001292616129, + 0.004863892216235399, + 0.00581320421770215, + 0.0063569266349077225, + 0.01055047195404768, + 0.009959983639419079, + 0.007338980678468943, + 0.016207650303840637, + 0.023542014881968498, + 0.009728840552270412, + 0.01074513141065836, + 0.09102631360292435, + 0.02627944014966488, + 0.0207972452044487, + 0.07463706284761429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.026388922706246376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081036999821663 + ], + "opposite_logprob_tokens": [ + -4.141080856323242 + ], + "logprob_reward": -0.016081036999821663, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081036999821663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10114863514900208, + 0.029283907264471054, + 0.07304280251264572, + 0.11981033533811569, + 0.18541359901428223, + 0.05819176137447357, + 0.08243414759635925, + 0.13350781798362732, + 0.09064805507659912, + 0.12651897966861725 + ], + "credit_map": [ + 0.012575571425259113, + 0.011535735800862312, + 0.011353016830980778, + 0.011053993366658688, + 0.014803297817707062, + 0.013276604935526848, + 0.01404237188398838, + 0.012508041225373745, + 0.003573479363694787, + 0.002476293360814452, + 0.0033080733846873045, + 0.0023978976532816887, + 0.004382534418255091, + 0.0052872709929943085, + 0.004010841250419617, + 0.003847516141831875, + 0.008189117535948753, + 0.005571596324443817, + 0.006496062967926264, + 0.006182998884469271, + 0.008031601086258888, + 0.012936708517372608, + 0.009365681558847427, + 0.016269037500023842, + 0.012050102464854717, + 0.011314637959003448, + 0.006824361626058817, + 0.007356515619903803, + 0.03616960346698761, + 0.009369478560984135, + 0.02266237884759903, + 0.014063259586691856, + 0.012878859415650368, + 0.022748248651623726, + 0.0195030365139246, + 0.01422137301415205, + 0.044828470796346664, + 0.03746741637587547, + 0.01856434904038906, + 0.015201844274997711, + 0.006444923579692841, + 0.008112090639770031, + 0.007696770131587982, + 0.005994526203721762, + 0.00885852426290512, + 0.0069277361035346985, + 0.007518178317695856, + 0.0066390144638717175, + 0.008503775112330914, + 0.01235705055296421, + 0.014814464375376701, + 0.010710072703659534, + 0.009894291870296001, + 0.007571201305836439, + 0.009908141568303108, + 0.008675153367221355, + 0.011895513162016869, + 0.01617189683020115, + 0.015223534777760506, + 0.015326154418289661, + 0.017932942137122154, + 0.022510917857289314, + 0.014826118014752865, + 0.019620737060904503, + 0.006713770795613527, + 0.013538659550249577, + 0.008154611103236675, + 0.007677392102777958, + 0.008402585983276367, + 0.021672554314136505, + 0.01396577712148428, + 0.010522700846195221, + 0.008010628633201122, + 0.016432475298643112, + 0.005985768046230078, + 0.006575531326234341, + 0.03837990388274193, + 0.011903897859156132, + 0.009076236747205257, + 0.030154548585414886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.016081036999821663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033725760877132416 + ], + "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": [ + 0.12999196350574493, + 0.11655406653881073, + 0.1063648983836174, + 0.2188473492860794, + 0.06822559982538223, + 0.061410918831825256, + 0.08356726169586182, + 0.054418571293354034, + 0.06603766232728958, + 0.09458168596029282 + ], + "credit_map": [ + 0.010997386649250984, + 0.020615197718143463, + 0.012589816935360432, + 0.01859108731150627, + 0.016991522163152695, + 0.016909558326005936, + 0.02086281217634678, + 0.012434583157300949, + 0.009330240078270435, + 0.007308358792215586, + 0.008169270120561123, + 0.010283777490258217, + 0.015849540010094643, + 0.028081702068448067, + 0.020640775561332703, + 0.016890399158000946, + 0.008258496411144733, + 0.00760298827663064, + 0.007247330620884895, + 0.007999727502465248, + 0.020589325577020645, + 0.014783911406993866, + 0.014921996742486954, + 0.024961136281490326, + 0.01705355755984783, + 0.021304015070199966, + 0.014895566739141941, + 0.015229648910462856, + 0.06816396862268448, + 0.017803389579057693, + 0.02626103162765503, + 0.038136180490255356, + 0.0074928211979568005, + 0.007851322181522846, + 0.009988575242459774, + 0.006181119941174984, + 0.011036555282771587, + 0.00990334339439869, + 0.008638975210487843, + 0.007132888771593571, + 0.005644689314067364, + 0.0065885065123438835, + 0.009128210134804249, + 0.006713365204632282, + 0.007535886950790882, + 0.008822117932140827, + 0.010158319026231766, + 0.006819824688136578, + 0.011298955418169498, + 0.015567099675536156, + 0.009555723518133163, + 0.010590454563498497, + 0.011736096814274788, + 0.008390529081225395, + 0.008311460725963116, + 0.008116942830383778, + 0.005349765531718731, + 0.0063935392536222935, + 0.005003079306334257, + 0.0050611295737326145, + 0.00667886296287179, + 0.009466957300901413, + 0.007676857057958841, + 0.008788385428488255, + 0.009571170434355736, + 0.008733130060136318, + 0.006161799654364586, + 0.006034140940755606, + 0.008173516020178795, + 0.008480080403387547, + 0.01041022315621376, + 0.008473598398268223, + 0.00633180933073163, + 0.016948599368333817, + 0.0068094185553491116, + 0.00765970628708601, + 0.018014905974268913, + 0.015758421272039413, + 0.010179558768868446, + 0.0128792654722929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.033725760877132416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695101737976074 + ], + "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": [ + 0.1437845230102539, + 0.039119914174079895, + 0.041173458099365234, + 0.28801947832107544, + 0.09079010039567947, + 0.042170360684394836, + 0.1536955088376999, + 0.033059362322092056, + 0.06708160787820816, + 0.10110579431056976 + ], + "credit_map": [ + 0.01418597437441349, + 0.01500994898378849, + 0.012577598914504051, + 0.018561163917183876, + 0.018308714032173157, + 0.01683664880692959, + 0.027933191508054733, + 0.020371267572045326, + 0.005446886643767357, + 0.0037685066927224398, + 0.004426990635693073, + 0.0031258065719157457, + 0.00661268038675189, + 0.004562695510685444, + 0.005829897243529558, + 0.005346445366740227, + 0.005072867497801781, + 0.0031665253918617964, + 0.0037574986927211285, + 0.003310005646198988, + 0.008660840801894665, + 0.0048875040374696255, + 0.005436387378722429, + 0.006881823297590017, + 0.020678622648119926, + 0.02554876171052456, + 0.014457708224654198, + 0.025828892365098, + 0.050724681466817856, + 0.037187833338975906, + 0.08411610871553421, + 0.029476838186383247, + 0.00748099759221077, + 0.008709395304322243, + 0.010707467794418335, + 0.008560572750866413, + 0.012339690700173378, + 0.013267727568745613, + 0.019624484702944756, + 0.010099751874804497, + 0.0035830221604555845, + 0.004443054087460041, + 0.00428241491317749, + 0.004870500415563583, + 0.008277458138763905, + 0.006276893895119429, + 0.005858288612216711, + 0.004578720778226852, + 0.014953967183828354, + 0.022651394829154015, + 0.017400942742824554, + 0.03105216845870018, + 0.014392789453268051, + 0.01355411484837532, + 0.025538308545947075, + 0.014151805080473423, + 0.0023637465201318264, + 0.0031617581844329834, + 0.002421442186459899, + 0.0032365440856665373, + 0.004287666641175747, + 0.004548670258373022, + 0.008993176743388176, + 0.0040463535115122795, + 0.0037210704758763313, + 0.00675270427018404, + 0.005921144038438797, + 0.004217806737869978, + 0.007548187393695116, + 0.018750006332993507, + 0.012535559013485909, + 0.007635115645825863, + 0.00468111876398325, + 0.0058448426425457, + 0.004137128125876188, + 0.003325567813590169, + 0.03466258570551872, + 0.00927051529288292, + 0.008616473525762558, + 0.030567551031708717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.033695101737976074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, full-body photograph captures the intense courtship ritual between two serrated river crabs, their bodies rendered in hyperrealistic CGI, highlighting every minute texture and damp sheen. They lie amidst moss-covered rocks in a clear, shallow stream, dappled sunlight filtering through the unseen forest canopy, casting a dancing, corroded pattern across the scene. One freshwater crab, its carapace a complex mosaic of dark brown and pale ochre, cautiously approaches the other with slow, deliberate sideways swoops. Its legs carefully rest on the velvety green moss clinging to the damp river rocks. Steadily and deliberately, it extends its large right claw, its claw slightly open, in a subtle, non-threatening display. The second crab remains motionless, like a cautious, observing statue, its eyestalks intently tracking the suitor's advances, its own claws firmly tucked against its body. Clear water flows gently around them, revealing the sandy, pebbly riverbed beneath, creating an atmosphere of palpable tranquility, a natural tension where the raw drama unfolds with lifelike precision.", + "seed": 17320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_2/index_2245_seed_17320_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": -0.0534260680899024, + "vlm_logprob": -0.0534260680899024, + "vlm_yes_no_margin": 3.424999976158142, + "vlm_token_credit": -0.0534260680899024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1634455919265747 + ], + "opposite_logprob_tokens": [ + -2.163445472717285 + ], + "logprob_reward": -0.1634455919265747, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1634455919265747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.168285071849823, + 0.05816575884819031, + 0.21549607813358307, + 0.05995781719684601, + 0.1529829353094101, + 0.04676422104239464, + 0.05959621071815491, + 0.10976842790842056, + 0.06791689991950989, + 0.06106666103005409 + ], + "credit_map": [ + 0.01907435618340969, + 0.024347223341464996, + 0.022001225501298904, + 0.018179060891270638, + 0.021326187998056412, + 0.018891775980591774, + 0.027128830552101135, + 0.017336392775177956, + 0.010423955507576466, + 0.006140663288533688, + 0.007047242950648069, + 0.006069579161703587, + 0.007678770925849676, + 0.007180697750300169, + 0.006810649298131466, + 0.006814193911850452, + 0.025758318603038788, + 0.01747601106762886, + 0.04399707913398743, + 0.019633516669273376, + 0.024065546691417694, + 0.03128965198993683, + 0.03155146539211273, + 0.021724458783864975, + 0.009289246052503586, + 0.006240266840904951, + 0.008547178469598293, + 0.010211301036179066, + 0.007650326006114483, + 0.00464979512616992, + 0.007760105188935995, + 0.005609592888504267, + 0.032601408660411835, + 0.007692248094826937, + 0.015887461602687836, + 0.014583323150873184, + 0.0076765199191868305, + 0.028247695416212082, + 0.014238514937460423, + 0.032055746763944626, + 0.007643763907253742, + 0.004039003513753414, + 0.009246346540749073, + 0.0036339748185127974, + 0.005182209890335798, + 0.005926669575273991, + 0.004617761820554733, + 0.006474488414824009, + 0.008988690562546253, + 0.007783675566315651, + 0.007706529460847378, + 0.00619112653657794, + 0.006551290862262249, + 0.007756329607218504, + 0.007699020206928253, + 0.0069195483811199665, + 0.01451774314045906, + 0.009119945578277111, + 0.012225266546010971, + 0.01330093014985323, + 0.0164271742105484, + 0.020801279693841934, + 0.010905969887971878, + 0.012470104731619358, + 0.009479796513915062, + 0.006132110953330994, + 0.010117940604686737, + 0.008200835436582565, + 0.008737991563975811, + 0.009717169217765331, + 0.009377076290547848, + 0.00615397235378623, + 0.00805512722581625, + 0.005078068934381008, + 0.006900382693856955, + 0.007026819046586752, + 0.01162020955234766, + 0.007568261120468378, + 0.008530602790415287, + 0.00628718500956893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.1634455919265747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016101565212011337 + ], + "opposite_logprob_tokens": [ + -4.141101360321045 + ], + "logprob_reward": -0.016101565212011337, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016101565212011337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14047716557979584, + 0.13760828971862793, + 0.13644787669181824, + 0.09100684523582458, + 0.11222738027572632, + 0.13306942582130432, + 0.06470867991447449, + 0.08733761310577393, + 0.033001963049173355, + 0.06411472707986832 + ], + "credit_map": [ + 0.016246400773525238, + 0.015689508989453316, + 0.020236697047948837, + 0.027687368914484978, + 0.012819776311516762, + 0.016525037586688995, + 0.01608903333544731, + 0.01518335472792387, + 0.015265689231455326, + 0.015089021064341068, + 0.02225377783179283, + 0.012929321266710758, + 0.01787898875772953, + 0.016224980354309082, + 0.02177434042096138, + 0.016192179173231125, + 0.014075829647481441, + 0.013514252379536629, + 0.023077892139554024, + 0.014561149291694164, + 0.020619487389922142, + 0.01752477139234543, + 0.02049603872001171, + 0.01257846225053072, + 0.008433876559138298, + 0.008509292267262936, + 0.014993082731962204, + 0.016737110912799835, + 0.011819923296570778, + 0.00893707200884819, + 0.013467528857290745, + 0.008108960464596748, + 0.0236631091684103, + 0.007340169511735439, + 0.011637814342975616, + 0.01597273163497448, + 0.007206652779132128, + 0.009701740927994251, + 0.011775209568440914, + 0.024929963052272797, + 0.02006583660840988, + 0.01966066285967827, + 0.014834224246442318, + 0.015190250240266323, + 0.012098977342247963, + 0.011757353320717812, + 0.01770205609500408, + 0.021760066971182823, + 0.008253989741206169, + 0.005495510995388031, + 0.010815691202878952, + 0.00721706310287118, + 0.005363167729228735, + 0.008075001649558544, + 0.011949594132602215, + 0.0075386702083051205, + 0.011101843789219856, + 0.008552691899240017, + 0.014985514804720879, + 0.011241866275668144, + 0.009367712773382664, + 0.012983987107872963, + 0.010262708179652691, + 0.008841292932629585, + 0.005592992063611746, + 0.0029314563143998384, + 0.00495848897844553, + 0.003601580858230591, + 0.004153277724981308, + 0.004103755112737417, + 0.00417693704366684, + 0.003483477747067809, + 0.008369739167392254, + 0.005380324088037014, + 0.009450029581785202, + 0.007188885472714901, + 0.009875417686998844, + 0.009263775311410427, + 0.007534329779446125, + 0.007052228320389986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016101565212011337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636430785059929 + ], + "opposite_logprob_tokens": [ + -3.651364326477051 + ], + "logprob_reward": -0.02636430785059929, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636430785059929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14733749628067017, + 0.06886228919029236, + 0.08360736817121506, + 0.04768233746290207, + 0.14945578575134277, + 0.1668599545955658, + 0.0969739556312561, + 0.13151489198207855, + 0.02828926220536232, + 0.0794166699051857 + ], + "credit_map": [ + 0.01832234486937523, + 0.012338804081082344, + 0.02142888307571411, + 0.019817808642983437, + 0.016141602769494057, + 0.021566642448306084, + 0.02158663421869278, + 0.01613476872444153, + 0.008732277899980545, + 0.0052802287973463535, + 0.014427214860916138, + 0.005743545014411211, + 0.009693557396531105, + 0.006561278365552425, + 0.01011257991194725, + 0.008311599493026733, + 0.013174545019865036, + 0.006032994482666254, + 0.015562763437628746, + 0.01005137525498867, + 0.008103449828922749, + 0.01109310518950224, + 0.010126863606274128, + 0.009462266229093075, + 0.005736491177231073, + 0.002756422618404031, + 0.007161984220147133, + 0.0059317853301763535, + 0.004185626283288002, + 0.0093306889757514, + 0.007853085175156593, + 0.0047262487933039665, + 0.026858877390623093, + 0.01232359278947115, + 0.01438857987523079, + 0.02071547321975231, + 0.015477387234568596, + 0.012561951763927937, + 0.018289713189005852, + 0.02884020283818245, + 0.019965387880802155, + 0.02355176955461502, + 0.021226102486252785, + 0.02084391377866268, + 0.01688527502119541, + 0.0169609896838665, + 0.024494625627994537, + 0.02293187379837036, + 0.013204731978476048, + 0.008604157716035843, + 0.015123231336474419, + 0.011472539976239204, + 0.007446476258337498, + 0.01066883560270071, + 0.01798984967172146, + 0.012464127503335476, + 0.014415879733860493, + 0.014527682214975357, + 0.024462617933750153, + 0.019012629985809326, + 0.01290775928646326, + 0.015765048563480377, + 0.017319325357675552, + 0.013103943318128586, + 0.00428621144965291, + 0.0028740086127072573, + 0.004728605039417744, + 0.0033917988184839487, + 0.003323083743453026, + 0.0028593388851732016, + 0.004086592234671116, + 0.002739620627835393, + 0.009977178648114204, + 0.007178342901170254, + 0.012236462906002998, + 0.01565827615559101, + 0.00982123427093029, + 0.0064943283796310425, + 0.009603794664144516, + 0.00844704732298851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.02636430785059929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048683051019907 + ], + "opposite_logprob_tokens": [ + -3.0486831665039062 + ], + "logprob_reward": -0.048683051019907, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048683051019907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1264360100030899, + 0.13770323991775513, + 0.1929023563861847, + 0.044911570847034454, + 0.15146473050117493, + 0.10404442250728607, + 0.05001559853553772, + 0.05968482047319412, + 0.036241352558135986, + 0.09659586101770401 + ], + "credit_map": [ + 0.016312243416905403, + 0.014025622978806496, + 0.01807720586657524, + 0.01855911687016487, + 0.015086743049323559, + 0.016817864030599594, + 0.012518311850726604, + 0.015038921497762203, + 0.02119787596166134, + 0.011777620762586594, + 0.021057700738310814, + 0.013883650302886963, + 0.01754981279373169, + 0.02250342071056366, + 0.015378624200820923, + 0.014354540966451168, + 0.02120940200984478, + 0.01687287539243698, + 0.03207317367196083, + 0.0199988204985857, + 0.025134816765785217, + 0.029029695317149162, + 0.029387809336185455, + 0.019195763394236565, + 0.005137674044817686, + 0.003328845137730241, + 0.008107482455670834, + 0.0054221246391534805, + 0.0063555617816746235, + 0.004943819250911474, + 0.006817714776843786, + 0.004798345733433962, + 0.04435165598988533, + 0.008864144794642925, + 0.01260502077639103, + 0.012248159386217594, + 0.00738038495182991, + 0.01138878334313631, + 0.011860357597470284, + 0.042766228318214417, + 0.019419005140662193, + 0.007762915920466185, + 0.01762448064982891, + 0.009430713951587677, + 0.009228304028511047, + 0.013231415301561356, + 0.008707687258720398, + 0.018639899790287018, + 0.005975399166345596, + 0.004535474348813295, + 0.007451483514159918, + 0.0048928954638540745, + 0.00436550285667181, + 0.0068601262755692005, + 0.009738816879689693, + 0.006195896305143833, + 0.0071387081407010555, + 0.005809538532048464, + 0.01395623479038477, + 0.006312535610049963, + 0.007461260072886944, + 0.005491213407367468, + 0.007264110259711742, + 0.0062512182630598545, + 0.005412707105278969, + 0.0036610157694667578, + 0.006121363025158644, + 0.003697221865877509, + 0.003911690320819616, + 0.006052733398973942, + 0.004088359884917736, + 0.003296259557828307, + 0.013095488771796227, + 0.008939133025705814, + 0.01585203967988491, + 0.015485581010580063, + 0.015302951447665691, + 0.009407649748027325, + 0.008485802449285984, + 0.010027209296822548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.048683051019907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535824440419674 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535824440419674, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535824440419674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09060832858085632, + 0.07904137670993805, + 0.0967477336525917, + 0.030372638255357742, + 0.23342376947402954, + 0.031318359076976776, + 0.034948885440826416, + 0.12557263672351837, + 0.12935233116149902, + 0.14861392974853516 + ], + "credit_map": [ + 0.011945954523980618, + 0.0093747079372406, + 0.011263960972428322, + 0.011055348441004753, + 0.008820944465696812, + 0.01590273156762123, + 0.011378968134522438, + 0.010865716263651848, + 0.010713275521993637, + 0.005947541445493698, + 0.020642129704356194, + 0.009167684242129326, + 0.006535873282700777, + 0.007532757241278887, + 0.009317666292190552, + 0.009184455499053001, + 0.013659256510436535, + 0.006576858926564455, + 0.028403636068105698, + 0.0076592047698795795, + 0.009362392127513885, + 0.013132067397236824, + 0.006997647229582071, + 0.010956679470837116, + 0.0033974256366491318, + 0.0022312686778604984, + 0.006050782278180122, + 0.003569191787391901, + 0.0032480591908097267, + 0.005306778475642204, + 0.0036413369234651327, + 0.0029277964495122433, + 0.05986333265900612, + 0.012969433329999447, + 0.014391940087080002, + 0.011150974780321121, + 0.02533806301653385, + 0.03138937056064606, + 0.01672571897506714, + 0.06159496679902077, + 0.006699512712657452, + 0.0031286210287362337, + 0.004804504103958607, + 0.0024428514298051596, + 0.002111733891069889, + 0.003613455919548869, + 0.002724549500271678, + 0.005793131422251463, + 0.004256955347955227, + 0.002833572681993246, + 0.0050058066844940186, + 0.00458957813680172, + 0.004282387439161539, + 0.00453026220202446, + 0.004864759277552366, + 0.004585569724440575, + 0.017564760521054268, + 0.011587994173169136, + 0.021578721702098846, + 0.00960083119571209, + 0.01492861844599247, + 0.02442266419529915, + 0.013144496828317642, + 0.012744557112455368, + 0.014470328576862812, + 0.009522453881800175, + 0.03019179403781891, + 0.019391745328903198, + 0.011040587909519672, + 0.01458221860229969, + 0.016633523628115654, + 0.013519682921469212, + 0.016467256471514702, + 0.014434529468417168, + 0.01542602013796568, + 0.028379617258906364, + 0.020299207419157028, + 0.018903734162449837, + 0.018681008368730545, + 0.01602255180478096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.012535824440419674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, full-body photograph captures the intense courtship ritual between two serrated river crabs, their bodies rendered in hyperrealistic CGI, highlighting every minute texture and damp sheen. They lie amidst moss-covered rocks in a clear, shallow stream, dappled sunlight filtering through the unseen forest canopy, casting a dancing, corroded pattern across the scene. One freshwater crab, its carapace a complex mosaic of dark brown and pale ochre, cautiously approaches the other with slow, deliberate sideways swoops. Its legs carefully rest on the velvety green moss clinging to the damp river rocks. Steadily and deliberately, it extends its large right claw, its claw slightly open, in a subtle, non-threatening display. The second crab remains motionless, like a cautious, observing statue, its eyestalks intently tracking the suitor's advances, its own claws firmly tucked against its body. Clear water flows gently around them, revealing the sandy, pebbly riverbed beneath, creating an atmosphere of palpable tranquility, a natural tension where the raw drama unfolds with lifelike precision.", + "seed": 17321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_2/index_2245_seed_17321_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": -0.08554193302989006, + "vlm_logprob": -0.08554193302989006, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.08554193302989006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.27619680762290955 + ], + "opposite_logprob_tokens": [ + -1.526196837425232 + ], + "logprob_reward": -0.27619680762290955, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.27619680762290955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07435882091522217, + 0.2364794760942459, + 0.07865135371685028, + 0.07410868257284164, + 0.1357637196779251, + 0.027308247983455658, + 0.017933452501893044, + 0.3284212350845337, + 0.011341377161443233, + 0.015633687376976013 + ], + "credit_map": [ + 0.007126685697585344, + 0.013073483482003212, + 0.005925596691668034, + 0.008896801620721817, + 0.010252431966364384, + 0.012418902479112148, + 0.007916809059679508, + 0.00874810479581356, + 0.007550622336566448, + 0.005686020012944937, + 0.0035992383491247892, + 0.00790189765393734, + 0.10900401324033737, + 0.08385258167982101, + 0.00819263607263565, + 0.010692465119063854, + 0.004758289083838463, + 0.007939008995890617, + 0.006693646311759949, + 0.011959974654018879, + 0.009213246405124664, + 0.02204733155667782, + 0.008775323629379272, + 0.007264531683176756, + 0.0066510955803096294, + 0.014356800355017185, + 0.00718996487557888, + 0.012071733362972736, + 0.008881088346242905, + 0.006616448052227497, + 0.00802307017147541, + 0.010318479500710964, + 0.006559669505804777, + 0.004826703108847141, + 0.00585572374984622, + 0.007559351157397032, + 0.01166883111000061, + 0.041131045669317245, + 0.04823414981365204, + 0.009928242303431034, + 0.0023150029592216015, + 0.003921117167919874, + 0.0024561635218560696, + 0.004595725331455469, + 0.0028886967338621616, + 0.0038123487029224634, + 0.004063664004206657, + 0.003255530260503292, + 0.002209549769759178, + 0.0017850997392088175, + 0.0015205774689093232, + 0.0018944451585412025, + 0.0019962817896157503, + 0.0018525595078244805, + 0.0039861733093857765, + 0.0026887673884630203, + 0.020481647923588753, + 0.009929414838552475, + 0.01910216733813286, + 0.020012695342302322, + 0.023849697783589363, + 0.10352908819913864, + 0.10896581411361694, + 0.022550715133547783, + 0.0022890407126396894, + 0.0011397266061976552, + 0.0013365326449275017, + 0.0009969596285372972, + 0.0012478540884330869, + 0.00162760098464787, + 0.0008718911558389664, + 0.0018317720387130976, + 0.002153858309611678, + 0.001676725223660469, + 0.0021276071202009916, + 0.001432752818800509, + 0.0018443629378452897, + 0.0021301177330315113, + 0.001453269855119288, + 0.0028149946592748165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.27619680762290955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018225114792585373 + ], + "opposite_logprob_tokens": [ + -4.018225193023682 + ], + "logprob_reward": -0.018225114792585373, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018225114792585373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10642058402299881, + 0.13959476351737976, + 0.04409392178058624, + 0.05812922120094299, + 0.23956617712974548, + 0.07620575278997421, + 0.045071035623550415, + 0.2214050441980362, + 0.016986608505249023, + 0.05252690613269806 + ], + "credit_map": [ + 0.009037371724843979, + 0.015775611624121666, + 0.010435025207698345, + 0.01437943521887064, + 0.013944958336651325, + 0.014795572496950626, + 0.015233505517244339, + 0.012819097377359867, + 0.00406067818403244, + 0.007286191917955875, + 0.003536112140864134, + 0.009272334165871143, + 0.049318116158246994, + 0.045288119465112686, + 0.007312591653317213, + 0.013520612381398678, + 0.0031478297896683216, + 0.0036241980269551277, + 0.005160057917237282, + 0.006112599745392799, + 0.007335982285439968, + 0.00915540475398302, + 0.003847071435302496, + 0.005710776429623365, + 0.003610410960391164, + 0.004595300182700157, + 0.002894214354455471, + 0.005946319550275803, + 0.01659875549376011, + 0.006826790980994701, + 0.01042879931628704, + 0.007228626403957605, + 0.011094632558524609, + 0.009227055124938488, + 0.010463308542966843, + 0.013268922455608845, + 0.021163782104849815, + 0.07321550697088242, + 0.0844685509800911, + 0.016664404422044754, + 0.005611303728073835, + 0.012180225923657417, + 0.005060269497334957, + 0.010301663540303707, + 0.008405877277255058, + 0.012992244213819504, + 0.010985037311911583, + 0.010669122450053692, + 0.004897735081613064, + 0.004831654950976372, + 0.005953716579824686, + 0.006147258915007114, + 0.004310418386012316, + 0.00636273343116045, + 0.006403453182429075, + 0.006164063233882189, + 0.01864999160170555, + 0.012574497610330582, + 0.017666740342974663, + 0.013343552127480507, + 0.017539946362376213, + 0.05958680808544159, + 0.06405989825725555, + 0.017983591184020042, + 0.0023755691945552826, + 0.0020908599253743887, + 0.0013898807810619473, + 0.0017788902623578906, + 0.0021729052532464266, + 0.0023849268909543753, + 0.0021802254486829042, + 0.0026133486535400152, + 0.0059741102159023285, + 0.00783119723200798, + 0.0076856412924826145, + 0.006117424927651882, + 0.0056037288159132, + 0.008479099720716476, + 0.005252080969512463, + 0.005583616904914379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.018225114792585373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590949803590775 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020590949803590775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590949803590775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056450821459293365, + 0.22156523168087006, + 0.13238726556301117, + 0.06586292386054993, + 0.028488965705037117, + 0.028456587344408035, + 0.028651710599660873, + 0.39642825722694397, + 0.013530182652175426, + 0.028178051114082336 + ], + "credit_map": [ + 0.004788744263350964, + 0.011191096156835556, + 0.005870083346962929, + 0.007784914691001177, + 0.006959011312574148, + 0.006174246780574322, + 0.005141087342053652, + 0.008541636168956757, + 0.00713872816413641, + 0.005730874836444855, + 0.004339541308581829, + 0.007221312262117863, + 0.09908570349216461, + 0.07685751467943192, + 0.00827208161354065, + 0.012919481843709946, + 0.00867968238890171, + 0.011572226881980896, + 0.010788456536829472, + 0.019163599237799644, + 0.016295015811920166, + 0.03944500535726547, + 0.014122816734015942, + 0.012320468202233315, + 0.0051347967237234116, + 0.009105185978114605, + 0.005052443593740463, + 0.01115309726446867, + 0.010895537212491035, + 0.005770506802946329, + 0.010098068974912167, + 0.008653291501104832, + 0.0033380533568561077, + 0.002549929777160287, + 0.0018071195809170604, + 0.003579626092687249, + 0.0036167865619063377, + 0.006108940578997135, + 0.004307599738240242, + 0.0031809101346880198, + 0.002550480654463172, + 0.004065396264195442, + 0.002880387008190155, + 0.004408132750540972, + 0.002930093789473176, + 0.003762799547985196, + 0.0031908713281154633, + 0.00466842669993639, + 0.0031882633920758963, + 0.0031274373177438974, + 0.002942522056400776, + 0.003554457798600197, + 0.003032877342775464, + 0.0036551696248352528, + 0.0051972949877381325, + 0.003953689709305763, + 0.022237908095121384, + 0.00969694647938013, + 0.023150628432631493, + 0.025834741070866585, + 0.028908537700772285, + 0.1269727498292923, + 0.13340561091899872, + 0.02622113935649395, + 0.0020656606648117304, + 0.0011740198824554682, + 0.0014471433823928237, + 0.0009643940138630569, + 0.002130518900230527, + 0.0019386970670893788, + 0.0015470715006813407, + 0.002262676600366831, + 0.0032593172509223223, + 0.00408319802954793, + 0.0028281984850764275, + 0.0027362490072846413, + 0.0031366529874503613, + 0.003781580599024892, + 0.005941093899309635, + 0.002411761786788702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.020590949803590775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033720459789037704 + ], + "opposite_logprob_tokens": [ + -3.4087204933166504 + ], + "logprob_reward": -0.033720459789037704, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033720459789037704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041885096579790115, + 0.0826975479722023, + 0.02989339455962181, + 0.07921730726957321, + 0.2196229100227356, + 0.018757537007331848, + 0.0233958438038826, + 0.4523240029811859, + 0.02601381577551365, + 0.02619258314371109 + ], + "credit_map": [ + 0.003227045526728034, + 0.007221598643809557, + 0.004482612945139408, + 0.005303496494889259, + 0.005718564614653587, + 0.0058232685551047325, + 0.00459114508703351, + 0.0055173663422465324, + 0.004648848436772823, + 0.007908714935183525, + 0.0057221148163080215, + 0.009335106238722801, + 0.019502777606248856, + 0.020944766700267792, + 0.006380428560078144, + 0.008254796266555786, + 0.002927452092990279, + 0.003633796004578471, + 0.002007340546697378, + 0.003698806744068861, + 0.006758254487067461, + 0.0036391981411725283, + 0.0027258291374891996, + 0.004502718802541494, + 0.0048356144689023495, + 0.009957613423466682, + 0.006059345789253712, + 0.00786521378904581, + 0.02136857807636261, + 0.007613830734044313, + 0.01317211240530014, + 0.008344996720552444, + 0.00992373563349247, + 0.007806265726685524, + 0.008601170964539051, + 0.011710058897733688, + 0.019083263352513313, + 0.06953846663236618, + 0.0772671103477478, + 0.01569281704723835, + 0.0020931579638272524, + 0.0017408998683094978, + 0.0019449430983513594, + 0.002254052320495248, + 0.0021342250984162092, + 0.003345798933878541, + 0.0022372996900230646, + 0.0030071609653532505, + 0.002822807291522622, + 0.002478774404153228, + 0.0022929899860173464, + 0.0026067811995744705, + 0.0029178194236010313, + 0.0033047732431441545, + 0.0039051352068781853, + 0.0030667635146528482, + 0.02510351687669754, + 0.0115080326795578, + 0.025373656302690506, + 0.029306616634130478, + 0.03316880017518997, + 0.14476744830608368, + 0.15387429296970367, + 0.029221612960100174, + 0.0029179451521486044, + 0.0032429208513349295, + 0.0026882747188210487, + 0.0024662897922098637, + 0.0042108409106731415, + 0.004322734661400318, + 0.003120338311418891, + 0.003044471377506852, + 0.0028292539063841105, + 0.003773677861317992, + 0.0025479760952293873, + 0.0028762798756361008, + 0.0029407693073153496, + 0.0030705085955560207, + 0.005965712014585733, + 0.0021884043235331774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.033720459789037704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 a static full-body photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two serrated river crabs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab approaching the other with slow, deliberate sideways movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second crab remaining motionless with its eyestalks tracking the first crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a clear, shallow stream with moss-covered rocks and dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789763331413269 + ], + "opposite_logprob_tokens": [ + -2.5789763927459717 + ], + "logprob_reward": -0.0789763331413269, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789763331413269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07133394479751587, + 0.18939520418643951, + 0.10230386257171631, + 0.08550446480512619, + 0.21480268239974976, + 0.052353568375110626, + 0.15655949711799622, + 0.08840068429708481, + 0.014193943701684475, + 0.025152182206511497 + ], + "credit_map": [ + 0.004222891293466091, + 0.011196271516382694, + 0.005959394387900829, + 0.007609406020492315, + 0.01065960805863142, + 0.010334235616028309, + 0.008989617228507996, + 0.012362510897219181, + 0.006022260058671236, + 0.008461665362119675, + 0.005832867696881294, + 0.018415598198771477, + 0.06570546329021454, + 0.05990566685795784, + 0.010483168996870518, + 0.01456849928945303, + 0.007983983494341373, + 0.0062294951640069485, + 0.00724629545584321, + 0.010312342084944248, + 0.017453927546739578, + 0.032128822058439255, + 0.009352456778287888, + 0.011596535332500935, + 0.0058255684562027454, + 0.009695813059806824, + 0.005790357012301683, + 0.012387145310640335, + 0.021237479522824287, + 0.00850003119558096, + 0.013415515422821045, + 0.008652549237012863, + 0.01957344077527523, + 0.010762382298707962, + 0.011984624899923801, + 0.015216648578643799, + 0.02251250669360161, + 0.04939379170536995, + 0.06634433567523956, + 0.01901494711637497, + 0.003934632521122694, + 0.0076883165165781975, + 0.005354671739041805, + 0.00839480571448803, + 0.005889970809221268, + 0.008262330666184425, + 0.007414902560412884, + 0.00541393319144845, + 0.018071651458740234, + 0.015989115461707115, + 0.01670440472662449, + 0.01962011307477951, + 0.015568208880722523, + 0.02041972056031227, + 0.02778904139995575, + 0.0223972387611866, + 0.009162874892354012, + 0.007487090770155191, + 0.007605139631778002, + 0.006574884057044983, + 0.008809753693640232, + 0.018846338614821434, + 0.01859487220644951, + 0.011319722980260849, + 0.0019204403506591916, + 0.0011529321782290936, + 0.0017752149142324924, + 0.0014447526773437858, + 0.0020153175573796034, + 0.0022151784505695105, + 0.0014697728911414742, + 0.0022003320045769215, + 0.0032713969703763723, + 0.0019480054033920169, + 0.0033162611071020365, + 0.0026897527277469635, + 0.00238120392896235, + 0.004595367703586817, + 0.0030922829173505306, + 0.003857910167425871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.0789763331413269 + } + ], + "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/0000128/group_5/index_tensor([5113, 5113])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_5/index_tensor([5113, 5113])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0ab3223bc401861e607b16dda3a5b85852c5bae2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_5/index_tensor([5113, 5113])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, slightly upward, medium close-up focuses on a middle-aged general with a stern, weathered face, ensconced within a dimly lit military tent. He wears an ancient suit of armor forged from a mixture of bronze and wrought iron. The shoulder plates are engraved with roaring animal faces, and the breastplate, inlaid with fine scales, reflects a cold, mottled luster in the flickering light of a single oil lamp. The edges of the armor are covered with battle scratches and slight rust. His graying hair is held in a high bun by an antique bronze hairpin, and his deep eyes exude a pensive and majestic air. He sits upright at a rough wooden low table, his hands steadying a simple bronze wine jug. Though the body is covered in green patina, the spout is polished to a shine. He carefully tilted the pot, and a stream of clear amber liquor flowed from the spout like a golden thread, precisely and steadily pouring into a stoneware bowl below. The glaze of the bowl shimmered as the liquor poured in. His gaze was completely focused on the flowing liquor, his expression focused and solemn. The warm, dim light carved a sharp contrast of light and dark on his deep features. The entire scene was dominated by the dark green of bronze, the iron gray of armor, and the dark canvas background. The amber liquor was the only bright spot, permeating the atmosphere of tranquility, solemnity, and contemplation of a post-war moment.", + "seed": 18995, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_5/index_5113_seed_18995_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": -0.02402571551501751, + "vlm_logprob": -0.02402571551501751, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02402571551501751, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055459555238485336 + ], + "opposite_logprob_tokens": [ + -3.180459499359131 + ], + "logprob_reward": -0.055459555238485336, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.055459555238485336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09150143712759018, + 0.06280918419361115, + 0.0967692881822586, + 0.07082637399435043, + 0.13567066192626953, + 0.07840824127197266, + 0.2123965620994568, + 0.08479826897382736, + 0.073649562895298, + 0.09317038208246231 + ], + "credit_map": [ + 0.014949191361665726, + 0.008990699425339699, + 0.008518079295754433, + 0.011738180182874203, + 0.011293867602944374, + 0.01392100378870964, + 0.011025885120034218, + 0.011064538732171059, + 0.005982572678476572, + 0.004293191712349653, + 0.0052616288885474205, + 0.00799532700330019, + 0.005672231316566467, + 0.010001244023442268, + 0.008822064846754074, + 0.01478093396872282, + 0.008445716463029385, + 0.009913683868944645, + 0.009105606935918331, + 0.013038378208875656, + 0.02111852914094925, + 0.013317194767296314, + 0.01100783422589302, + 0.010822348296642303, + 0.006532973609864712, + 0.005872931331396103, + 0.006831859704107046, + 0.006936452817171812, + 0.013574443757534027, + 0.009746001102030277, + 0.008817700669169426, + 0.012514014728367329, + 0.013622277416288853, + 0.009351249784231186, + 0.01301553100347519, + 0.013432797975838184, + 0.022185619920492172, + 0.01684005931019783, + 0.020024949684739113, + 0.0271981842815876, + 0.012737596407532692, + 0.006708318833261728, + 0.007141432724893093, + 0.007351344916969538, + 0.010891922749578953, + 0.008749576285481453, + 0.012169124558568, + 0.01265893504023552, + 0.06131554767489433, + 0.01085070613771677, + 0.011398490518331528, + 0.014458592049777508, + 0.015651943162083626, + 0.06448112428188324, + 0.0170898474752903, + 0.01715032011270523, + 0.014468829147517681, + 0.00900362990796566, + 0.006071764975786209, + 0.013678045943379402, + 0.0113400649279356, + 0.01192599069327116, + 0.009267286397516727, + 0.009042659774422646, + 0.013205405324697495, + 0.007212190423160791, + 0.006717746611684561, + 0.007269435562193394, + 0.00910120364278555, + 0.009340295568108559, + 0.011177877895534039, + 0.00962541252374649, + 0.014401672407984734, + 0.008137118071317673, + 0.009602604433894157, + 0.010672547854483128, + 0.016433952376246452, + 0.011647229082882404, + 0.013021519407629967, + 0.00925374310463667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.055459555238485336 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033714696764945984 + ], + "opposite_logprob_tokens": [ + -3.408714771270752 + ], + "logprob_reward": -0.033714696764945984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033714696764945984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12847407162189484, + 0.09641669690608978, + 0.15457402169704437, + 0.04491563141345978, + 0.060431621968746185, + 0.03684587404131889, + 0.12290634959936142, + 0.13246862590312958, + 0.08146391808986664, + 0.1415031999349594 + ], + "credit_map": [ + 0.019319504499435425, + 0.010718347504734993, + 0.009002541191875935, + 0.013621260412037373, + 0.017293987795710564, + 0.02046356350183487, + 0.016559220850467682, + 0.021495655179023743, + 0.011094437912106514, + 0.007877357304096222, + 0.007909066043794155, + 0.011616754345595837, + 0.01170907448977232, + 0.015206714160740376, + 0.01325572282075882, + 0.01774757355451584, + 0.01290967408567667, + 0.011907133273780346, + 0.014738786965608597, + 0.0227807629853487, + 0.019177205860614777, + 0.02382068522274494, + 0.01742991991341114, + 0.03180984780192375, + 0.004147741477936506, + 0.0035226852633059025, + 0.004644873086363077, + 0.0058486503548920155, + 0.007952176034450531, + 0.005946675315499306, + 0.007173077668994665, + 0.005679752677679062, + 0.006638784892857075, + 0.003759989980608225, + 0.006727525033056736, + 0.006588345859199762, + 0.009375989437103271, + 0.007103307172656059, + 0.009513959288597107, + 0.01072372030466795, + 0.0063591948710381985, + 0.002314171753823757, + 0.00224836403504014, + 0.004467461723834276, + 0.005082378629595041, + 0.004454712849110365, + 0.005082327406853437, + 0.0068372637033462524, + 0.03383345529437065, + 0.007326678838580847, + 0.00698969466611743, + 0.008466197177767754, + 0.007777678780257702, + 0.03504754975438118, + 0.008857335895299911, + 0.014607754535973072, + 0.029526393860578537, + 0.012498445808887482, + 0.008635174483060837, + 0.018770353868603706, + 0.012928755022585392, + 0.022113846614956856, + 0.013179946690797806, + 0.014815716072916985, + 0.01783290132880211, + 0.008840287104249, + 0.007121505681425333, + 0.00605488196015358, + 0.010225970298051834, + 0.01139998622238636, + 0.010524661280214787, + 0.009463733062148094, + 0.02712445706129074, + 0.012814992107450962, + 0.01144096627831459, + 0.012823566794395447, + 0.014659573324024677, + 0.023496538400650024, + 0.02435116097331047, + 0.014791948720812798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.033714696764945984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748029962182045 + ], + "opposite_logprob_tokens": [ + -5.006748199462891 + ], + "logprob_reward": -0.006748029962182045, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748029962182045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12871064245700836, + 0.06819949299097061, + 0.14559492468833923, + 0.07637819647789001, + 0.12342958152294159, + 0.05040958151221275, + 0.10373014211654663, + 0.1460423320531845, + 0.07404173165559769, + 0.0834634006023407 + ], + "credit_map": [ + 0.019897794350981712, + 0.01204911433160305, + 0.009401671588420868, + 0.018163857981562614, + 0.016367482021450996, + 0.017983241006731987, + 0.01659102737903595, + 0.018256457522511482, + 0.005224325694143772, + 0.008696089498698711, + 0.00642387056723237, + 0.009823830798268318, + 0.008045395836234093, + 0.0095455851405859, + 0.008050169795751572, + 0.012390225194394588, + 0.01246716920286417, + 0.014464805833995342, + 0.016808003187179565, + 0.021816866472363472, + 0.022319916635751724, + 0.021602656692266464, + 0.014624234288930893, + 0.021491259336471558, + 0.007086404133588076, + 0.006559328641742468, + 0.005346687510609627, + 0.009758307598531246, + 0.013246996328234673, + 0.009698777459561825, + 0.010810032486915588, + 0.013871664181351662, + 0.015087625943124294, + 0.010203049518167973, + 0.012305956333875656, + 0.013779270462691784, + 0.021590104326605797, + 0.01576060988008976, + 0.019118724390864372, + 0.015584239736199379, + 0.008473562076687813, + 0.0035487241111695766, + 0.004474898334592581, + 0.00632390147075057, + 0.008137659169733524, + 0.006677965633571148, + 0.005566624458879232, + 0.007206244394183159, + 0.018326422199606895, + 0.0091929342597723, + 0.010622340254485607, + 0.012899246998131275, + 0.012141586281359196, + 0.015018983744084835, + 0.014643318951129913, + 0.010885313153266907, + 0.023364029824733734, + 0.01422442402690649, + 0.01058335229754448, + 0.019246401265263557, + 0.02235102653503418, + 0.021390726789832115, + 0.01873416267335415, + 0.016148215159773827, + 0.008599203079938889, + 0.006258566398173571, + 0.005880595184862614, + 0.009867646731436253, + 0.01503677573055029, + 0.009397851303219795, + 0.011411551386117935, + 0.007589541841298342, + 0.01176149956882, + 0.007748214993625879, + 0.008885861374437809, + 0.010381168685853481, + 0.015772942453622818, + 0.008923976682126522, + 0.011560742743313313, + 0.00842899177223444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.006748029962182045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058674581348896 + ], + "opposite_logprob_tokens": [ + -3.8955867290496826 + ], + "logprob_reward": -0.02058674581348896, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058674581348896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08631175756454468, + 0.09233459830284119, + 0.10184382647275925, + 0.13273167610168457, + 0.04720322787761688, + 0.057329922914505005, + 0.13170626759529114, + 0.21162953972816467, + 0.07649888843297958, + 0.062410254031419754 + ], + "credit_map": [ + 0.010392038151621819, + 0.008020900189876556, + 0.007594664115458727, + 0.0128102358430624, + 0.012062424793839455, + 0.01241752877831459, + 0.010018506087362766, + 0.012995473109185696, + 0.007550922688096762, + 0.00676034577190876, + 0.007225369568914175, + 0.010683548636734486, + 0.016035793349146843, + 0.014570998959243298, + 0.015199721790850163, + 0.014307908713817596, + 0.009924523532390594, + 0.008935521356761456, + 0.012403692118823528, + 0.01301526091992855, + 0.01825181394815445, + 0.015275965444743633, + 0.012482143938541412, + 0.011554920114576817, + 0.012982350774109364, + 0.012199895456433296, + 0.009347568266093731, + 0.016398292034864426, + 0.029614314436912537, + 0.018745673820376396, + 0.019826924428343773, + 0.013616673648357391, + 0.006064501591026783, + 0.0038107673171907663, + 0.0036928036715835333, + 0.004936589859426022, + 0.008837606757879257, + 0.005431762896478176, + 0.006829721853137016, + 0.0075994799844920635, + 0.00821008626371622, + 0.005449688993394375, + 0.005691869650036097, + 0.005161578767001629, + 0.010925566777586937, + 0.005366305820643902, + 0.0085054412484169, + 0.008019394241273403, + 0.020757196471095085, + 0.015747148543596268, + 0.01400147657841444, + 0.012749412097036839, + 0.01676003262400627, + 0.01985965482890606, + 0.016582975164055824, + 0.015248379670083523, + 0.041560158133506775, + 0.020133377984166145, + 0.011857060715556145, + 0.030322307720780373, + 0.028679074719548225, + 0.0409209281206131, + 0.019533950835466385, + 0.018622703850269318, + 0.0149535546079278, + 0.00890117883682251, + 0.006330325733870268, + 0.007338776718825102, + 0.008370655588805676, + 0.011217457242310047, + 0.011482461355626583, + 0.007904483005404472, + 0.007958890870213509, + 0.006596500985324383, + 0.0066226208582520485, + 0.010004411451518536, + 0.008177421987056732, + 0.006667230278253555, + 0.010474110022187233, + 0.005909077823162079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.02058674581348896 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619549795985222 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619549795985222, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619549795985222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0851348266005516, + 0.08109278231859207, + 0.07272939383983612, + 0.039392679929733276, + 0.06887631118297577, + 0.12107227742671967, + 0.08740626275539398, + 0.07973110675811768, + 0.2640899419784546, + 0.10047442466020584 + ], + "credit_map": [ + 0.010756519623100758, + 0.01131407730281353, + 0.008518693968653679, + 0.010457869619131088, + 0.013121661730110645, + 0.011355490423738956, + 0.009085321798920631, + 0.010525190271437168, + 0.007864491082727909, + 0.005978161934763193, + 0.007301083765923977, + 0.009218797087669373, + 0.008749560452997684, + 0.012205235660076141, + 0.010566437616944313, + 0.019209010526537895, + 0.013572121039032936, + 0.004410173278301954, + 0.005733570083975792, + 0.009546404704451561, + 0.00557288620620966, + 0.011972016654908657, + 0.006054166238754988, + 0.015868056565523148, + 0.005110682919621468, + 0.004328821320086718, + 0.0035942825488746166, + 0.004055211320519447, + 0.0065942746587097645, + 0.005512769799679518, + 0.004587820265442133, + 0.005608818959444761, + 0.0071344212628901005, + 0.0046536740846931934, + 0.006409714464098215, + 0.006882566958665848, + 0.00980278942734003, + 0.008498292416334152, + 0.01336562167853117, + 0.012129231356084347, + 0.02307739667594433, + 0.008678453974425793, + 0.013667810708284378, + 0.011149688623845577, + 0.01694834791123867, + 0.01621861942112446, + 0.0158817321062088, + 0.015450223349034786, + 0.01093878224492073, + 0.007374056614935398, + 0.00919931847602129, + 0.009845222346484661, + 0.017680911347270012, + 0.011253178119659424, + 0.010645515285432339, + 0.010469276458024979, + 0.013803215697407722, + 0.008774819783866405, + 0.005105677526444197, + 0.009962004609405994, + 0.008689641021192074, + 0.0117481155321002, + 0.008056586608290672, + 0.01359104085713625, + 0.0572940818965435, + 0.02137940376996994, + 0.022152129560709, + 0.017519483342766762, + 0.035622384399175644, + 0.03779810667037964, + 0.03540715202689171, + 0.036917198449373245, + 0.014872841536998749, + 0.008824306540191174, + 0.011816899292171001, + 0.013986846432089806, + 0.014924075454473495, + 0.010808843187987804, + 0.01566052809357643, + 0.009580082260072231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.003619549795985222 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, slightly upward, medium close-up focuses on a middle-aged general with a stern, weathered face, ensconced within a dimly lit military tent. He wears an ancient suit of armor forged from a mixture of bronze and wrought iron. The shoulder plates are engraved with roaring animal faces, and the breastplate, inlaid with fine scales, reflects a cold, mottled luster in the flickering light of a single oil lamp. The edges of the armor are covered with battle scratches and slight rust. His graying hair is held in a high bun by an antique bronze hairpin, and his deep eyes exude a pensive and majestic air. He sits upright at a rough wooden low table, his hands steadying a simple bronze wine jug. Though the body is covered in green patina, the spout is polished to a shine. He carefully tilted the pot, and a stream of clear amber liquor flowed from the spout like a golden thread, precisely and steadily pouring into a stoneware bowl below. The glaze of the bowl shimmered as the liquor poured in. His gaze was completely focused on the flowing liquor, his expression focused and solemn. The warm, dim light carved a sharp contrast of light and dark on his deep features. The entire scene was dominated by the dark green of bronze, the iron gray of armor, and the dark canvas background. The amber liquor was the only bright spot, permeating the atmosphere of tranquility, solemnity, and contemplation of a post-war moment.", + "seed": 18996, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_5/index_5113_seed_18996_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": -0.024242278840392827, + "vlm_logprob": -0.024242278840392827, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.024242278840392827, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06267247349023819 + ], + "opposite_logprob_tokens": [ + -3.0626723766326904 + ], + "logprob_reward": -0.06267247349023819, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06267247349023819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09034623205661774, + 0.07065863907337189, + 0.08084233850240707, + 0.045272428542375565, + 0.11822369694709778, + 0.10117638111114502, + 0.15083880722522736, + 0.03675668686628342, + 0.1390867978334427, + 0.16679798066616058 + ], + "credit_map": [ + 0.009597362950444221, + 0.012881845235824585, + 0.014016778208315372, + 0.007578246295452118, + 0.012919538654386997, + 0.010792120359838009, + 0.012410474941134453, + 0.010149864479899406, + 0.007859934121370316, + 0.00941852480173111, + 0.008539310656487942, + 0.007588948123157024, + 0.007196654565632343, + 0.009642488323152065, + 0.012588118202984333, + 0.00782465934753418, + 0.008923635818064213, + 0.011013106442987919, + 0.005412485916167498, + 0.007488641422241926, + 0.014901372604072094, + 0.010250572115182877, + 0.013805169612169266, + 0.009047359228134155, + 0.005485339555889368, + 0.005892663262784481, + 0.0029069571755826473, + 0.0035920203663408756, + 0.007346763741225004, + 0.00652972562238574, + 0.007787374779582024, + 0.005731587763875723, + 0.018077701330184937, + 0.011124366894364357, + 0.007163657806813717, + 0.008951269090175629, + 0.016472503542900085, + 0.018465610221028328, + 0.01347105111926794, + 0.024497533217072487, + 0.01095170434564352, + 0.010225440375506878, + 0.009952322579920292, + 0.009897898882627487, + 0.017184708267450333, + 0.013143494725227356, + 0.01517315674573183, + 0.014647658914327621, + 0.03754511475563049, + 0.026935523375868797, + 0.0077834478579461575, + 0.007685683201998472, + 0.016139637678861618, + 0.022606568410992622, + 0.017258314415812492, + 0.01488452311605215, + 0.004887375049293041, + 0.004488423466682434, + 0.003812558948993683, + 0.004646473564207554, + 0.003313619876280427, + 0.004954247735440731, + 0.003878437913954258, + 0.006775554269552231, + 0.017000019550323486, + 0.01477478165179491, + 0.01249265018850565, + 0.01769230142235756, + 0.014896392822265625, + 0.02031788043677807, + 0.01818685233592987, + 0.023725928738713264, + 0.014691730961203575, + 0.012981317937374115, + 0.009723401628434658, + 0.036264512687921524, + 0.016846193000674248, + 0.024177186191082, + 0.018026893958449364, + 0.03408673405647278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.06267247349023819 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550539337098598 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.012550539337098598, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550539337098598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21598142385482788, + 0.11248359829187393, + 0.0713617131114006, + 0.03137921541929245, + 0.08696217089891434, + 0.08135268837213516, + 0.1075635477900505, + 0.09624266624450684, + 0.08584828674793243, + 0.11082468181848526 + ], + "credit_map": [ + 0.02251332812011242, + 0.031761329621076584, + 0.031004710122942924, + 0.03319660574197769, + 0.022537991404533386, + 0.022680507972836494, + 0.029787609353661537, + 0.022499343380331993, + 0.009917010553181171, + 0.014501726254820824, + 0.011849354021251202, + 0.009030753746628761, + 0.01878761313855648, + 0.014583131298422813, + 0.022342929616570473, + 0.01147108431905508, + 0.007515407167375088, + 0.013235382735729218, + 0.0059514837339520454, + 0.00738784484565258, + 0.0108925960958004, + 0.010207189247012138, + 0.010757573880255222, + 0.005414233542978764, + 0.0032433844171464443, + 0.0035556545481085777, + 0.00304663204587996, + 0.004279676359146833, + 0.0045151072554290295, + 0.004231313243508339, + 0.004035153891891241, + 0.004472296219319105, + 0.01756628043949604, + 0.008348798379302025, + 0.00563290249556303, + 0.007345619145780802, + 0.005948763806372881, + 0.014350776560604572, + 0.007047839462757111, + 0.0207211896777153, + 0.009958960115909576, + 0.007534832693636417, + 0.00665441295132041, + 0.0054825046099722385, + 0.011457584798336029, + 0.010079787112772465, + 0.014605523087084293, + 0.015579085797071457, + 0.012591412290930748, + 0.011310386471450329, + 0.00971140619367361, + 0.01292591542005539, + 0.015015925280749798, + 0.013674029149115086, + 0.01628774404525757, + 0.01604672148823738, + 0.012492110021412373, + 0.01260156836360693, + 0.007991321384906769, + 0.011454903520643711, + 0.015110314823687077, + 0.013882542960345745, + 0.01014566607773304, + 0.012564246542751789, + 0.013099397532641888, + 0.012531922198832035, + 0.007198021747171879, + 0.009250459261238575, + 0.01087847352027893, + 0.012299055233597755, + 0.009743914008140564, + 0.010847039520740509, + 0.013024000450968742, + 0.01241893321275711, + 0.010641041211783886, + 0.016304627060890198, + 0.013244797475636005, + 0.014905956573784351, + 0.01209922693669796, + 0.01818610541522503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012550539337098598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790142998099327 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.009790142998099327, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790142998099327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1760399490594864, + 0.128215029835701, + 0.05729454383254051, + 0.08677898347377777, + 0.07714282721281052, + 0.07779080420732498, + 0.1634342074394226, + 0.07909414172172546, + 0.050557851791381836, + 0.10365168750286102 + ], + "credit_map": [ + 0.023456009104847908, + 0.02606366015970707, + 0.017723621800541878, + 0.022363558411598206, + 0.017269643023610115, + 0.021458590403199196, + 0.02632104977965355, + 0.021383799612522125, + 0.013675467111170292, + 0.015399434603750706, + 0.014740144833922386, + 0.014211977832019329, + 0.0179483313113451, + 0.018142998218536377, + 0.020167754963040352, + 0.013928915373980999, + 0.005587227642536163, + 0.007467041723430157, + 0.006948335561901331, + 0.006531249266117811, + 0.007945668883621693, + 0.007781732361763716, + 0.00875889789313078, + 0.0062743863090872765, + 0.008375370875000954, + 0.011257406324148178, + 0.007450972218066454, + 0.008482072502374649, + 0.011404707096517086, + 0.012352986261248589, + 0.014898587018251419, + 0.012556871399283409, + 0.013918767683207989, + 0.0086587555706501, + 0.005154977086931467, + 0.007809541653841734, + 0.009267883375287056, + 0.01009207684546709, + 0.009387065656483173, + 0.012853756546974182, + 0.009260648861527443, + 0.00716741057112813, + 0.009122400544583797, + 0.0065210130997002125, + 0.01353316567838192, + 0.00866510346531868, + 0.012196776457130909, + 0.011324276216328144, + 0.04318264499306679, + 0.03221886232495308, + 0.009174566715955734, + 0.009887606836855412, + 0.014093889854848385, + 0.02599470503628254, + 0.013006630353629589, + 0.015875281766057014, + 0.01210677158087492, + 0.008963377214968204, + 0.005270964000374079, + 0.011991335079073906, + 0.00817170925438404, + 0.00989849865436554, + 0.009802880696952343, + 0.012888602912425995, + 0.0054399799555540085, + 0.006542910356074572, + 0.005102343857288361, + 0.004921865649521351, + 0.007528853137046099, + 0.007486197631806135, + 0.007589112035930157, + 0.005946584511548281, + 0.0099828215315938, + 0.011344287544488907, + 0.008714234456419945, + 0.01561945304274559, + 0.01535892579704523, + 0.012981690466403961, + 0.014351628720760345, + 0.015298636630177498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.009790142998099327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337059386074543 + ], + "opposite_logprob_tokens": [ + -3.408705949783325 + ], + "logprob_reward": -0.0337059386074543, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337059386074543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11073224991559982, + 0.17072537541389465, + 0.02967459335923195, + 0.05531163886189461, + 0.07519374787807465, + 0.06872353702783585, + 0.04047195613384247, + 0.052595216780900955, + 0.09281516820192337, + 0.3037565350532532 + ], + "credit_map": [ + 0.012395194731652737, + 0.01655665971338749, + 0.010012448765337467, + 0.012142375111579895, + 0.012964857742190361, + 0.014777404256165028, + 0.018757369369268417, + 0.01312593836337328, + 0.015552729368209839, + 0.01842227205634117, + 0.020340126007795334, + 0.024776805192232132, + 0.022898515686392784, + 0.024182401597499847, + 0.027204571291804314, + 0.01734795980155468, + 0.0025175309274345636, + 0.0050141639076173306, + 0.0022456995211541653, + 0.002885442925617099, + 0.004071454983204603, + 0.004457186907529831, + 0.0053732553496956825, + 0.0031098583713173866, + 0.006123803555965424, + 0.007441333495080471, + 0.006713024340569973, + 0.0068726311437785625, + 0.007173039019107819, + 0.006918956059962511, + 0.008279431611299515, + 0.005789417773485184, + 0.011539081111550331, + 0.007795576937496662, + 0.005373186897486448, + 0.006599735002964735, + 0.00834911223500967, + 0.010392853990197182, + 0.009675189852714539, + 0.015469015575945377, + 0.009622886776924133, + 0.006427137181162834, + 0.0097827622666955, + 0.008253112435340881, + 0.007957155816257, + 0.00883412640541792, + 0.009380803443491459, + 0.008465549908578396, + 0.005805402994155884, + 0.004941317718476057, + 0.0032594630029052496, + 0.004145471379160881, + 0.0050794100388884544, + 0.005348086357116699, + 0.0055463495664298534, + 0.006346455775201321, + 0.004776780493557453, + 0.005331970285624266, + 0.0032020965591073036, + 0.005106536205857992, + 0.011377514339983463, + 0.00688083516433835, + 0.00790813285857439, + 0.0080113485455513, + 0.01007310301065445, + 0.01105143316090107, + 0.007473784498870373, + 0.007631157990545034, + 0.015263285487890244, + 0.01207469217479229, + 0.01666494645178318, + 0.012582764029502869, + 0.03342321142554283, + 0.023744123056530952, + 0.01888294704258442, + 0.06761046499013901, + 0.025760335847735405, + 0.046656616032123566, + 0.024060240015387535, + 0.06361860036849976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0337059386074543 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024922997690737247 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024922997690737247, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024922997690737247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10565219819545746, + 0.054528385400772095, + 0.038151711225509644, + 0.04419882223010063, + 0.06255726516246796, + 0.05719104781746864, + 0.12504060566425323, + 0.14567671716213226, + 0.25337114930152893, + 0.11363207548856735 + ], + "credit_map": [ + 0.009564393199980259, + 0.017910202965140343, + 0.010978437960147858, + 0.011740094982087612, + 0.014168144203722477, + 0.01261229533702135, + 0.013131571933627129, + 0.01554705947637558, + 0.004882441833615303, + 0.008302431553602219, + 0.004477128852158785, + 0.006466636899858713, + 0.0063772182911634445, + 0.009538333863019943, + 0.0078009190037846565, + 0.006683278363198042, + 0.003398916916921735, + 0.005235088057816029, + 0.003302830271422863, + 0.0038138735108077526, + 0.0057898638769984245, + 0.005288716871291399, + 0.007670266553759575, + 0.003652155166491866, + 0.0035909991711378098, + 0.005058160051703453, + 0.0019780576694756746, + 0.003998730331659317, + 0.006610595155507326, + 0.006801936775445938, + 0.010484615340828896, + 0.005675726104527712, + 0.009714631363749504, + 0.007868866436183453, + 0.0053724381141364574, + 0.007960501126945019, + 0.006606911309063435, + 0.008572656661272049, + 0.005878943484276533, + 0.01058232318609953, + 0.010949662886559963, + 0.005838876124471426, + 0.006697706412523985, + 0.004328764043748379, + 0.0063933213241398335, + 0.007652852684259415, + 0.007690731901675463, + 0.0076391310431063175, + 0.030986392870545387, + 0.024498924612998962, + 0.007737494073808193, + 0.007363774348050356, + 0.011879242956638336, + 0.020140109583735466, + 0.0111854812130332, + 0.011249188333749771, + 0.019048573449254036, + 0.015935318544507027, + 0.01437532901763916, + 0.022161679342389107, + 0.02086838334798813, + 0.018965480849146843, + 0.014507465995848179, + 0.019814496859908104, + 0.029102832078933716, + 0.025279242545366287, + 0.021258831024169922, + 0.03380883112549782, + 0.02481582574546337, + 0.03881382569670677, + 0.030941836535930634, + 0.04934992641210556, + 0.012712052091956139, + 0.011986230500042439, + 0.006824231706559658, + 0.018582705408334732, + 0.01262847799807787, + 0.017356855794787407, + 0.012549077160656452, + 0.020992446690797806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.0024922997690737247 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, slightly upward, medium close-up focuses on a middle-aged general with a stern, weathered face, ensconced within a dimly lit military tent. He wears an ancient suit of armor forged from a mixture of bronze and wrought iron. The shoulder plates are engraved with roaring animal faces, and the breastplate, inlaid with fine scales, reflects a cold, mottled luster in the flickering light of a single oil lamp. The edges of the armor are covered with battle scratches and slight rust. His graying hair is held in a high bun by an antique bronze hairpin, and his deep eyes exude a pensive and majestic air. He sits upright at a rough wooden low table, his hands steadying a simple bronze wine jug. Though the body is covered in green patina, the spout is polished to a shine. He carefully tilted the pot, and a stream of clear amber liquor flowed from the spout like a golden thread, precisely and steadily pouring into a stoneware bowl below. The glaze of the bowl shimmered as the liquor poured in. His gaze was completely focused on the flowing liquor, his expression focused and solemn. The warm, dim light carved a sharp contrast of light and dark on his deep features. The entire scene was dominated by the dark green of bronze, the iron gray of armor, and the dark canvas background. The amber liquor was the only bright spot, permeating the atmosphere of tranquility, solemnity, and contemplation of a post-war moment.", + "seed": 18997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_5/index_5113_seed_18997_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": -0.024530993402004243, + "vlm_logprob": -0.024530993402004243, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.024530993402004243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045648396015167236 + ], + "opposite_logprob_tokens": [ + -3.2956483364105225 + ], + "logprob_reward": -0.045648396015167236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.045648396015167236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2543976604938507, + 0.03676464781165123, + 0.08774741739034653, + 0.06466932594776154, + 0.10741414874792099, + 0.06609150767326355, + 0.11199528723955154, + 0.11663071811199188, + 0.09229441732168198, + 0.061994876712560654 + ], + "credit_map": [ + 0.025070305913686752, + 0.02590673416852951, + 0.040091950446367264, + 0.05033828690648079, + 0.025494296103715897, + 0.024249467998743057, + 0.024488454684615135, + 0.03875817731022835, + 0.00488158967345953, + 0.003628081874921918, + 0.0042142062447965145, + 0.004092435352504253, + 0.006052167620509863, + 0.002713751047849655, + 0.006743053440004587, + 0.00443936325609684, + 0.012013035826385021, + 0.006016280967742205, + 0.014984799548983574, + 0.005641760304570198, + 0.009275737218558788, + 0.006951221264898777, + 0.008894375525414944, + 0.023970207199454308, + 0.010689917951822281, + 0.00407798308879137, + 0.0069273691624403, + 0.005055537912994623, + 0.005777423270046711, + 0.006760070566087961, + 0.006541105452924967, + 0.018839925527572632, + 0.019038403406739235, + 0.00955894123762846, + 0.011707628145813942, + 0.009781423024833202, + 0.015088182874023914, + 0.019153879955410957, + 0.010701566003262997, + 0.012384118512272835, + 0.008013410493731499, + 0.008677693083882332, + 0.007897138595581055, + 0.0070171114057302475, + 0.012165405787527561, + 0.00847514346241951, + 0.007293862756341696, + 0.00655174208804965, + 0.011741024442017078, + 0.019435113295912743, + 0.010959134437143803, + 0.009710791520774364, + 0.01953546702861786, + 0.015851350501179695, + 0.015301086939871311, + 0.009461320005357265, + 0.026121074333786964, + 0.027648214250802994, + 0.008223768323659897, + 0.007435952313244343, + 0.009043495170772076, + 0.02036128379404545, + 0.010318503715097904, + 0.007478424347937107, + 0.012117463164031506, + 0.012705539353191853, + 0.01168327871710062, + 0.01252527255564928, + 0.011122584342956543, + 0.011632760986685753, + 0.011568134650588036, + 0.008939383551478386, + 0.013451666571199894, + 0.012253264896571636, + 0.004902983084321022, + 0.005872500594705343, + 0.005368581507354975, + 0.00908179022371769, + 0.005854860879480839, + 0.005209226161241531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.045648396015167236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192103132605553 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192103132605553, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192103132605553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13479463756084442, + 0.1486244797706604, + 0.10335920751094818, + 0.11903057992458344, + 0.07740478217601776, + 0.07976945489645004, + 0.045802224427461624, + 0.06092988699674606, + 0.04087239131331444, + 0.1894124299287796 + ], + "credit_map": [ + 0.014569857157766819, + 0.012303224764764309, + 0.015234439633786678, + 0.018025850877165794, + 0.019152073189616203, + 0.01564122550189495, + 0.012342114932835102, + 0.027525832876563072, + 0.020400794222950935, + 0.009048744104802608, + 0.020840875804424286, + 0.010714310221374035, + 0.015059592202305794, + 0.015699341893196106, + 0.0384073443710804, + 0.01845346949994564, + 0.01869260147213936, + 0.006906693335622549, + 0.019763624295592308, + 0.009419961832463741, + 0.010445017367601395, + 0.007150439545512199, + 0.015168692916631699, + 0.01581217162311077, + 0.020211176946759224, + 0.010458482429385185, + 0.013089407235383987, + 0.015955621376633644, + 0.012692169286310673, + 0.016999604180455208, + 0.014411158859729767, + 0.015212946571409702, + 0.011857918463647366, + 0.008452135138213634, + 0.008229092694818974, + 0.009707563556730747, + 0.012931807897984982, + 0.011299850419163704, + 0.006316658109426498, + 0.008609750308096409, + 0.009367788210511208, + 0.01117205061018467, + 0.011334409937262535, + 0.009769811294972897, + 0.011533048003911972, + 0.009586276486515999, + 0.008154500275850296, + 0.008851563557982445, + 0.006211787462234497, + 0.009847518987953663, + 0.005767008755356073, + 0.003562481142580509, + 0.004951884504407644, + 0.007061201147735119, + 0.00518152117729187, + 0.003218820784240961, + 0.015574312768876553, + 0.011189800687134266, + 0.00452991621568799, + 0.004357682075351477, + 0.0061765615828335285, + 0.010496759787201881, + 0.0053108069114387035, + 0.0032940427772700787, + 0.006476882379502058, + 0.0052292002364993095, + 0.005565799307078123, + 0.004949967842549086, + 0.00438526039943099, + 0.00566292367875576, + 0.004451930522918701, + 0.004150424152612686, + 0.03944271430373192, + 0.036798249930143356, + 0.018505332991480827, + 0.015867646783590317, + 0.015010054223239422, + 0.02826358750462532, + 0.020021218806505203, + 0.015503622591495514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.018192103132605553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607118360698223 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.01607118360698223, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607118360698223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23674419522285461, + 0.05507471784949303, + 0.14725197851657867, + 0.15290802717208862, + 0.07494225353002548, + 0.11732316017150879, + 0.037152502685785294, + 0.05750637874007225, + 0.0325193889439106, + 0.08857738226652145 + ], + "credit_map": [ + 0.02209094539284706, + 0.02561642788350582, + 0.028916962444782257, + 0.048444584012031555, + 0.022589590400457382, + 0.023976201191544533, + 0.026488356292247772, + 0.038621120154857635, + 0.008088315837085247, + 0.004295137245208025, + 0.0076657626777887344, + 0.005445951130241156, + 0.0058454000391066074, + 0.004011065699160099, + 0.011467045173048973, + 0.008256037719547749, + 0.018117111176252365, + 0.010761896148324013, + 0.03448571264743805, + 0.010811839252710342, + 0.011290153488516808, + 0.010991731658577919, + 0.018500689417123795, + 0.03229285404086113, + 0.0254092738032341, + 0.010177033953368664, + 0.018062826246023178, + 0.008155319839715958, + 0.023106364533305168, + 0.013574853539466858, + 0.01525889802724123, + 0.03916345164179802, + 0.011762185022234917, + 0.006172263529151678, + 0.007299408782273531, + 0.007152428384870291, + 0.00908155832439661, + 0.015158537775278091, + 0.005772725213319063, + 0.01254314836114645, + 0.012514220550656319, + 0.014746662229299545, + 0.012782988138496876, + 0.009881684556603432, + 0.025289777666330338, + 0.015324694104492664, + 0.016216397285461426, + 0.010566738434135914, + 0.004838488530367613, + 0.005238757003098726, + 0.0038652706425637007, + 0.0023580060806125402, + 0.007455084938555956, + 0.005190243478864431, + 0.005960346665233374, + 0.002246306976303458, + 0.006241371389478445, + 0.009430800564587116, + 0.004463498480618, + 0.00534672848880291, + 0.011868059635162354, + 0.008498148061335087, + 0.00790033582597971, + 0.0037574362941086292, + 0.003817459335550666, + 0.004229913465678692, + 0.0032931233290582895, + 0.002513454994186759, + 0.006784006487578154, + 0.004941424820572138, + 0.005130236968398094, + 0.0018097717547789216, + 0.023099565878510475, + 0.019482634961605072, + 0.007517771329730749, + 0.0068958778865635395, + 0.005039687734097242, + 0.014561152085661888, + 0.00543451402336359, + 0.006546184420585632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.01607118360698223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038104183971881866 + ], + "opposite_logprob_tokens": [ + -3.288104295730591 + ], + "logprob_reward": -0.038104183971881866, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038104183971881866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05598394200205803, + 0.06559617817401886, + 0.10583069920539856, + 0.12789379060268402, + 0.15095962584018707, + 0.21243955194950104, + 0.03082086890935898, + 0.058019522577524185, + 0.027293849736452103, + 0.16516199707984924 + ], + "credit_map": [ + 0.0066925049759447575, + 0.003953489009290934, + 0.011404897086322308, + 0.004510177299380302, + 0.00710246292874217, + 0.005745638161897659, + 0.005321124102920294, + 0.011253646574914455, + 0.01020355150103569, + 0.0058395820669829845, + 0.007194636855274439, + 0.0077947210520505905, + 0.009910057298839092, + 0.005566355772316456, + 0.011361346580088139, + 0.007725928910076618, + 0.012676609680056572, + 0.008583455346524715, + 0.021236393600702286, + 0.010171938687562943, + 0.008678329177200794, + 0.007528502028435469, + 0.012718000449240208, + 0.024237461388111115, + 0.021608319133520126, + 0.011304352432489395, + 0.016038214787840843, + 0.016237160190939903, + 0.01233090739697218, + 0.019653892144560814, + 0.01616566814482212, + 0.014555273577570915, + 0.018739329650998116, + 0.014306662604212761, + 0.022533252835273743, + 0.02011941187083721, + 0.022191781550645828, + 0.021454494446516037, + 0.013834464363753796, + 0.01778022199869156, + 0.02338108792901039, + 0.02510048821568489, + 0.021944155916571617, + 0.016033465042710304, + 0.04909183830022812, + 0.027674218639731407, + 0.03446933999657631, + 0.014744958840310574, + 0.004096194636076689, + 0.0042310138233006, + 0.002812302438542247, + 0.002096688374876976, + 0.005535061005502939, + 0.004924715496599674, + 0.005101012997329235, + 0.0020238813012838364, + 0.015657205134630203, + 0.01025678962469101, + 0.005805125925689936, + 0.003332540625706315, + 0.004244901705533266, + 0.011254467070102692, + 0.00432015024125576, + 0.0031483485363423824, + 0.0030712990555912256, + 0.0031283998396247625, + 0.003118740161880851, + 0.0035479478538036346, + 0.004315936006605625, + 0.0041021923534572124, + 0.003429792821407318, + 0.002579543273895979, + 0.042695146054029465, + 0.03748650848865509, + 0.013196361251175404, + 0.013665663078427315, + 0.009472423233091831, + 0.02689213678240776, + 0.009971950203180313, + 0.011781800538301468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.038104183971881866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639100283384323 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639100283384323, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639100283384323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04225945472717285, + 0.024625379592180252, + 0.034372761845588684, + 0.023769620805978775, + 0.035041164606809616, + 0.03045157715678215, + 0.030592646449804306, + 0.02796454355120659, + 0.02061644196510315, + 0.7303063869476318 + ], + "credit_map": [ + 0.005873365327715874, + 0.0032553423661738634, + 0.00836947001516819, + 0.0032885349355638027, + 0.0058888220228254795, + 0.004438912030309439, + 0.006150253117084503, + 0.004994760267436504, + 0.0037963639479130507, + 0.001619429443962872, + 0.0030975693371146917, + 0.002092110924422741, + 0.0035942746326327324, + 0.001709440373815596, + 0.005500183440744877, + 0.0032160079572349787, + 0.005949751939624548, + 0.002957561518996954, + 0.0043675899505615234, + 0.002722641918808222, + 0.002831029938533902, + 0.003038339549675584, + 0.0034130611456930637, + 0.009092788212001324, + 0.003628240432590246, + 0.0014918853994458914, + 0.0034901273902505636, + 0.003081222763285041, + 0.0028084497898817062, + 0.0018135206773877144, + 0.0025723278522491455, + 0.004883849062025547, + 0.006172743160277605, + 0.0031135117169469595, + 0.003254410345107317, + 0.003145396476611495, + 0.0047905053943395615, + 0.006159504409879446, + 0.00506553566083312, + 0.0033395634964108467, + 0.003889748128131032, + 0.004149063024669886, + 0.0031911658588796854, + 0.0022663897834718227, + 0.006411134265363216, + 0.0046066660434007645, + 0.003833833849057555, + 0.0021035801619291306, + 0.003236042335629463, + 0.0037469163071364164, + 0.0036179590970277786, + 0.0035551146138459444, + 0.004239901900291443, + 0.004172105807811022, + 0.005377737805247307, + 0.002646871143952012, + 0.006385316606611013, + 0.004469508305191994, + 0.002619450679048896, + 0.0027124248445034027, + 0.0026717307046055794, + 0.004883330315351486, + 0.002485918579623103, + 0.001736869104206562, + 0.0028356576804071665, + 0.004161662422120571, + 0.002143253805115819, + 0.0023227338679134846, + 0.0020401382353156805, + 0.0026526751462370157, + 0.0023833971936255693, + 0.0020769264083355665, + 0.18850870430469513, + 0.16218958795070648, + 0.059813618659973145, + 0.060808535665273666, + 0.04094360023736954, + 0.11759573221206665, + 0.0443614199757576, + 0.05608522146940231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.004639100283384323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, slightly upward, medium close-up focuses on a middle-aged general with a stern, weathered face, ensconced within a dimly lit military tent. He wears an ancient suit of armor forged from a mixture of bronze and wrought iron. The shoulder plates are engraved with roaring animal faces, and the breastplate, inlaid with fine scales, reflects a cold, mottled luster in the flickering light of a single oil lamp. The edges of the armor are covered with battle scratches and slight rust. His graying hair is held in a high bun by an antique bronze hairpin, and his deep eyes exude a pensive and majestic air. He sits upright at a rough wooden low table, his hands steadying a simple bronze wine jug. Though the body is covered in green patina, the spout is polished to a shine. He carefully tilted the pot, and a stream of clear amber liquor flowed from the spout like a golden thread, precisely and steadily pouring into a stoneware bowl below. The glaze of the bowl shimmered as the liquor poured in. His gaze was completely focused on the flowing liquor, his expression focused and solemn. The warm, dim light carved a sharp contrast of light and dark on his deep features. The entire scene was dominated by the dark green of bronze, the iron gray of armor, and the dark canvas background. The amber liquor was the only bright spot, permeating the atmosphere of tranquility, solemnity, and contemplation of a post-war moment.", + "seed": 18998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_5/index_5113_seed_18998_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": -0.028209206042811276, + "vlm_logprob": -0.028209206042811276, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.028209206042811276, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07849262654781342 + ], + "opposite_logprob_tokens": [ + -2.7034926414489746 + ], + "logprob_reward": -0.07849262654781342, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07849262654781342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23744294047355652, + 0.09911264479160309, + 0.06708236783742905, + 0.10318775475025177, + 0.14627107977867126, + 0.09302294254302979, + 0.045047976076602936, + 0.06920430064201355, + 0.04775978624820709, + 0.09186822175979614 + ], + "credit_map": [ + 0.025141611695289612, + 0.018270792439579964, + 0.046648625284433365, + 0.032084111124277115, + 0.033351361751556396, + 0.032164305448532104, + 0.018773149698972702, + 0.031008992344141006, + 0.011318324133753777, + 0.009900073520839214, + 0.015925390645861626, + 0.012505372054874897, + 0.017431657761335373, + 0.01717878133058548, + 0.006783589720726013, + 0.008069455623626709, + 0.00826172437518835, + 0.006106750573962927, + 0.00963582657277584, + 0.0073469714261591434, + 0.011224868707358837, + 0.008654465898871422, + 0.007033444941043854, + 0.008818321861326694, + 0.011336891911923885, + 0.01167216058820486, + 0.01637585274875164, + 0.010979917831718922, + 0.018187806010246277, + 0.012528245337307453, + 0.012318708933889866, + 0.009788175113499165, + 0.01734568178653717, + 0.01636739820241928, + 0.017302801832556725, + 0.01205849926918745, + 0.03616055101156235, + 0.0178737323731184, + 0.013299946673214436, + 0.01586245931684971, + 0.01465721894055605, + 0.010762897320091724, + 0.014736752025783062, + 0.008751587942242622, + 0.011036382056772709, + 0.014226235449314117, + 0.010225428268313408, + 0.008626438677310944, + 0.0064142243936657906, + 0.007539719808846712, + 0.005772473756223917, + 0.004528710152953863, + 0.005013884045183659, + 0.006083938758820295, + 0.0051049175672233105, + 0.004590109921991825, + 0.009137382730841637, + 0.013260501436889172, + 0.009107149206101894, + 0.006620513275265694, + 0.007938527502119541, + 0.01040320098400116, + 0.00633765896782279, + 0.006399362348020077, + 0.005352432373911142, + 0.006139381323009729, + 0.006063494831323624, + 0.007946701720356941, + 0.006205392070114613, + 0.0057741678319871426, + 0.004995009861886501, + 0.005283209960907698, + 0.009017289616167545, + 0.012984445318579674, + 0.009377424605190754, + 0.017843205481767654, + 0.009957807138562202, + 0.012810765765607357, + 0.011834193021059036, + 0.00804308895021677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.07849262654781342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361869648098946 + ], + "opposite_logprob_tokens": [ + -3.6513619422912598 + ], + "logprob_reward": -0.026361869648098946, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361869648098946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23064714670181274, + 0.06269285827875137, + 0.10129231959581375, + 0.09272104501724243, + 0.05730170011520386, + 0.06227584555745125, + 0.09298492968082428, + 0.11800896376371384, + 0.1250050961971283, + 0.05707011744379997 + ], + "credit_map": [ + 0.023497454822063446, + 0.015245012938976288, + 0.040308281779289246, + 0.03122391551733017, + 0.05438784882426262, + 0.02533743716776371, + 0.022950150072574615, + 0.0176970437169075, + 0.008142588660120964, + 0.007619292475283146, + 0.00853040348738432, + 0.0067922575399279594, + 0.010114866308867931, + 0.008288010954856873, + 0.007191385142505169, + 0.006014047656208277, + 0.0090408930554986, + 0.008921205997467041, + 0.012095934711396694, + 0.009616459719836712, + 0.012600102461874485, + 0.011694764718413353, + 0.014376113191246986, + 0.022946830838918686, + 0.011244431138038635, + 0.009758763015270233, + 0.012247108854353428, + 0.008034031838178635, + 0.01755058951675892, + 0.013260633684694767, + 0.010140851140022278, + 0.010484622791409492, + 0.005931470077484846, + 0.008373282849788666, + 0.006131243426352739, + 0.004625317174941301, + 0.012959767132997513, + 0.007236355450004339, + 0.006755429785698652, + 0.005288825370371342, + 0.008321127854287624, + 0.007592617999762297, + 0.008849888108670712, + 0.009238685481250286, + 0.006906850263476372, + 0.007743261754512787, + 0.007833210751414299, + 0.0057901982218027115, + 0.0112300468608737, + 0.010339903645217419, + 0.012892876751720905, + 0.016638249158859253, + 0.012417527846992016, + 0.012283611111342907, + 0.010470031760632992, + 0.006712683476507664, + 0.016526425257325172, + 0.02333669550716877, + 0.016160324215888977, + 0.008229807950556278, + 0.011832555755972862, + 0.020878510549664497, + 0.012526479549705982, + 0.008518165908753872, + 0.020555520430207253, + 0.018368525430560112, + 0.019484257325530052, + 0.016205105930566788, + 0.011842343956232071, + 0.016451751813292503, + 0.011463306844234467, + 0.010634281672537327, + 0.007310531567782164, + 0.0086613018065691, + 0.006682541221380234, + 0.006721053272485733, + 0.006760371383279562, + 0.008114799857139587, + 0.008159012533724308, + 0.004660507198423147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.026361869648098946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637348957359791 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637348957359791, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637348957359791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15919233858585358, + 0.12980957329273224, + 0.1301421821117401, + 0.07003938406705856, + 0.10994476824998856, + 0.08103131502866745, + 0.08982884883880615, + 0.07400812953710556, + 0.07830734550952911, + 0.07769611477851868 + ], + "credit_map": [ + 0.01980559714138508, + 0.013129827566444874, + 0.016972919926047325, + 0.018685920163989067, + 0.02703380584716797, + 0.01762326806783676, + 0.02987365610897541, + 0.016067340970039368, + 0.015561849810183048, + 0.00924873910844326, + 0.02303062193095684, + 0.007392581552267075, + 0.01923827826976776, + 0.03566433861851692, + 0.0089044701308012, + 0.01076868362724781, + 0.013798916712403297, + 0.012044924311339855, + 0.02000923827290535, + 0.011974017135798931, + 0.027933944016695023, + 0.01700824312865734, + 0.014085205271840096, + 0.013287671841681004, + 0.008141580037772655, + 0.00954347662627697, + 0.010609565302729607, + 0.004537919536232948, + 0.009191748686134815, + 0.009531859308481216, + 0.010010926052927971, + 0.008472306653857231, + 0.011686134152114391, + 0.013300697319209576, + 0.014871183782815933, + 0.012502969242632389, + 0.02144155651330948, + 0.013488663360476494, + 0.013388196006417274, + 0.009265356697142124, + 0.012778722681105137, + 0.009853695519268513, + 0.012394278310239315, + 0.013957828283309937, + 0.008825419470667839, + 0.010464230552315712, + 0.008157961070537567, + 0.004599167499691248, + 0.010586881078779697, + 0.010785527527332306, + 0.014831040054559708, + 0.012841525487601757, + 0.009061205200850964, + 0.011981719173491001, + 0.010586610995233059, + 0.009154331870377064, + 0.010526436381042004, + 0.009236793033778667, + 0.009621510282158852, + 0.013432752341032028, + 0.006485404912382364, + 0.009902061894536018, + 0.006537475623190403, + 0.008265684358775616, + 0.013598546385765076, + 0.009165671654045582, + 0.010151044465601444, + 0.012460656464099884, + 0.008262046612799168, + 0.011269496753811836, + 0.007728608325123787, + 0.005671264603734016, + 0.00861001294106245, + 0.011209615506231785, + 0.011082863435149193, + 0.01150244940072298, + 0.006127007305622101, + 0.012174941599369049, + 0.007689723279327154, + 0.009299494326114655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.007637348957359791 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353858411312103 + ], + "opposite_logprob_tokens": [ + -3.6513538360595703 + ], + "logprob_reward": -0.026353858411312103, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353858411312103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1316438913345337, + 0.16550980508327484, + 0.14124390482902527, + 0.08568094670772552, + 0.12402678281068802, + 0.032260481268167496, + 0.08862202614545822, + 0.061431508511304855, + 0.03131075203418732, + 0.13826994597911835 + ], + "credit_map": [ + 0.01414589211344719, + 0.010823068208992481, + 0.02362823113799095, + 0.027985144406557083, + 0.016861697658896446, + 0.014118754304945469, + 0.013248843140900135, + 0.01083226315677166, + 0.01863222010433674, + 0.012492797337472439, + 0.02947409637272358, + 0.011268331669270992, + 0.023225024342536926, + 0.047846805304288864, + 0.011384105309844017, + 0.011186428368091583, + 0.0140170156955719, + 0.011329462751746178, + 0.01606631837785244, + 0.014703912660479546, + 0.025061311200261116, + 0.01657804846763611, + 0.013624018058180809, + 0.029863815754652023, + 0.008853206411004066, + 0.008654286153614521, + 0.011786540038883686, + 0.008497724309563637, + 0.01479258295148611, + 0.010826237499713898, + 0.009042675606906414, + 0.013227698393166065, + 0.016733307391405106, + 0.021444937214255333, + 0.01578282006084919, + 0.012058376334607601, + 0.016592247411608696, + 0.0168417077511549, + 0.012463697232306004, + 0.01210970152169466, + 0.004806187469512224, + 0.00342159578576684, + 0.0037853901740163565, + 0.006272255443036556, + 0.0031114083249121904, + 0.004952494986355305, + 0.002987137297168374, + 0.0029240092262625694, + 0.011724726296961308, + 0.011114528402686119, + 0.01224383432418108, + 0.01257918681949377, + 0.012288684025406837, + 0.011988064274191856, + 0.010556585155427456, + 0.006126415450125933, + 0.006445075385272503, + 0.007819853723049164, + 0.008605070412158966, + 0.010019549168646336, + 0.009385582059621811, + 0.008016650564968586, + 0.005937944166362286, + 0.005201785359531641, + 0.004738662391901016, + 0.004390001762658358, + 0.004465149715542793, + 0.00419773580506444, + 0.0031280573457479477, + 0.004982836544513702, + 0.003017512382939458, + 0.002390798879787326, + 0.012647358700633049, + 0.023404523730278015, + 0.01817845180630684, + 0.009600178338587284, + 0.01983274146914482, + 0.02228032797574997, + 0.020595386624336243, + 0.011730973608791828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.026353858411312103 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged general wearing ancient armor made of bronze and wrought iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a dimly lit military tent with a dark canvas background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera framed as a steady, slightly upward medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general tilt a bronze wine jug to pour clear amber liquor into a stoneware bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and dim, creating a sharp contrast of light and dark on the subject's features?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022003266494721174 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.0022003266494721174, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022003266494721174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14578291773796082, + 0.07352712750434875, + 0.057151615619659424, + 0.06392449140548706, + 0.05975857377052307, + 0.09638912230730057, + 0.031440507620573044, + 0.3909352719783783, + 0.0260261632502079, + 0.055064182728528976 + ], + "credit_map": [ + 0.017504841089248657, + 0.01313302107155323, + 0.03042955882847309, + 0.011123103089630604, + 0.017597094178199768, + 0.0212127435952425, + 0.017372507601976395, + 0.017410049214959145, + 0.007769584655761719, + 0.005900677293539047, + 0.014229019172489643, + 0.0046961246989667416, + 0.011802079156041145, + 0.014081446453928947, + 0.007859851233661175, + 0.0071883476339280605, + 0.006652615498751402, + 0.005419362802058458, + 0.006881326902657747, + 0.005996593739837408, + 0.010417900048196316, + 0.008117863908410072, + 0.006958695128560066, + 0.006707258056849241, + 0.007353485561907291, + 0.007487634662538767, + 0.011076800525188446, + 0.004615952726453543, + 0.009823456406593323, + 0.00899674091488123, + 0.006998464465141296, + 0.007571962662041187, + 0.005675411783158779, + 0.006263840012252331, + 0.0073000844568014145, + 0.006154647562652826, + 0.01245917845517397, + 0.008097924292087555, + 0.0064332676120102406, + 0.007374223787337542, + 0.012865596450865269, + 0.010603864677250385, + 0.01453925296664238, + 0.012076681479811668, + 0.014857538044452667, + 0.013212570920586586, + 0.01037536934018135, + 0.007858255878090858, + 0.0049167051911354065, + 0.004234082531183958, + 0.004472593776881695, + 0.0045290314592421055, + 0.0034169533755630255, + 0.003908292856067419, + 0.003310450119897723, + 0.002652401104569435, + 0.06763368099927902, + 0.06018226221203804, + 0.050804369151592255, + 0.04551747813820839, + 0.04174898937344551, + 0.05620908364653587, + 0.04625289514660835, + 0.022586561739444733, + 0.002368554938584566, + 0.0027541876770555973, + 0.003030460560694337, + 0.004477718845009804, + 0.004144362639635801, + 0.003230697475373745, + 0.0035231478977948427, + 0.0024970355443656445, + 0.005564714781939983, + 0.008723272942006588, + 0.0073644183576107025, + 0.005963553674519062, + 0.006685145664960146, + 0.008793708868324757, + 0.006684595253318548, + 0.0052847773768007755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.0022003266494721174 + } + ], + "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/0000128/group_6/index_tensor([4822, 4822])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_6/index_tensor([4822, 4822])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b476be99ca2cf93d22a4ccbb28a43635481b9e05 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_6/index_tensor([4822, 4822])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle tracking shot glides smoothly alongside a cyclist, capturing her energetic movements along a classic Parisian cobblestone street, the camera keeping pace with her rapid progress. The cyclist is a young woman, her brown hair pulled back in a slightly messy bun, with a few strands escaping and dancing in the wind. Dressed in a timeless blue and white striped long-sleeved shirt and dark pants, she leans forward over the handlebars of her vintage bicycle, striking a focused pose. Her legs beat in a powerful, steady rhythm, her body swaying slightly with the effort, the wheel bouncing gently on the uneven, time-worn stones of the street. In the background, the majestic Eiffel Tower looms majestically, its iconic grid structure dominating the upper frame, slightly softened by the golden haze of a sunny afternoon. The bright, warm sunlight casts long, dancing shadows and creates subtle lens flares, bathing the entire scene in a vibrant golden light, evoking a joyful and exciting atmosphere of freedom.", + "seed": 14380, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_6/index_4822_seed_14380_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": -0.18292296417057513, + "vlm_logprob": -0.18292296417057513, + "vlm_yes_no_margin": 3.174999976158142, + "vlm_token_credit": -0.18292296417057513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4998016953468323 + ], + "opposite_logprob_tokens": [ + -1.3748016357421875 + ], + "logprob_reward": -0.4998016953468323, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.4998016953468323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6027792692184448, + 0.04581139236688614, + 0.05442285165190697, + 0.021338025107979774, + 0.0808737501502037, + 0.020231273025274277, + 0.07836833596229553, + 0.02617703564465046, + 0.03893275558948517, + 0.031065259128808975 + ], + "credit_map": [ + 0.03692179545760155, + 0.020012954249978065, + 0.23606672883033752, + 0.034028444439172745, + 0.019712716341018677, + 0.027085598558187485, + 0.024720748886466026, + 0.20423030853271484, + 0.003986257594078779, + 0.006213690619915724, + 0.0077186389826238155, + 0.006039874162524939, + 0.0052400981076061726, + 0.006634279619902372, + 0.005403578747063875, + 0.0045749773271381855, + 0.005859069991856813, + 0.009176871739327908, + 0.01168838795274496, + 0.005127325188368559, + 0.004202712327241898, + 0.004837547428905964, + 0.005598139017820358, + 0.00793280079960823, + 0.0022210420574992895, + 0.003971250727772713, + 0.0029661227017641068, + 0.0020517290104180574, + 0.003288297913968563, + 0.0018188239773735404, + 0.0030710557475686073, + 0.0019497039029374719, + 0.012404162436723709, + 0.008207944221794605, + 0.010630267672240734, + 0.010021855123341084, + 0.009473597630858421, + 0.0091552110388875, + 0.012073216959834099, + 0.008907507173717022, + 0.0034403649624437094, + 0.003070281818509102, + 0.002853056648746133, + 0.001817787648178637, + 0.0021189048420637846, + 0.0016323121963068843, + 0.0030424101278185844, + 0.0022561573423445225, + 0.01497802883386612, + 0.011444134637713432, + 0.01187538169324398, + 0.0066780950874090195, + 0.008390821516513824, + 0.007034108508378267, + 0.006589876022189856, + 0.011377893388271332, + 0.0031223814003169537, + 0.0031718576792627573, + 0.003727603005245328, + 0.002878113416954875, + 0.0020875022746622562, + 0.0023564219009131193, + 0.006257012020796537, + 0.002576146274805069, + 0.002467745216563344, + 0.0027248968835920095, + 0.004401535727083683, + 0.002780917566269636, + 0.004159125499427319, + 0.007875756360590458, + 0.0039571658708155155, + 0.010565614327788353, + 0.0038033134769648314, + 0.00419809902086854, + 0.004536048974841833, + 0.004394779913127422, + 0.0017827190458774567, + 0.0032188326586037874, + 0.005823219660669565, + 0.003308249171823263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.4998016953468323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871386647224426 + ], + "opposite_logprob_tokens": [ + -1.1371386051177979 + ], + "logprob_reward": -0.3871386647224426, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3871386647224426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17867837846279144, + 0.05223703756928444, + 0.024059079587459564, + 0.0656365305185318, + 0.08999047428369522, + 0.03616666793823242, + 0.193674698472023, + 0.12031389027833939, + 0.17217925190925598, + 0.06706396490335464 + ], + "credit_map": [ + 0.011602809652686119, + 0.021218543872237206, + 0.048574142158031464, + 0.010737051256000996, + 0.01699279434978962, + 0.009785130620002747, + 0.016063516959547997, + 0.04370439052581787, + 0.005315838381648064, + 0.00878229457885027, + 0.008459662087261677, + 0.004775250796228647, + 0.006837158929556608, + 0.005650458391755819, + 0.006385822780430317, + 0.006030550692230463, + 0.0020770984701812267, + 0.003070011967793107, + 0.004906798247247934, + 0.002782914787530899, + 0.002597576705738902, + 0.002196046058088541, + 0.0032342218328267336, + 0.0031944098882377148, + 0.007684763520956039, + 0.017261484637856483, + 0.007237488403916359, + 0.0072358110919594765, + 0.007056795991957188, + 0.007520855870097876, + 0.005048385821282864, + 0.006590947043150663, + 0.007400371599942446, + 0.010541182942688465, + 0.008554869331419468, + 0.015405350364744663, + 0.012372699566185474, + 0.00750983040779829, + 0.01705610379576683, + 0.011150065809488297, + 0.007793414406478405, + 0.004436255432665348, + 0.005276271142065525, + 0.0024970583617687225, + 0.003626454621553421, + 0.003181559732183814, + 0.004509181249886751, + 0.004846473224461079, + 0.05213943496346474, + 0.019983477890491486, + 0.029688114300370216, + 0.0131293423473835, + 0.020864488556981087, + 0.02076602727174759, + 0.01549777016043663, + 0.02160602994263172, + 0.021698690950870514, + 0.021670861169695854, + 0.013240200467407703, + 0.011207975447177887, + 0.015003264881670475, + 0.011641581542789936, + 0.015971871092915535, + 0.009879445657134056, + 0.010116029530763626, + 0.008827182464301586, + 0.01919684000313282, + 0.009685369208455086, + 0.01360337994992733, + 0.03446364402770996, + 0.022038834169507027, + 0.054247964173555374, + 0.004571184050291777, + 0.008760809898376465, + 0.006415438838303089, + 0.007866433821618557, + 0.010676407255232334, + 0.008568485267460346, + 0.013369763270020485, + 0.00683544110506773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.3871386647224426 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748977117240429 + ], + "opposite_logprob_tokens": [ + -5.006749153137207 + ], + "logprob_reward": -0.006748977117240429, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748977117240429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6414625644683838, + 0.05076378211379051, + 0.05895392969250679, + 0.014442832209169865, + 0.04803632199764252, + 0.0245500635355711, + 0.08091074228286743, + 0.04689730703830719, + 0.01763959228992462, + 0.016342833638191223 + ], + "credit_map": [ + 0.0420057587325573, + 0.019596125930547714, + 0.25923022627830505, + 0.032372381538152695, + 0.016705632209777832, + 0.028034456074237823, + 0.018607934936881065, + 0.22491003572940826, + 0.003946287557482719, + 0.00703892158344388, + 0.005902314558625221, + 0.006018841173499823, + 0.006692508235573769, + 0.007328088860958815, + 0.00842612236738205, + 0.005410703830420971, + 0.004741130396723747, + 0.006189926061779261, + 0.011260847561061382, + 0.006656288169324398, + 0.0073790946044027805, + 0.005108213517814875, + 0.009244563058018684, + 0.008373867720365524, + 0.001250913250260055, + 0.0023579068947583437, + 0.001565330196171999, + 0.0021213816944509745, + 0.0016762677114456892, + 0.0020324867218732834, + 0.0016838093288242817, + 0.0017547371098771691, + 0.007096495013684034, + 0.006030866876244545, + 0.0057142386212944984, + 0.005983822513371706, + 0.006308764219284058, + 0.005225513130426407, + 0.006116609089076519, + 0.005560016725212336, + 0.005830984562635422, + 0.002818629378452897, + 0.004160263109952211, + 0.0017417453927919269, + 0.002090304857119918, + 0.0016264008590951562, + 0.0036680654156953096, + 0.002613670425489545, + 0.008738277480006218, + 0.011738341301679611, + 0.008950899355113506, + 0.0076795099303126335, + 0.01639196090400219, + 0.009203426539897919, + 0.009339661337435246, + 0.008868671953678131, + 0.009223595261573792, + 0.006758328061550856, + 0.006238406989723444, + 0.003306090133264661, + 0.004454120993614197, + 0.004276154097169638, + 0.008957826532423496, + 0.003682787762954831, + 0.0014991648495197296, + 0.0016050264239311218, + 0.0017083819257095456, + 0.0019273809157311916, + 0.0026098834350705147, + 0.002329216105863452, + 0.0031374001409858465, + 0.002823140937834978, + 0.001337077235803008, + 0.0024896450340747833, + 0.0019713840447366238, + 0.0018989868694916368, + 0.0018053230596706271, + 0.0019151108572259545, + 0.0029497595969587564, + 0.001975546358153224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.006748977117240429 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190834015607834 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190834015607834, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190834015607834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06302265077829361, + 0.13372358679771423, + 0.058860909193754196, + 0.08050483465194702, + 0.10923171043395996, + 0.11258988082408905, + 0.07647430896759033, + 0.02586929127573967, + 0.3243362009525299, + 0.015386655926704407 + ], + "credit_map": [ + 0.005628910847008228, + 0.007252523209899664, + 0.010057322680950165, + 0.006985762622207403, + 0.009976032190024853, + 0.004774549975991249, + 0.010277063585817814, + 0.008070483803749084, + 0.012595608830451965, + 0.024699218571186066, + 0.02268940955400467, + 0.012320387177169323, + 0.013584173284471035, + 0.01595892943441868, + 0.015559428371489048, + 0.016316430643200874, + 0.008235104382038116, + 0.016260359436273575, + 0.008794928900897503, + 0.004448503255844116, + 0.004389347042888403, + 0.0063913743942976, + 0.004204504191875458, + 0.006136787123978138, + 0.009040920995175838, + 0.021460581570863724, + 0.009311346337199211, + 0.008493270725011826, + 0.008802921511232853, + 0.00988889392465353, + 0.005721422843635082, + 0.0077854725532233715, + 0.011890416964888573, + 0.011482586152851582, + 0.012675185687839985, + 0.011043204925954342, + 0.014894509688019753, + 0.014942568726837635, + 0.017361072823405266, + 0.014942164532840252, + 0.028481340035796165, + 0.01616024225950241, + 0.020142944529652596, + 0.006602199282497168, + 0.008217215538024902, + 0.01034222636371851, + 0.008099407888948917, + 0.014544298872351646, + 0.016071703284978867, + 0.01087454054504633, + 0.010346881113946438, + 0.005221858620643616, + 0.007860054261982441, + 0.007852429524064064, + 0.0057324315421283245, + 0.01251441054046154, + 0.0025138775818049908, + 0.0035997030790895224, + 0.003869140287861228, + 0.0030198104213923216, + 0.0036363506224006414, + 0.0023828439880162477, + 0.004624560009688139, + 0.0022230048198252916, + 0.01959831267595291, + 0.0174049474298954, + 0.03464599326252937, + 0.016242777928709984, + 0.02513451687991619, + 0.07055380940437317, + 0.03388402611017227, + 0.10687178373336792, + 0.0023793529253453016, + 0.0031588159035891294, + 0.0017910812748596072, + 0.0014469717862084508, + 0.001256601302884519, + 0.0017267928924411535, + 0.0021776899229735136, + 0.0014493487542495131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.014190834015607834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734649650752544 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734649650752544, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734649650752544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3165976107120514, + 0.03132513165473938, + 0.021171795204281807, + 0.03394771367311478, + 0.06507892161607742, + 0.08208553493022919, + 0.39006155729293823, + 0.012272887863218784, + 0.03199031203985214, + 0.015468559227883816 + ], + "credit_map": [ + 0.021694350987672806, + 0.010214054025709629, + 0.12473288923501968, + 0.016535528004169464, + 0.011045791208744049, + 0.015434560365974903, + 0.010226373560726643, + 0.10671405494213104, + 0.003032319014891982, + 0.003051259322091937, + 0.006701376289129257, + 0.0030800187960267067, + 0.003530884860083461, + 0.0036692682188004255, + 0.004147201310843229, + 0.004112803842872381, + 0.0023952913470566273, + 0.004737779963761568, + 0.0035916303750127554, + 0.0016301824944093823, + 0.0016074914019554853, + 0.002547814277932048, + 0.0017186199547722936, + 0.0029429851565510035, + 0.0039010026957839727, + 0.008367056958377361, + 0.003734137862920761, + 0.0037397227715700865, + 0.004029244650155306, + 0.002982217585667968, + 0.0033202364575117826, + 0.0038740970194339752, + 0.012912756763398647, + 0.009375925175845623, + 0.008375093340873718, + 0.004967748653143644, + 0.007083601783961058, + 0.008421496488153934, + 0.00756586529314518, + 0.006376428529620171, + 0.020568156614899635, + 0.012338997796177864, + 0.014431309886276722, + 0.005226758774369955, + 0.00611340394243598, + 0.00724249891936779, + 0.006282295566052198, + 0.009882119484245777, + 0.11848045885562897, + 0.05234631523489952, + 0.06283382326364517, + 0.023658810183405876, + 0.024244150146842003, + 0.03437019884586334, + 0.023123396560549736, + 0.05100439861416817, + 0.0027855075895786285, + 0.0013734045205637813, + 0.0015383469872176647, + 0.000990389846265316, + 0.0010836449218913913, + 0.0011056590592488647, + 0.002370154485106468, + 0.0010257798712700605, + 0.0030316822230815887, + 0.003207054454833269, + 0.003615091321989894, + 0.0030441591516137123, + 0.004670613445341587, + 0.004156142007559538, + 0.00430514058098197, + 0.005960427690297365, + 0.004115489311516285, + 0.0017990544438362122, + 0.002336464822292328, + 0.0012076866114512086, + 0.001225514686666429, + 0.001967751421034336, + 0.0010404710192233324, + 0.0017761267954483628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.006734649650752544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot glides smoothly alongside a cyclist, capturing her energetic movements along a classic Parisian cobblestone street, the camera keeping pace with her rapid progress. The cyclist is a young woman, her brown hair pulled back in a slightly messy bun, with a few strands escaping and dancing in the wind. Dressed in a timeless blue and white striped long-sleeved shirt and dark pants, she leans forward over the handlebars of her vintage bicycle, striking a focused pose. Her legs beat in a powerful, steady rhythm, her body swaying slightly with the effort, the wheel bouncing gently on the uneven, time-worn stones of the street. In the background, the majestic Eiffel Tower looms majestically, its iconic grid structure dominating the upper frame, slightly softened by the golden haze of a sunny afternoon. The bright, warm sunlight casts long, dancing shadows and creates subtle lens flares, bathing the entire scene in a vibrant golden light, evoking a joyful and exciting atmosphere of freedom.", + "seed": 14381, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_6/index_4822_seed_14381_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": -0.1174314002506435, + "vlm_logprob": -0.1174314002506435, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.1174314002506435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22202283143997192 + ], + "opposite_logprob_tokens": [ + -2.222022771835327 + ], + "logprob_reward": -0.22202283143997192, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.22202283143997192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12734894454479218, + 0.07164066284894943, + 0.07462821900844574, + 0.05296015366911888, + 0.0824308767914772, + 0.17537005245685577, + 0.05985986068844795, + 0.1566474437713623, + 0.1293148249387741, + 0.06979893147945404 + ], + "credit_map": [ + 0.016777269542217255, + 0.011394436471164227, + 0.03137936070561409, + 0.00900205411016941, + 0.012735809199512005, + 0.01808750629425049, + 0.013016748242080212, + 0.014955763705074787, + 0.009018870070576668, + 0.006447337567806244, + 0.00837141927331686, + 0.007221614010632038, + 0.012608066201210022, + 0.007174431346356869, + 0.013905787840485573, + 0.0068931314162909985, + 0.00977136567234993, + 0.00845272559672594, + 0.009308942593634129, + 0.005948411766439676, + 0.008058623410761356, + 0.015194646082818508, + 0.005328093655407429, + 0.012565414421260357, + 0.005864489357918501, + 0.004107981454581022, + 0.012057904154062271, + 0.007220847997814417, + 0.004526244476437569, + 0.007823538035154343, + 0.004716232419013977, + 0.006642916239798069, + 0.00866620521992445, + 0.009374898858368397, + 0.01341450959444046, + 0.007778011728078127, + 0.009148071520030499, + 0.015039539895951748, + 0.008571034297347069, + 0.010438607074320316, + 0.017995862290263176, + 0.02156461961567402, + 0.015518232248723507, + 0.0308462455868721, + 0.017203811556100845, + 0.02062537893652916, + 0.03221248835325241, + 0.019403407350182533, + 0.007624596357345581, + 0.007662900257855654, + 0.007231080438941717, + 0.00849882047623396, + 0.0052721090614795685, + 0.008474785834550858, + 0.007031734101474285, + 0.00806383416056633, + 0.01314258761703968, + 0.012827309779822826, + 0.04999155551195145, + 0.011577085591852665, + 0.013078351505100727, + 0.009167575277388096, + 0.011080298572778702, + 0.03578268736600876, + 0.015355250798165798, + 0.02775220200419426, + 0.01298526395112276, + 0.01339144166558981, + 0.013544206507503986, + 0.011186539195477962, + 0.011902523227036, + 0.023197397589683533, + 0.009307951666414738, + 0.007657923270016909, + 0.008619970642030239, + 0.010022531263530254, + 0.0055710067972540855, + 0.008663976565003395, + 0.013751129619777203, + 0.00620444118976593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.22202283143997192 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31351494789123535 + ], + "opposite_logprob_tokens": [ + -1.3135149478912354 + ], + "logprob_reward": -0.31351494789123535, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31351494789123535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10661476850509644, + 0.07420095801353455, + 0.03147018328309059, + 0.03961940109729767, + 0.06086791306734085, + 0.17552582919597626, + 0.025613337755203247, + 0.28753313422203064, + 0.08235369622707367, + 0.11620070785284042 + ], + "credit_map": [ + 0.014860628172755241, + 0.010857462882995605, + 0.017363887280225754, + 0.0070607843808829784, + 0.016416240483522415, + 0.014218069612979889, + 0.011937051080167294, + 0.013900645077228546, + 0.008213918656110764, + 0.0067417798563838005, + 0.010222770273685455, + 0.008593575097620487, + 0.011898764409124851, + 0.00942150317132473, + 0.01124060433357954, + 0.007868043147027493, + 0.004069668240845203, + 0.005845445208251476, + 0.0050826892256736755, + 0.002864283276721835, + 0.002854257356375456, + 0.0035066911950707436, + 0.0029831030406057835, + 0.004264047369360924, + 0.005131533369421959, + 0.00570217240601778, + 0.00751589797437191, + 0.004455392248928547, + 0.00445970194414258, + 0.004425754304975271, + 0.0028470298275351524, + 0.0050819204188883305, + 0.006684710271656513, + 0.009638906456530094, + 0.010084094479680061, + 0.009130256250500679, + 0.005602904595434666, + 0.007147754542529583, + 0.007205118425190449, + 0.005374170374125242, + 0.018238665536046028, + 0.01963627152144909, + 0.020459521561861038, + 0.02700091525912285, + 0.017126789316534996, + 0.018513882532715797, + 0.03473629057407379, + 0.019813498482108116, + 0.003120679408311844, + 0.0028092877473682165, + 0.003551259869709611, + 0.0020792214199900627, + 0.0038516265340149403, + 0.0026602752041071653, + 0.003519719000905752, + 0.004021269269287586, + 0.021582970395684242, + 0.0227093193680048, + 0.10036961734294891, + 0.017272021621465683, + 0.019805951043963432, + 0.011882144957780838, + 0.014482279308140278, + 0.07942885905504227, + 0.012423033826053143, + 0.006915308069437742, + 0.014141758903861046, + 0.007073015905916691, + 0.006790406070649624, + 0.007609962020069361, + 0.006667580921202898, + 0.02073262818157673, + 0.013032826595008373, + 0.012307161465287209, + 0.01845518872141838, + 0.021614275872707367, + 0.008904434740543365, + 0.011816821061074734, + 0.01810714788734913, + 0.011962848715484142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.31351494789123535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198120683431625 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198120683431625, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198120683431625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21991388499736786, + 0.10988809168338776, + 0.24332574009895325, + 0.02318732626736164, + 0.034467123448848724, + 0.04342258721590042, + 0.04641813412308693, + 0.14942698180675507, + 0.0647195503115654, + 0.0652305856347084 + ], + "credit_map": [ + 0.031855300068855286, + 0.03476174175739288, + 0.022716784849762917, + 0.020257078111171722, + 0.04007558897137642, + 0.02570335939526558, + 0.026620138436555862, + 0.01792389340698719, + 0.01649453118443489, + 0.014657019637525082, + 0.01155888382345438, + 0.009516073390841484, + 0.02510232850909233, + 0.009503738954663277, + 0.01076668594032526, + 0.012288841418921947, + 0.03975767269730568, + 0.07550332695245743, + 0.018629571422934532, + 0.030782412737607956, + 0.025616008788347244, + 0.018479134887456894, + 0.016620920971035957, + 0.017936697229743004, + 0.002739817602559924, + 0.0035876601468771696, + 0.0026233417447656393, + 0.0025196392089128494, + 0.0030274412129074335, + 0.0038522412069141865, + 0.0022441770415753126, + 0.0025930071715265512, + 0.003847508691251278, + 0.004725733306258917, + 0.0063501326367259026, + 0.0045388625003397465, + 0.002963179023936391, + 0.004233114421367645, + 0.002845302689820528, + 0.004963289014995098, + 0.004408207256346941, + 0.004910002462565899, + 0.00409207446500659, + 0.006712460890412331, + 0.004515489563345909, + 0.0046887388452887535, + 0.009471896104514599, + 0.0046237194910645485, + 0.0059907264076173306, + 0.00582222081720829, + 0.006945461500436068, + 0.004641035106033087, + 0.004623770713806152, + 0.005786910653114319, + 0.006697129923850298, + 0.005910877138376236, + 0.01259537786245346, + 0.01150935422629118, + 0.04947517067193985, + 0.009173997677862644, + 0.008991598151624203, + 0.007481357082724571, + 0.011800485663115978, + 0.03839964047074318, + 0.008399838581681252, + 0.009729534387588501, + 0.00560954213142395, + 0.008058031089603901, + 0.007202610373497009, + 0.008315560407936573, + 0.006394420750439167, + 0.011010011658072472, + 0.009139942936599255, + 0.007450804580003023, + 0.010262571275234222, + 0.011146222241222858, + 0.004652731586247683, + 0.0076422314159572124, + 0.008779898285865784, + 0.0061561851762235165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.014198120683431625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793808236718178 + ], + "opposite_logprob_tokens": [ + -3.5297937393188477 + ], + "logprob_reward": -0.029793808236718178, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793808236718178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28005602955818176, + 0.10339953750371933, + 0.1631440669298172, + 0.039217427372932434, + 0.04111498221755028, + 0.10160557925701141, + 0.0809917002916336, + 0.07481997460126877, + 0.06332360953092575, + 0.05232710763812065 + ], + "credit_map": [ + 0.044170476496219635, + 0.024341963231563568, + 0.028672760352492332, + 0.013782764784991741, + 0.057272203266620636, + 0.048793893307447433, + 0.023218756541609764, + 0.03980322927236557, + 0.015372496098279953, + 0.008399849757552147, + 0.011677864007651806, + 0.008797851391136646, + 0.02181999757885933, + 0.01182604394853115, + 0.013263314962387085, + 0.012242123484611511, + 0.018306151032447815, + 0.009445849806070328, + 0.020559685304760933, + 0.013726022094488144, + 0.01976354420185089, + 0.040787920355796814, + 0.01303186547011137, + 0.027523016557097435, + 0.004079912323504686, + 0.0033558388240635395, + 0.008444634266197681, + 0.0029468215070664883, + 0.005118366796523333, + 0.006466131191700697, + 0.0032207623589783907, + 0.005584959406405687, + 0.005112907849252224, + 0.008815585635602474, + 0.005115985404700041, + 0.005702463444322348, + 0.0035379831679165363, + 0.005162097979336977, + 0.0029377411119639874, + 0.004730219952762127, + 0.010513444431126118, + 0.011433040723204613, + 0.009072979912161827, + 0.01605003885924816, + 0.009711976163089275, + 0.010973628610372543, + 0.0206845011562109, + 0.013165969401597977, + 0.010469249449670315, + 0.0093391677364707, + 0.012298832647502422, + 0.00947535503655672, + 0.008667300455272198, + 0.009355107322335243, + 0.010548719204962254, + 0.010837975889444351, + 0.006241868250072002, + 0.006176154129207134, + 0.021198159083724022, + 0.006841672584414482, + 0.004906215239316225, + 0.004086937755346298, + 0.006770172156393528, + 0.01859879307448864, + 0.009289971552789211, + 0.01208183541893959, + 0.007148881442844868, + 0.007052559405565262, + 0.005096090491861105, + 0.006238551810383797, + 0.006148043554276228, + 0.010267678648233414, + 0.009070956148207188, + 0.006623605731874704, + 0.007477341219782829, + 0.006552352104336023, + 0.004849689546972513, + 0.007843338884413242, + 0.005513929296284914, + 0.004395895171910524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.029793808236718178 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762729300186038 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.00762729300186038, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762729300186038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1941012293100357, + 0.06052490323781967, + 0.0642905905842781, + 0.21403442323207855, + 0.034863799810409546, + 0.01999819651246071, + 0.06702462583780289, + 0.2141561210155487, + 0.07021500915288925, + 0.06079109013080597 + ], + "credit_map": [ + 0.02526722475886345, + 0.02158704586327076, + 0.04788545146584511, + 0.011577091179788113, + 0.02175016887485981, + 0.027076102793216705, + 0.01815887913107872, + 0.020799262449145317, + 0.007549075875431299, + 0.0056898570619523525, + 0.005446828436106443, + 0.007439908571541309, + 0.009490306489169598, + 0.007342572323977947, + 0.011035973206162453, + 0.006530380807816982, + 0.007336308714002371, + 0.005833894945681095, + 0.00747317261993885, + 0.006063681561499834, + 0.008340755477547646, + 0.016528381034731865, + 0.005831835325807333, + 0.006882561836391687, + 0.0216290932148695, + 0.014883138239383698, + 0.036809004843235016, + 0.029717804864048958, + 0.03182539716362953, + 0.03495633974671364, + 0.016886917874217033, + 0.02732672169804573, + 0.004073138348758221, + 0.0038192023057490587, + 0.005441858433187008, + 0.003537616925314069, + 0.0036500361748039722, + 0.004725880455225706, + 0.0033193116541951895, + 0.00629675667732954, + 0.0019985674880445004, + 0.0028945140074938536, + 0.0026774003636091948, + 0.002667561871930957, + 0.0017149156192317605, + 0.002265467308461666, + 0.0024252997245639563, + 0.0033544700127094984, + 0.008188006468117237, + 0.006777075584977865, + 0.00894199125468731, + 0.00824249442666769, + 0.008870155550539494, + 0.007366099860519171, + 0.00850154273211956, + 0.010137260891497135, + 0.01757918857038021, + 0.015529925003647804, + 0.06196962669491768, + 0.02436576597392559, + 0.01469704881310463, + 0.011084700003266335, + 0.024170750752091408, + 0.0447591207921505, + 0.010801992379128933, + 0.008337810635566711, + 0.013475053943693638, + 0.007863455452024937, + 0.008722210302948952, + 0.007436427287757397, + 0.006936841644346714, + 0.006641216576099396, + 0.010234554298222065, + 0.007683971896767616, + 0.011260082945227623, + 0.007878591306507587, + 0.00402737595140934, + 0.0083165243268013, + 0.005072989501059055, + 0.006316997576504946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.00762729300186038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot glides smoothly alongside a cyclist, capturing her energetic movements along a classic Parisian cobblestone street, the camera keeping pace with her rapid progress. The cyclist is a young woman, her brown hair pulled back in a slightly messy bun, with a few strands escaping and dancing in the wind. Dressed in a timeless blue and white striped long-sleeved shirt and dark pants, she leans forward over the handlebars of her vintage bicycle, striking a focused pose. Her legs beat in a powerful, steady rhythm, her body swaying slightly with the effort, the wheel bouncing gently on the uneven, time-worn stones of the street. In the background, the majestic Eiffel Tower looms majestically, its iconic grid structure dominating the upper frame, slightly softened by the golden haze of a sunny afternoon. The bright, warm sunlight casts long, dancing shadows and creates subtle lens flares, bathing the entire scene in a vibrant golden light, evoking a joyful and exciting atmosphere of freedom.", + "seed": 14382, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_6/index_4822_seed_14382_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": -0.0783571746200323, + "vlm_logprob": -0.0783571746200323, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.0783571746200323, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20070014894008636 + ], + "opposite_logprob_tokens": [ + -2.700700044631958 + ], + "logprob_reward": -0.20070014894008636, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.20070014894008636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4006293714046478, + 0.1151193231344223, + 0.05841107666492462, + 0.08710961788892746, + 0.042281195521354675, + 0.07341521978378296, + 0.08285470306873322, + 0.07553227990865707, + 0.042505644261837006, + 0.022141529247164726 + ], + "credit_map": [ + 0.040699541568756104, + 0.08893577009439468, + 0.03717062249779701, + 0.04954591020941734, + 0.04689715430140495, + 0.042542941868305206, + 0.04637272655963898, + 0.04846469312906265, + 0.012674016878008842, + 0.011398633010685444, + 0.016030801460146904, + 0.016257166862487793, + 0.00798166822642088, + 0.010153031907975674, + 0.01715962588787079, + 0.023464377969503403, + 0.008581883274018764, + 0.006613467819988728, + 0.008600323460996151, + 0.007456749211996794, + 0.005976191256195307, + 0.005784470122307539, + 0.008329318836331367, + 0.0070686726830899715, + 0.013772801496088505, + 0.010344339534640312, + 0.027134142816066742, + 0.01027081348001957, + 0.005373755935579538, + 0.004308292642235756, + 0.007577979005873203, + 0.008327482268214226, + 0.0035603505093604326, + 0.007407199591398239, + 0.006410530302673578, + 0.0058950805105268955, + 0.005129095632582903, + 0.005122524686157703, + 0.0038181075360625982, + 0.004938303492963314, + 0.006682118400931358, + 0.010594304651021957, + 0.015475139953196049, + 0.009736568666994572, + 0.007014414295554161, + 0.005635532550513744, + 0.009167519398033619, + 0.009109620936214924, + 0.006363985128700733, + 0.008111203089356422, + 0.020835556089878082, + 0.015467096120119095, + 0.008288722485303879, + 0.005339060444384813, + 0.009610057808458805, + 0.008839021436870098, + 0.009558825753629208, + 0.008968456648290157, + 0.011215489357709885, + 0.009072907269001007, + 0.008271139115095139, + 0.007566293701529503, + 0.011240046471357346, + 0.009639116935431957, + 0.008977344259619713, + 0.00627258513122797, + 0.005347016733139753, + 0.004394212272018194, + 0.0044495719484984875, + 0.0048204162158071995, + 0.00485755130648613, + 0.0033869417384266853, + 0.004800823982805014, + 0.002884233370423317, + 0.0028644183184951544, + 0.002170031890273094, + 0.0019809401128441095, + 0.0022354002576321363, + 0.0019260604167357087, + 0.0032796200830489397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.20070014894008636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003812700510025 + ], + "opposite_logprob_tokens": [ + -2.350381374359131 + ], + "logprob_reward": -0.1003812700510025, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003812700510025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17527073621749878, + 0.22527986764907837, + 0.09495233744382858, + 0.14006635546684265, + 0.07282236963510513, + 0.04510354623198509, + 0.031615596264600754, + 0.12384338676929474, + 0.05851761996746063, + 0.03252812474966049 + ], + "credit_map": [ + 0.010793778114020824, + 0.026064883917570114, + 0.022158758714795113, + 0.03508903831243515, + 0.017076341435313225, + 0.017215969040989876, + 0.022927584126591682, + 0.023944390937685966, + 0.026827802881598473, + 0.020194344222545624, + 0.027514982968568802, + 0.02805369347333908, + 0.025669801980257034, + 0.020270192995667458, + 0.029167698696255684, + 0.04758135974407196, + 0.013053685426712036, + 0.009594805538654327, + 0.013756212778389454, + 0.012393105775117874, + 0.007962711155414581, + 0.013543453998863697, + 0.012805362232029438, + 0.011842998676002026, + 0.02945120446383953, + 0.009764906018972397, + 0.050255049020051956, + 0.015278041362762451, + 0.0068037500604987144, + 0.007750231772661209, + 0.008007238619029522, + 0.012755929492413998, + 0.005933684762567282, + 0.01543022133409977, + 0.012199567630887032, + 0.012441865168511868, + 0.010933993384242058, + 0.004075333010405302, + 0.006648792419582605, + 0.005158909130841494, + 0.004156406503170729, + 0.006224177777767181, + 0.00874748732894659, + 0.004906865768134594, + 0.0054477802477777, + 0.004623054061084986, + 0.00489761121571064, + 0.006100166589021683, + 0.0033086342737078667, + 0.003807042259722948, + 0.0037532015703618526, + 0.005090515594929457, + 0.0036128333304077387, + 0.0030061567667871714, + 0.005252595990896225, + 0.00378461554646492, + 0.01639634557068348, + 0.016592221334576607, + 0.017101941630244255, + 0.013714435510337353, + 0.0163930281996727, + 0.01271804142743349, + 0.019365821033716202, + 0.011561551131308079, + 0.010432152077555656, + 0.00916854664683342, + 0.007800286170095205, + 0.007603300269693136, + 0.007590859197080135, + 0.006480233743786812, + 0.005431173834949732, + 0.004011068493127823, + 0.0047228978946805, + 0.00579111510887742, + 0.002500776667147875, + 0.003674544394016266, + 0.004836574196815491, + 0.0037177661433815956, + 0.004754090216010809, + 0.0025303615257143974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.1003812700510025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606719382107258 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.01606719382107258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606719382107258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08961953967809677, + 0.1326737254858017, + 0.15437990427017212, + 0.1306666135787964, + 0.06708993017673492, + 0.05581791326403618, + 0.04317064955830574, + 0.1565895974636078, + 0.13636094331741333, + 0.033631131052970886 + ], + "credit_map": [ + 0.0057126726023852825, + 0.010404606349766254, + 0.010839341208338737, + 0.01571478694677353, + 0.007930557243525982, + 0.012141451239585876, + 0.013199986889958382, + 0.013676140457391739, + 0.013323287479579449, + 0.014783584512770176, + 0.015318149700760841, + 0.01692678965628147, + 0.013579212129116058, + 0.010540260933339596, + 0.01737205684185028, + 0.030830370262265205, + 0.022658348083496094, + 0.020904704928398132, + 0.02300497330725193, + 0.0178615041077137, + 0.01720292866230011, + 0.011797267012298107, + 0.020116610452532768, + 0.020833557471632957, + 0.028117995709180832, + 0.016231469810009003, + 0.03336726874113083, + 0.01603052392601967, + 0.008198211900889874, + 0.007836163975298405, + 0.01122359000146389, + 0.009661395102739334, + 0.005505668465048075, + 0.008877677842974663, + 0.012577474117279053, + 0.011071909219026566, + 0.008812865242362022, + 0.005409264471381903, + 0.007083456963300705, + 0.007751611527055502, + 0.0049672541208565235, + 0.008648259565234184, + 0.012512440793216228, + 0.007020954973995686, + 0.005303522106260061, + 0.003788650967180729, + 0.005823563784360886, + 0.007753266487270594, + 0.0038783736526966095, + 0.004475922789424658, + 0.0052507854998111725, + 0.006140757352113724, + 0.006497862748801708, + 0.0030103903263807297, + 0.008129733614623547, + 0.0057868254370987415, + 0.019049884751439095, + 0.018897825852036476, + 0.01910465583205223, + 0.017370175570249557, + 0.026716988533735275, + 0.018465187400579453, + 0.021273983642458916, + 0.015710901468992233, + 0.024182751774787903, + 0.020781749859452248, + 0.029472289606928825, + 0.011171089485287666, + 0.01062756683677435, + 0.015827981755137444, + 0.007614159025251865, + 0.016683358699083328, + 0.005789770279079676, + 0.006219979375600815, + 0.002632797695696354, + 0.003621983341872692, + 0.00562251964583993, + 0.0033801079262048006, + 0.003276588162407279, + 0.003087384393438697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.01606719382107258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000136375427246 + ], + "opposite_logprob_tokens": [ + -2.6950013637542725 + ], + "logprob_reward": -0.07000136375427246, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000136375427246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11379453539848328, + 0.047109730541706085, + 0.07383593171834946, + 0.18310751020908356, + 0.10760696977376938, + 0.02393016591668129, + 0.03175749257206917, + 0.04889100044965744, + 0.3562443256378174, + 0.013722273521125317 + ], + "credit_map": [ + 0.010157833807170391, + 0.028942445293068886, + 0.01051260344684124, + 0.013672670349478722, + 0.013652907684445381, + 0.010209480300545692, + 0.012036649510264397, + 0.014609944075345993, + 0.00662420317530632, + 0.005871559493243694, + 0.006291867699474096, + 0.006101248320192099, + 0.004610375966876745, + 0.006377713289111853, + 0.004159093368798494, + 0.007073672022670507, + 0.0102015919983387, + 0.00851505622267723, + 0.012057038024067879, + 0.00786347221583128, + 0.007831628434360027, + 0.007809899281710386, + 0.00960079487413168, + 0.009956453926861286, + 0.016142096370458603, + 0.027924852445721626, + 0.050354860723018646, + 0.029679903760552406, + 0.013450627215206623, + 0.011455989442765713, + 0.013950320892035961, + 0.020148877054452896, + 0.010276522487401962, + 0.028609678149223328, + 0.017667260020971298, + 0.014814922586083412, + 0.014471624977886677, + 0.006992317270487547, + 0.007318607997149229, + 0.007456042338162661, + 0.0025341350119560957, + 0.0041582523845136166, + 0.005366906989365816, + 0.0028291698545217514, + 0.0019835939165204763, + 0.001732144388370216, + 0.001896137953735888, + 0.0034298256505280733, + 0.003653776366263628, + 0.004427285864949226, + 0.00458345003426075, + 0.004500855691730976, + 0.0036668262910097837, + 0.0028907067608088255, + 0.0029300523456186056, + 0.00510454410687089, + 0.007245476823300123, + 0.007899136282503605, + 0.007634191773831844, + 0.005343918688595295, + 0.0050170887261629105, + 0.00506680179387331, + 0.005290175788104534, + 0.005394212901592255, + 0.06272950768470764, + 0.047232735902071, + 0.08526770025491714, + 0.03695908188819885, + 0.025441350415349007, + 0.043640416115522385, + 0.02285975217819214, + 0.03211382031440735, + 0.002741039963439107, + 0.0029376812744885683, + 0.0013941709185019135, + 0.001663503353483975, + 0.0015691743465140462, + 0.0015390198677778244, + 0.0008248768863268197, + 0.001052808715030551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.07000136375427246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.4573425352573395, + 0.0518476776778698, + 0.026080166921019554, + 0.1083734929561615, + 0.08107806742191315, + 0.039207201451063156, + 0.04370403662323952, + 0.021912695840001106, + 0.14446000754833221, + 0.025994084775447845 + ], + "credit_map": [ + 0.050050605088472366, + 0.03893865644931793, + 0.06630349904298782, + 0.0999377965927124, + 0.03450382873415947, + 0.06119599565863609, + 0.057221755385398865, + 0.0491904579102993, + 0.006791216786950827, + 0.006048437207937241, + 0.006176165770739317, + 0.0055180205963552, + 0.0046367645263671875, + 0.005146829877048731, + 0.0059956153854727745, + 0.011534629389643669, + 0.005338088143616915, + 0.0024599626194685698, + 0.0033974964171648026, + 0.0031182225793600082, + 0.0022116932086646557, + 0.0031974816229194403, + 0.0034705193247646093, + 0.0028867044020444155, + 0.023071415722370148, + 0.013916015625, + 0.02988947182893753, + 0.012541725300252438, + 0.005126579664647579, + 0.006350911688059568, + 0.009129436686635017, + 0.008347944356501102, + 0.007648695260286331, + 0.011889324523508549, + 0.018430717289447784, + 0.016317348927259445, + 0.008440792560577393, + 0.004604834597557783, + 0.0062344796024262905, + 0.00751187838613987, + 0.003570745699107647, + 0.006288031116127968, + 0.008823610842227936, + 0.004952682182192802, + 0.004586515948176384, + 0.0024380211252719164, + 0.0032572709023952484, + 0.005290324799716473, + 0.00371462176553905, + 0.00418801698833704, + 0.010681034065783024, + 0.009202024899423122, + 0.004184954799711704, + 0.0031115112360566854, + 0.004376514814794064, + 0.004245362710207701, + 0.0028960825875401497, + 0.0025614509359002113, + 0.0038658399134874344, + 0.0026312428526580334, + 0.0023712296970188618, + 0.00237843906506896, + 0.0036148668732494116, + 0.001593545195646584, + 0.0320294052362442, + 0.024030769243836403, + 0.018226290121674538, + 0.012335214763879776, + 0.012902374379336834, + 0.015255584381520748, + 0.011491118930280209, + 0.018189266324043274, + 0.004968723747879267, + 0.0031823047902435064, + 0.004822953138500452, + 0.00202065403573215, + 0.0021883209701627493, + 0.002727261045947671, + 0.0019607313442975283, + 0.004123138729482889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.004635896533727646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot glides smoothly alongside a cyclist, capturing her energetic movements along a classic Parisian cobblestone street, the camera keeping pace with her rapid progress. The cyclist is a young woman, her brown hair pulled back in a slightly messy bun, with a few strands escaping and dancing in the wind. Dressed in a timeless blue and white striped long-sleeved shirt and dark pants, she leans forward over the handlebars of her vintage bicycle, striking a focused pose. Her legs beat in a powerful, steady rhythm, her body swaying slightly with the effort, the wheel bouncing gently on the uneven, time-worn stones of the street. In the background, the majestic Eiffel Tower looms majestically, its iconic grid structure dominating the upper frame, slightly softened by the golden haze of a sunny afternoon. The bright, warm sunlight casts long, dancing shadows and creates subtle lens flares, bathing the entire scene in a vibrant golden light, evoking a joyful and exciting atmosphere of freedom.", + "seed": 14383, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_6/index_4822_seed_14383_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": -0.08808224871754647, + "vlm_logprob": -0.08808224871754647, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.08808224871754647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10773075371980667 + ], + "opposite_logprob_tokens": [ + -3.6077308654785156 + ], + "logprob_reward": -0.10773075371980667, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.10773075371980667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25616323947906494, + 0.12641911208629608, + 0.062216926366090775, + 0.11535593867301941, + 0.0841386541724205, + 0.03539475426077843, + 0.0946459174156189, + 0.096415676176548, + 0.07496998459100723, + 0.05427980422973633 + ], + "credit_map": [ + 0.023345407098531723, + 0.022512901574373245, + 0.047111205756664276, + 0.048337437212467194, + 0.021594632416963577, + 0.032069142907857895, + 0.029811672866344452, + 0.031380828469991684, + 0.011420131660997868, + 0.009227041155099869, + 0.022978506982326508, + 0.018974417820572853, + 0.0118815703317523, + 0.02226847968995571, + 0.013760779052972794, + 0.01590818352997303, + 0.006723865866661072, + 0.006822231225669384, + 0.012031483463943005, + 0.00844532623887062, + 0.007532334886491299, + 0.008022978901863098, + 0.007085863035172224, + 0.00555284321308136, + 0.011242141015827656, + 0.02829037792980671, + 0.01608526147902012, + 0.011009951122105122, + 0.013805918395519257, + 0.01505926065146923, + 0.01003112643957138, + 0.009831897914409637, + 0.006293271668255329, + 0.01036643236875534, + 0.024573329836130142, + 0.015413008630275726, + 0.008888914249837399, + 0.008664296939969063, + 0.0038923718966543674, + 0.006047029048204422, + 0.0040097651071846485, + 0.003884131321683526, + 0.004818479064851999, + 0.0035100742243230343, + 0.00636468967422843, + 0.0055955504067242146, + 0.003059557406231761, + 0.004152506589889526, + 0.008460473269224167, + 0.013520663604140282, + 0.017612015828490257, + 0.010238399729132652, + 0.011290748603641987, + 0.019173478707671165, + 0.00624039163812995, + 0.008109748363494873, + 0.015313384123146534, + 0.014786073006689548, + 0.013120205141603947, + 0.012253718450665474, + 0.012530547566711903, + 0.01343039982020855, + 0.006064494606107473, + 0.008916852995753288, + 0.007685605902224779, + 0.006004689261317253, + 0.020345965400338173, + 0.01500952523201704, + 0.00541986059397459, + 0.008361701853573322, + 0.0043134428560733795, + 0.007829193957149982, + 0.005555193405598402, + 0.006413432769477367, + 0.011879107914865017, + 0.0070503088645637035, + 0.005178052466362715, + 0.00526511250063777, + 0.006329735741019249, + 0.006608857773244381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.10773075371980667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22563211619853973 + ], + "opposite_logprob_tokens": [ + -1.6006320714950562 + ], + "logprob_reward": -0.22563211619853973, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22563211619853973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16269247233867645, + 0.08971551805734634, + 0.060048434883356094, + 0.08563614636659622, + 0.07181315869092941, + 0.07416772842407227, + 0.19773149490356445, + 0.10699144750833511, + 0.07382997870445251, + 0.07737360894680023 + ], + "credit_map": [ + 0.01751871034502983, + 0.018915977329015732, + 0.02337314374744892, + 0.024648992344737053, + 0.01739463023841381, + 0.01712816208600998, + 0.023324202746152878, + 0.02038864977657795, + 0.011046758852899075, + 0.00699130492284894, + 0.013177063316106796, + 0.016665225848555565, + 0.009178715758025646, + 0.013629693537950516, + 0.008180541917681694, + 0.010846221819519997, + 0.006530593615025282, + 0.005771775264292955, + 0.013559037819504738, + 0.008306801319122314, + 0.0072037456557154655, + 0.007329854648560286, + 0.005919034127146006, + 0.0054275961592793465, + 0.010254022665321827, + 0.02204553782939911, + 0.011865610256791115, + 0.008494109846651554, + 0.008689841255545616, + 0.008850195445120335, + 0.007791017647832632, + 0.007645812351256609, + 0.008958922699093819, + 0.011989014223217964, + 0.013264627195894718, + 0.009120994247496128, + 0.0070251645520329475, + 0.006881347391754389, + 0.004895100835710764, + 0.009677986614406109, + 0.010054589249193668, + 0.012388224713504314, + 0.01576697640120983, + 0.008730974048376083, + 0.007078232709318399, + 0.010579202324151993, + 0.003434048732742667, + 0.006135475356131792, + 0.016827821731567383, + 0.01699233055114746, + 0.0509750097990036, + 0.01980827935039997, + 0.015745345503091812, + 0.05672389641404152, + 0.008815338835120201, + 0.011843469925224781, + 0.013029338791966438, + 0.017433276399970055, + 0.019412873312830925, + 0.012868938967585564, + 0.013438792899250984, + 0.014471407048404217, + 0.006578837055712938, + 0.009757984429597855, + 0.009169896133244038, + 0.007588466163724661, + 0.01201589684933424, + 0.013863404281437397, + 0.0068209003657102585, + 0.007611533161252737, + 0.006157893221825361, + 0.010601981543004513, + 0.00515742227435112, + 0.007051380351185799, + 0.018992789089679718, + 0.010186114348471165, + 0.006149795837700367, + 0.007460560649633408, + 0.008869512006640434, + 0.013506029732525349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.22563211619853973 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358967646956444 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026358967646956444, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358967646956444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07041341066360474, + 0.05846092477440834, + 0.2909710705280304, + 0.09783462435007095, + 0.057709481567144394, + 0.08456581085920334, + 0.19142621755599976, + 0.06580829620361328, + 0.028659136965870857, + 0.054150983691215515 + ], + "credit_map": [ + 0.008553940802812576, + 0.005440648645162582, + 0.01509874127805233, + 0.011903318576514721, + 0.005382022820413113, + 0.00855608657002449, + 0.0065142689272761345, + 0.008964378386735916, + 0.005043873563408852, + 0.004721738398075104, + 0.009195085614919662, + 0.009843661449849606, + 0.005694652907550335, + 0.0098571153357625, + 0.006200010888278484, + 0.007904785685241222, + 0.024233857169747353, + 0.020199524238705635, + 0.06385887414216995, + 0.03781842812895775, + 0.039471667259931564, + 0.04155280813574791, + 0.032794371247291565, + 0.03104156069457531, + 0.006653987802565098, + 0.012735887430608273, + 0.0159466415643692, + 0.009928788058459759, + 0.012543260119855404, + 0.01874164491891861, + 0.010726477950811386, + 0.01055793184787035, + 0.00591845391318202, + 0.005594263318926096, + 0.007501883897930384, + 0.004283742979168892, + 0.01409873180091381, + 0.008505016565322876, + 0.005520706530660391, + 0.006286683492362499, + 0.010617188178002834, + 0.013329890556633472, + 0.011750641278922558, + 0.008462941274046898, + 0.014731502160429955, + 0.01334542315453291, + 0.005117841064929962, + 0.007210386451333761, + 0.00847633183002472, + 0.01468160841614008, + 0.04046451672911644, + 0.02491586096584797, + 0.018879089504480362, + 0.06258603185415268, + 0.009150072000920773, + 0.012272707186639309, + 0.011944741010665894, + 0.00614787545055151, + 0.010023614391684532, + 0.006971498019993305, + 0.011650476604700089, + 0.007872398942708969, + 0.004477368667721748, + 0.006720319390296936, + 0.0020972101483494043, + 0.0023469440639019012, + 0.0060538905672729015, + 0.005707297474145889, + 0.0046943845227360725, + 0.003496757708489895, + 0.0015812072670087218, + 0.00268144765868783, + 0.004640588071197271, + 0.004818798042833805, + 0.012022706679999828, + 0.006927705369889736, + 0.006761931348592043, + 0.006070280447602272, + 0.0048184082843363285, + 0.008090564049780369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.026358967646956444 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897545397281647 + ], + "opposite_logprob_tokens": [ + -2.5789754390716553 + ], + "logprob_reward": -0.07897545397281647, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897545397281647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08704831451177597, + 0.14441744983196259, + 0.043287649750709534, + 0.04504245147109032, + 0.15331415832042694, + 0.05760601535439491, + 0.21779274940490723, + 0.10870475322008133, + 0.022980837151408195, + 0.11980562657117844 + ], + "credit_map": [ + 0.00835050456225872, + 0.009399465285241604, + 0.012409068644046783, + 0.014224141836166382, + 0.007549361791461706, + 0.009205535054206848, + 0.013818490318953991, + 0.01209175679832697, + 0.011749010533094406, + 0.010711860843002796, + 0.022327899932861328, + 0.01822766289114952, + 0.014053473249077797, + 0.027987390756607056, + 0.018325410783290863, + 0.021034736186265945, + 0.005961801391094923, + 0.005389200057834387, + 0.009021596051752567, + 0.007960098795592785, + 0.0041768504306674, + 0.003868602216243744, + 0.0030878346879035234, + 0.0038216637913137674, + 0.0045623695477843285, + 0.009267338551580906, + 0.006365260574966669, + 0.005635865498334169, + 0.004837351851165295, + 0.005992515943944454, + 0.003966276999562979, + 0.004415476229041815, + 0.024078404530882835, + 0.026731979101896286, + 0.02961072325706482, + 0.022472351789474487, + 0.011451116763055325, + 0.017495835199952126, + 0.009141271933913231, + 0.012332472950220108, + 0.006620045751333237, + 0.012110786512494087, + 0.009853176772594452, + 0.006770612206310034, + 0.0047348877415061, + 0.011270145885646343, + 0.0024215986486524343, + 0.003824761835858226, + 0.013264765962958336, + 0.02512974664568901, + 0.04685166850686073, + 0.029973838478326797, + 0.011491741053760052, + 0.07020223885774612, + 0.009281005710363388, + 0.01159773301333189, + 0.016693979501724243, + 0.013862155377864838, + 0.020471548661589622, + 0.01643495075404644, + 0.01280215848237276, + 0.01370253600180149, + 0.005602406803518534, + 0.009135018102824688, + 0.0038140364922583103, + 0.003620115341618657, + 0.003227743785828352, + 0.0032368008978664875, + 0.001864609308540821, + 0.0026055884081870317, + 0.0013721500290557742, + 0.0032397923059761524, + 0.014866504818201065, + 0.012574578635394573, + 0.028109990060329437, + 0.015371158719062805, + 0.009157399646937847, + 0.015935324132442474, + 0.007644158788025379, + 0.01614651083946228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.07897545397281647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyclist a young woman with brown hair in a messy bun wearing a blue and white striped long-sleeved shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a classic Parisian cobblestone street with the Eiffel Tower visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright warm sunlight creating long shadows and a golden haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001713952049612999 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.001713952049612999, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001713952049612999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2120530605316162, + 0.06649943441152573, + 0.030844086781144142, + 0.044687192887067795, + 0.19241946935653687, + 0.02891835756599903, + 0.09557855129241943, + 0.23288531601428986, + 0.03993910178542137, + 0.056175388395786285 + ], + "credit_map": [ + 0.018678033724427223, + 0.020962487906217575, + 0.038484275341033936, + 0.047781918197870255, + 0.01611587218940258, + 0.013048357330262661, + 0.03510812669992447, + 0.021873988211154938, + 0.008053746074438095, + 0.006209664046764374, + 0.009750103577971458, + 0.00959098245948553, + 0.00575080793350935, + 0.010443402454257011, + 0.008303769864141941, + 0.008396957069635391, + 0.0040727113373577595, + 0.004153629299253225, + 0.004286252427846193, + 0.0030672892462462187, + 0.00412572268396616, + 0.005076968111097813, + 0.0027769755106419325, + 0.0032845374662429094, + 0.005139401648193598, + 0.007612213026732206, + 0.006717788986861706, + 0.004709216300398111, + 0.005388004705309868, + 0.0065380847081542015, + 0.0039700912311673164, + 0.004612390883266926, + 0.022507444024086, + 0.02932882495224476, + 0.05021596699953079, + 0.03646358847618103, + 0.012806719169020653, + 0.019446080550551414, + 0.009334503673017025, + 0.012316357344388962, + 0.005581938195973635, + 0.004117867443710566, + 0.0038500449154525995, + 0.003486324567347765, + 0.0028944092337042093, + 0.004720257129520178, + 0.0016853832639753819, + 0.00258213235065341, + 0.009879518300294876, + 0.016724565997719765, + 0.015075813978910446, + 0.012510915286839008, + 0.008853388950228691, + 0.02193472906947136, + 0.004425393417477608, + 0.0061742281541228294, + 0.0314784049987793, + 0.04954766482114792, + 0.034291766583919525, + 0.02739293873310089, + 0.020919837057590485, + 0.03581152856349945, + 0.014018080197274685, + 0.019425103440880775, + 0.005144872236996889, + 0.004161692690104246, + 0.0067258006893098354, + 0.008996030315756798, + 0.0028484135400503874, + 0.004304102156311274, + 0.0025670151226222515, + 0.005191177129745483, + 0.005618325434625149, + 0.006039460655301809, + 0.00793592631816864, + 0.007987730205059052, + 0.00797083880752325, + 0.006961210630834103, + 0.004064394626766443, + 0.009597500786185265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.001713952049612999 + } + ], + "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/0000128/group_7/index_tensor([9709, 9709])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_tensor([9709, 9709])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..79df34c36a5febea9537364a03f1537f9cd8b299 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_tensor([9709, 9709])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion close-up shot, shot from a slightly upward angle and employing an extremely shallow depth of field, precisely captures an exquisite antique watch being gently tossed into the air. The watch features a polished rose gold case, an ivory dial adorned with slender Roman numerals and delicate blued-steel Breguet hands, and a rich, dark brown alligator leather strap. As the watch gracefully and slowly turns freely in the air, its domed sapphire crystal and metal case capture and reflect the warm surrounding light, shimmering fleeting, dazzling bursts of light with each rotation. In the background, the figures of friends laughing and chatting are blurred into soft, dynamic silhouettes, while the warm interior lighting transforms into captivating circular circles of light, creating an atmosphere of intimacy, joy, and the essence of life, capturing the beauty of this moment in eternal joy.", + "seed": 25087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_9709_seed_25087_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": -0.018278746865689754, + "vlm_logprob": -0.018278746865689754, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.018278746865689754, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008273139595985413 + ], + "opposite_logprob_tokens": [ + -5.383273124694824 + ], + "logprob_reward": -0.008273139595985413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.008273139595985413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13337105512619019, + 0.07830598205327988, + 0.07856243848800659, + 0.13390718400478363, + 0.09411569684743881, + 0.07845890522003174, + 0.10360460728406906, + 0.049394942820072174, + 0.05731411278247833, + 0.19296503067016602 + ], + "credit_map": [ + 0.012156604789197445, + 0.011463978327810764, + 0.01882481575012207, + 0.014804992824792862, + 0.010843336582183838, + 0.0162524227052927, + 0.01220627874135971, + 0.03681863471865654, + 0.005665095057338476, + 0.007447005249559879, + 0.011393708176910877, + 0.007394369225949049, + 0.007169422693550587, + 0.023481886833906174, + 0.009027646854519844, + 0.00672684982419014, + 0.01050336193293333, + 0.012843282893300056, + 0.006072469986975193, + 0.014065824449062347, + 0.004521598573774099, + 0.011944886296987534, + 0.008984854444861412, + 0.00962616316974163, + 0.01641167886555195, + 0.016058389097452164, + 0.017374277114868164, + 0.011388340964913368, + 0.018654385581612587, + 0.01759500801563263, + 0.022127507254481316, + 0.014297602698206902, + 0.012522066943347454, + 0.011824163608253002, + 0.009147394448518753, + 0.011806623078882694, + 0.015671193599700928, + 0.01071404293179512, + 0.01307247020304203, + 0.009357747621834278, + 0.010571314953267574, + 0.007188515737652779, + 0.009232538752257824, + 0.009686041623353958, + 0.008008410222828388, + 0.007096403744071722, + 0.019377613440155983, + 0.007298073265701532, + 0.01065613329410553, + 0.008736792951822281, + 0.012451895512640476, + 0.019909165799617767, + 0.009417415596544743, + 0.008527289144694805, + 0.022921407595276833, + 0.010984514839947224, + 0.00624444056302309, + 0.006493505090475082, + 0.008881988003849983, + 0.004945294000208378, + 0.005820435471832752, + 0.005094318185001612, + 0.004513389430940151, + 0.007401579059660435, + 0.006890461314469576, + 0.007306769955903292, + 0.008648816496133804, + 0.011695926077663898, + 0.00478646345436573, + 0.00638474291190505, + 0.006797766778618097, + 0.004803170450031757, + 0.03800032287836075, + 0.02764810249209404, + 0.06709189713001251, + 0.012123598717153072, + 0.011053348891437054, + 0.012913533486425877, + 0.013967497274279594, + 0.010166737250983715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.008273139595985413 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209873043000698 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209873043000698, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209873043000698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10967544466257095, + 0.08872276544570923, + 0.0462084636092186, + 0.13629095256328583, + 0.041574496775865555, + 0.10243304818868637, + 0.07724777609109879, + 0.069783516228199, + 0.03360528126358986, + 0.2944582402706146 + ], + "credit_map": [ + 0.008228362537920475, + 0.007593842223286629, + 0.01611446402966976, + 0.011602886021137238, + 0.010348482057452202, + 0.01449313759803772, + 0.010258465074002743, + 0.031035801395773888, + 0.012010855600237846, + 0.013429185375571251, + 0.008920539170503616, + 0.01510629616677761, + 0.0055931331589818, + 0.008520075120031834, + 0.014351650141179562, + 0.010791031643748283, + 0.005214954260736704, + 0.00664616608992219, + 0.003938114736229181, + 0.00480742147192359, + 0.0024863339494913816, + 0.012507312931120396, + 0.004811776801943779, + 0.005796381272375584, + 0.020513560622930527, + 0.008939016610383987, + 0.0388796404004097, + 0.01125436369329691, + 0.014334643259644508, + 0.014815800823271275, + 0.012362579815089703, + 0.01519134920090437, + 0.005862120538949966, + 0.003988619893789291, + 0.005183848086744547, + 0.004252231679856777, + 0.0051003131084144115, + 0.006187953986227512, + 0.005514703691005707, + 0.005484702996909618, + 0.011651060543954372, + 0.009555868804454803, + 0.011775223538279533, + 0.014200171455740929, + 0.011979633942246437, + 0.010387827642261982, + 0.021252736449241638, + 0.01163051463663578, + 0.008643434382975101, + 0.008342577144503593, + 0.01051684096455574, + 0.017269428819417953, + 0.004985131788998842, + 0.006691256072372198, + 0.012916624546051025, + 0.007882476784288883, + 0.00664796307682991, + 0.008598476648330688, + 0.010342190973460674, + 0.011178772896528244, + 0.005789384711533785, + 0.006153016816824675, + 0.009676665998995304, + 0.011397041380405426, + 0.0042688618414103985, + 0.004099286627024412, + 0.005091659724712372, + 0.005244697909802198, + 0.0030714578460901976, + 0.004448742140084505, + 0.0031558361370116472, + 0.004224738571792841, + 0.0467454232275486, + 0.018830953165888786, + 0.10433176159858704, + 0.017058243975043297, + 0.014271890744566917, + 0.06262798607349396, + 0.016253970563411713, + 0.014338008128106594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.014209873043000698 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589547231793404 + ], + "opposite_logprob_tokens": [ + -3.895589590072632 + ], + "logprob_reward": -0.020589547231793404, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589547231793404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17806479334831238, + 0.07831715047359467, + 0.11113817244768143, + 0.13580287992954254, + 0.055224109441041946, + 0.07807638496160507, + 0.14870023727416992, + 0.03425457701086998, + 0.043630585074424744, + 0.13679106533527374 + ], + "credit_map": [ + 0.014462104998528957, + 0.01844974234700203, + 0.024160511791706085, + 0.01918352022767067, + 0.014679458923637867, + 0.02593257650732994, + 0.015116417780518532, + 0.046080462634563446, + 0.005621709860861301, + 0.007647059857845306, + 0.009166533127427101, + 0.00934841763228178, + 0.006692848633974791, + 0.014705464243888855, + 0.01322257798165083, + 0.011912534944713116, + 0.02549624629318714, + 0.017675306648015976, + 0.009059419855475426, + 0.01469150185585022, + 0.006482328753918409, + 0.012316934764385223, + 0.012264804914593697, + 0.013151627033948898, + 0.019420092925429344, + 0.014000860042870045, + 0.035631123930215836, + 0.01451870333403349, + 0.014189209789037704, + 0.012139436788856983, + 0.012874497100710869, + 0.013028965331614017, + 0.007802142761647701, + 0.0056215254589915276, + 0.006771840620785952, + 0.0067048026248812675, + 0.007100180257111788, + 0.006776961963623762, + 0.006269178818911314, + 0.00817747414112091, + 0.009611615911126137, + 0.011927708052098751, + 0.008759303018450737, + 0.011280438862740993, + 0.011859267018735409, + 0.006218424066901207, + 0.011758733540773392, + 0.006660899147391319, + 0.016875486820936203, + 0.014422942884266376, + 0.020094413310289383, + 0.03007274493575096, + 0.011713051237165928, + 0.012788081541657448, + 0.02894500270485878, + 0.013788511045277119, + 0.004114638548344374, + 0.005171966273337603, + 0.0045867920853197575, + 0.004472643136978149, + 0.0032670460641384125, + 0.003277123672887683, + 0.003842266509309411, + 0.005522097926586866, + 0.004864523187279701, + 0.006279489025473595, + 0.005109380930662155, + 0.007198680657893419, + 0.003287214320152998, + 0.004944521468132734, + 0.004902516026049852, + 0.007044260390102863, + 0.016081225126981735, + 0.01618151366710663, + 0.0272192545235157, + 0.017604315653443336, + 0.012828630395233631, + 0.012765870429575443, + 0.02303723245859146, + 0.011073029600083828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.020589547231793404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306785389780998 + ], + "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": [ + 0.09949057549238205, + 0.03438953310251236, + 0.16609147191047668, + 0.15716658532619476, + 0.04927889630198479, + 0.3448165953159332, + 0.025459084659814835, + 0.07101687043905258, + 0.0337851382791996, + 0.01850522682070732 + ], + "credit_map": [ + 0.00732376566156745, + 0.008929774165153503, + 0.01229364424943924, + 0.017138471826910973, + 0.006955749820917845, + 0.012867063283920288, + 0.008040986023843288, + 0.025941114872694016, + 0.0031498167663812637, + 0.003462577937170863, + 0.004992371890693903, + 0.004884128924459219, + 0.002030099043622613, + 0.007493501529097557, + 0.004697505384683609, + 0.0036795325577259064, + 0.02076585404574871, + 0.017643745988607407, + 0.01655380241572857, + 0.016356784850358963, + 0.011193727143108845, + 0.04470393806695938, + 0.017428157851099968, + 0.02144547738134861, + 0.018332481384277344, + 0.01622849516570568, + 0.02498515322804451, + 0.019175156950950623, + 0.020698262378573418, + 0.018627341836690903, + 0.02397436462342739, + 0.015145326964557171, + 0.0064042191952466965, + 0.006166120059788227, + 0.006478165742009878, + 0.006804360076785088, + 0.007498112041503191, + 0.005179801024496555, + 0.007167561911046505, + 0.0035805560182780027, + 0.03887706995010376, + 0.050196584314107895, + 0.03757384046912193, + 0.04606148228049278, + 0.04312857612967491, + 0.038057051599025726, + 0.04698719084262848, + 0.04393480345606804, + 0.0032585863955318928, + 0.00219933595508337, + 0.003881514072418213, + 0.004664351232349873, + 0.0023294296115636826, + 0.0023659146390855312, + 0.004799804650247097, + 0.001960149733349681, + 0.008655036799609661, + 0.011691205203533173, + 0.009426509030163288, + 0.0060895648784935474, + 0.004797023255378008, + 0.011419866234064102, + 0.006468812469393015, + 0.012468847446143627, + 0.003386199474334717, + 0.0034756180830299854, + 0.004441528581082821, + 0.00456243148073554, + 0.003363092662766576, + 0.003954882267862558, + 0.0060035758651793, + 0.004597811494022608, + 0.0020143426954746246, + 0.0021742188837379217, + 0.002616152400150895, + 0.0021024232264608145, + 0.0021776703651994467, + 0.0024176754523068666, + 0.00331155676394701, + 0.0016911870334297419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.04306785389780998 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253320559859276 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253320559859276, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253320559859276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.121302030980587, + 0.04400673881173134, + 0.09892268478870392, + 0.22195148468017578, + 0.08787067979574203, + 0.0804186463356018, + 0.11655856668949127, + 0.05852004140615463, + 0.07083159685134888, + 0.09961753338575363 + ], + "credit_map": [ + 0.009005597792565823, + 0.017743628472089767, + 0.013688593171536922, + 0.022073442116379738, + 0.011895122937858105, + 0.016092991456389427, + 0.01424347423017025, + 0.016559185460209846, + 0.0034200926311314106, + 0.0030108727514743805, + 0.00637826370075345, + 0.0034743091091513634, + 0.003921398427337408, + 0.01313007716089487, + 0.005585890728980303, + 0.005085833836346865, + 0.01433755736798048, + 0.01678479090332985, + 0.008528354577720165, + 0.010308613069355488, + 0.007019799668341875, + 0.019070563837885857, + 0.009896841831505299, + 0.012976161204278469, + 0.02641659788787365, + 0.01782808266580105, + 0.05837061628699303, + 0.021392373368144035, + 0.02272975631058216, + 0.02261042222380638, + 0.030947471037507057, + 0.02165617235004902, + 0.012078017927706242, + 0.00818453449755907, + 0.012438494712114334, + 0.009159077890217304, + 0.010932506993412971, + 0.013246235437691212, + 0.009351968765258789, + 0.012479838915169239, + 0.009165404364466667, + 0.011496009305119514, + 0.008951895870268345, + 0.010645303875207901, + 0.010315481573343277, + 0.006294182501733303, + 0.014284014701843262, + 0.009266352280974388, + 0.013868882320821285, + 0.012054414488375187, + 0.014558770693838596, + 0.021868448704481125, + 0.011120794340968132, + 0.009340370073914528, + 0.022240055724978447, + 0.011506826616823673, + 0.010132530704140663, + 0.010080770589411259, + 0.006884477101266384, + 0.006901668384671211, + 0.004013963043689728, + 0.005976369604468346, + 0.0059505607932806015, + 0.008579700253903866, + 0.005769453942775726, + 0.009711368940770626, + 0.008044721558690071, + 0.012153743766248226, + 0.007169029209762812, + 0.010923407971858978, + 0.010420657694339752, + 0.006639210507273674, + 0.016992051154375076, + 0.007098548114299774, + 0.03562465310096741, + 0.0063570719212293625, + 0.004939710721373558, + 0.01774846576154232, + 0.006732551846653223, + 0.0041244844906032085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.005253320559859276 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up shot, shot from a slightly upward angle and employing an extremely shallow depth of field, precisely captures an exquisite antique watch being gently tossed into the air. The watch features a polished rose gold case, an ivory dial adorned with slender Roman numerals and delicate blued-steel Breguet hands, and a rich, dark brown alligator leather strap. As the watch gracefully and slowly turns freely in the air, its domed sapphire crystal and metal case capture and reflect the warm surrounding light, shimmering fleeting, dazzling bursts of light with each rotation. In the background, the figures of friends laughing and chatting are blurred into soft, dynamic silhouettes, while the warm interior lighting transforms into captivating circular circles of light, creating an atmosphere of intimacy, joy, and the essence of life, capturing the beauty of this moment in eternal joy.", + "seed": 25088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_9709_seed_25088_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": -0.027448378456756472, + "vlm_logprob": -0.027448378456756472, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.027448378456756472, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009155198000371456 + ], + "opposite_logprob_tokens": [ + -5.1341552734375 + ], + "logprob_reward": -0.009155198000371456, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009155198000371456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1570701003074646, + 0.03300047293305397, + 0.0761728286743164, + 0.1908939927816391, + 0.0752454474568367, + 0.10955547541379929, + 0.09507597982883453, + 0.047880206257104874, + 0.10864624381065369, + 0.10645929723978043 + ], + "credit_map": [ + 0.015654239803552628, + 0.02467353641986847, + 0.018637800589203835, + 0.017041336745023727, + 0.01566990092396736, + 0.018628904595971107, + 0.022701341658830643, + 0.024063032120466232, + 0.002779944334179163, + 0.003077559405937791, + 0.004812287166714668, + 0.003592673921957612, + 0.004401788115501404, + 0.0042778742499649525, + 0.003845733357593417, + 0.006212610751390457, + 0.008505074307322502, + 0.009939953684806824, + 0.010156659409403801, + 0.009455167688429356, + 0.013504886999726295, + 0.008524834178388119, + 0.009223958477377892, + 0.006862292066216469, + 0.020597483962774277, + 0.012608829885721207, + 0.027534469962120056, + 0.019277919083833694, + 0.010885221883654594, + 0.04414073005318642, + 0.014973162673413754, + 0.04087618738412857, + 0.013154670596122742, + 0.005123791750520468, + 0.014004097320139408, + 0.005603218916803598, + 0.0041154916398227215, + 0.01470168586820364, + 0.010260910727083683, + 0.008281577378511429, + 0.01230813842266798, + 0.005107174161821604, + 0.00676005519926548, + 0.006314344238489866, + 0.005636816378682852, + 0.033603716641664505, + 0.006192141678184271, + 0.03363308683037758, + 0.01214966457337141, + 0.008350018411874771, + 0.013988811522722244, + 0.007856958545744419, + 0.013775584287941456, + 0.014677532948553562, + 0.01617470011115074, + 0.008102701976895332, + 0.005208943039178848, + 0.00468413857743144, + 0.005284890532493591, + 0.006125719286501408, + 0.006249105092138052, + 0.0062688724137842655, + 0.0059596518985927105, + 0.008098884485661983, + 0.017335955053567886, + 0.01018841378390789, + 0.020073166117072105, + 0.012967121787369251, + 0.011402461677789688, + 0.012944280169904232, + 0.01145878341048956, + 0.012276064604520798, + 0.007589440327137709, + 0.011548769660294056, + 0.015311595983803272, + 0.011477884836494923, + 0.016960741952061653, + 0.016466191038489342, + 0.01010556984692812, + 0.016999099403619766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.009155198000371456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255148183554411 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.01255148183554411, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255148183554411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08713974803686142, + 0.031120698899030685, + 0.09846878051757812, + 0.1857212334871292, + 0.09895563870668411, + 0.16223756968975067, + 0.04257356375455856, + 0.050099849700927734, + 0.19030195474624634, + 0.05338101461529732 + ], + "credit_map": [ + 0.008481213822960854, + 0.013618430122733116, + 0.011226513423025608, + 0.00956166535615921, + 0.009026477113366127, + 0.009304650127887726, + 0.012265772558748722, + 0.013655029237270355, + 0.0038090210873633623, + 0.003945694770663977, + 0.00474102608859539, + 0.003306163242086768, + 0.0037955152802169323, + 0.002870735013857484, + 0.004006340634077787, + 0.004646202549338341, + 0.014293143525719643, + 0.01424955204129219, + 0.01290693785995245, + 0.009083648212254047, + 0.013354546390473843, + 0.011568759568035603, + 0.01073491107672453, + 0.012277282774448395, + 0.02016451768577099, + 0.010837540030479431, + 0.03129952773451805, + 0.017396418377757072, + 0.009141026996076107, + 0.0412876233458519, + 0.015141800045967102, + 0.04045278578996658, + 0.015615733340382576, + 0.006626871880143881, + 0.02183166705071926, + 0.005955095402896404, + 0.005140399094671011, + 0.025342633947730064, + 0.006216107867658138, + 0.012227126397192478, + 0.03586699813604355, + 0.007913707755506039, + 0.006247719284147024, + 0.008701636455953121, + 0.010100701823830605, + 0.058979831635951996, + 0.007554986514151096, + 0.026871992275118828, + 0.004447032231837511, + 0.0035652685910463333, + 0.006678253877907991, + 0.0036876664962619543, + 0.00609248224645853, + 0.005672180559486151, + 0.008133758790791035, + 0.004296920727938414, + 0.009182783775031567, + 0.00517839752137661, + 0.009632185101509094, + 0.006378910504281521, + 0.005280285142362118, + 0.004195057787001133, + 0.004553901497274637, + 0.005698330234736204, + 0.014816970564424992, + 0.01842113398015499, + 0.019754109904170036, + 0.021997777745127678, + 0.025482473894953728, + 0.024241557344794273, + 0.02237824909389019, + 0.043209683150053024, + 0.005665796343237162, + 0.005524938926100731, + 0.0070830341428518295, + 0.006969434209167957, + 0.005803078878670931, + 0.008471188135445118, + 0.006034554913640022, + 0.007828989066183567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.01255148183554411 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329564467072487 + ], + "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": [ + 0.13907967507839203, + 0.08156333863735199, + 0.040587764233350754, + 0.2513314187526703, + 0.059012364596128464, + 0.09574389457702637, + 0.09543219208717346, + 0.09360450506210327, + 0.06744179129600525, + 0.07620308548212051 + ], + "credit_map": [ + 0.016817791387438774, + 0.023213408887386322, + 0.016804758459329605, + 0.015727965161204338, + 0.01713951863348484, + 0.016274468973279, + 0.01706383377313614, + 0.016037924215197563, + 0.007022767793387175, + 0.006857695989310741, + 0.011132140643894672, + 0.012445841915905476, + 0.015167892910540104, + 0.009059760719537735, + 0.010750905610620975, + 0.009126324206590652, + 0.005537906661629677, + 0.005863273981958628, + 0.004773546475917101, + 0.005071292165666819, + 0.005246944725513458, + 0.004957609809935093, + 0.005331107880920172, + 0.0038060788065195084, + 0.02709193527698517, + 0.018270641565322876, + 0.03416099399328232, + 0.028356630355119705, + 0.016412587836384773, + 0.0545201301574707, + 0.022645404562354088, + 0.049873076379299164, + 0.006715388037264347, + 0.005864548496901989, + 0.008599347434937954, + 0.004919243510812521, + 0.006038538180291653, + 0.006760706659406424, + 0.01182869728654623, + 0.008285892195999622, + 0.02143518254160881, + 0.005144280381500721, + 0.006493036635220051, + 0.006928042508661747, + 0.005423130467534065, + 0.03144542872905731, + 0.00429695937782526, + 0.014577826485037804, + 0.015116628259420395, + 0.00971303042024374, + 0.02611490711569786, + 0.009112298488616943, + 0.010813819244503975, + 0.007874942384660244, + 0.00880031380802393, + 0.007886244915425777, + 0.015261140652000904, + 0.010029108263552189, + 0.014909952878952026, + 0.007132034283131361, + 0.01030633132904768, + 0.010471134446561337, + 0.009265843778848648, + 0.01622895523905754, + 0.008905604481697083, + 0.0074114068411290646, + 0.012019769288599491, + 0.007531848736107349, + 0.009548285976052284, + 0.00812279898673296, + 0.005413191393017769, + 0.008488881401717663, + 0.006021176464855671, + 0.006473875138908625, + 0.0096724983304739, + 0.009289880283176899, + 0.008803507313132286, + 0.01040769461542368, + 0.012496454641222954, + 0.013037985190749168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "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 shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904688060283661 + ], + "opposite_logprob_tokens": [ + -2.4640469551086426 + ], + "logprob_reward": -0.08904688060283661, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904688060283661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05387963727116585, + 0.04036164656281471, + 0.052514009177684784, + 0.3221026062965393, + 0.242609903216362, + 0.1116449162364006, + 0.06374509632587433, + 0.021475231274962425, + 0.06673115491867065, + 0.024935785681009293 + ], + "credit_map": [ + 0.005385907832533121, + 0.008864963427186012, + 0.0078002712689340115, + 0.007122809998691082, + 0.007307907100766897, + 0.0052941241301596165, + 0.0056259180419147015, + 0.00647773826494813, + 0.004016028251498938, + 0.004884897265583277, + 0.006286409683525562, + 0.0055821058340370655, + 0.0052042207680642605, + 0.004219162743538618, + 0.004739386960864067, + 0.005429434590041637, + 0.005092105362564325, + 0.005445367190986872, + 0.0064811441116034985, + 0.0059353504329919815, + 0.009486082009971142, + 0.005159766413271427, + 0.008413348346948624, + 0.00650084437802434, + 0.03483336418867111, + 0.021697303280234337, + 0.04869143292307854, + 0.030984237790107727, + 0.01855555549263954, + 0.07117876410484314, + 0.028258949518203735, + 0.06790299713611603, + 0.038267020136117935, + 0.013295403681695461, + 0.05608721077442169, + 0.014476620592176914, + 0.010833102278411388, + 0.06649695336818695, + 0.013378782197833061, + 0.029774796217679977, + 0.010987880639731884, + 0.005328758154064417, + 0.0073689064010977745, + 0.0061678276397287846, + 0.005008751060813665, + 0.04013875871896744, + 0.00650163646787405, + 0.030142389237880707, + 0.011389623396098614, + 0.006352867931127548, + 0.019666342064738274, + 0.007165607530623674, + 0.004391251131892204, + 0.0046699331142008305, + 0.004527844954282045, + 0.0055816262029111385, + 0.002763921394944191, + 0.0022909706458449364, + 0.0030843301210552454, + 0.003697217209264636, + 0.0028675503563135862, + 0.0022779430728405714, + 0.0021497858688235283, + 0.0023435107432305813, + 0.00903348345309496, + 0.006650905590504408, + 0.01219523698091507, + 0.0076814391650259495, + 0.0072171650826931, + 0.007920431904494762, + 0.007054863963276148, + 0.00897762831300497, + 0.0019386730855330825, + 0.0025464019272476435, + 0.0037863939069211483, + 0.003165049012750387, + 0.0030800264794379473, + 0.003685146104544401, + 0.0028516005259007215, + 0.0038824959192425013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.08904688060283661 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12485427409410477, + 0.03452502191066742, + 0.039855893701314926, + 0.1313890963792801, + 0.08989860862493515, + 0.0806116908788681, + 0.16352055966854095, + 0.11041510850191116, + 0.1496310532093048, + 0.0752987340092659 + ], + "credit_map": [ + 0.01495484821498394, + 0.025353219360113144, + 0.01796358823776245, + 0.013472228311002254, + 0.011208781972527504, + 0.013192460872232914, + 0.014938383363187313, + 0.013770770281553268, + 0.003832901595160365, + 0.004842334892600775, + 0.0045415242202579975, + 0.003927279729396105, + 0.003860929748043418, + 0.005428576376289129, + 0.002637522527948022, + 0.005453954450786114, + 0.006063321605324745, + 0.0043157897889614105, + 0.00553558487445116, + 0.004246570635586977, + 0.004695715382695198, + 0.004616692196577787, + 0.00501128938049078, + 0.005370929837226868, + 0.014330276288092136, + 0.00787990540266037, + 0.024553021416068077, + 0.012206071056425571, + 0.007373607251793146, + 0.026305891573429108, + 0.010331545025110245, + 0.028408780694007874, + 0.014529929496347904, + 0.008007477037608624, + 0.016206441447138786, + 0.0066961790435016155, + 0.010557249188423157, + 0.011430379934608936, + 0.01414281316101551, + 0.008328127674758434, + 0.015733858570456505, + 0.004883484449237585, + 0.005132483784109354, + 0.007389052305370569, + 0.005009805783629417, + 0.020457478240132332, + 0.006085983011871576, + 0.015919538214802742, + 0.022819548845291138, + 0.014491577632725239, + 0.03351627662777901, + 0.014946170151233673, + 0.021267639473080635, + 0.01897256448864937, + 0.019257880747318268, + 0.018248897045850754, + 0.024529622867703438, + 0.010358856990933418, + 0.02796698361635208, + 0.005381887778639793, + 0.00917830504477024, + 0.009155305102467537, + 0.008703198283910751, + 0.015140950679779053, + 0.01729387789964676, + 0.013097664341330528, + 0.028699547052383423, + 0.016715912148356438, + 0.018840456381440163, + 0.018234416842460632, + 0.017341114580631256, + 0.01940806582570076, + 0.007517052348703146, + 0.00808388739824295, + 0.010275491513311863, + 0.00939792301505804, + 0.01021911483258009, + 0.011450347490608692, + 0.007504330016672611, + 0.010850585997104645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.00319268717430532 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up shot, shot from a slightly upward angle and employing an extremely shallow depth of field, precisely captures an exquisite antique watch being gently tossed into the air. The watch features a polished rose gold case, an ivory dial adorned with slender Roman numerals and delicate blued-steel Breguet hands, and a rich, dark brown alligator leather strap. As the watch gracefully and slowly turns freely in the air, its domed sapphire crystal and metal case capture and reflect the warm surrounding light, shimmering fleeting, dazzling bursts of light with each rotation. In the background, the figures of friends laughing and chatting are blurred into soft, dynamic silhouettes, while the warm interior lighting transforms into captivating circular circles of light, creating an atmosphere of intimacy, joy, and the essence of life, capturing the beauty of this moment in eternal joy.", + "seed": 25089, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_9709_seed_25089_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": -0.04636559477075934, + "vlm_logprob": -0.04636559477075934, + "vlm_yes_no_margin": 3.7749999523162843, + "vlm_token_credit": -0.04636559477075934, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014285323210060596 + ], + "opposite_logprob_tokens": [ + -4.514285087585449 + ], + "logprob_reward": -0.014285323210060596, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014285323210060596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16043563187122345, + 0.15984037518501282, + 0.07061921060085297, + 0.08463165163993835, + 0.06599798053503036, + 0.03847809135913849, + 0.12280815094709396, + 0.06499005109071732, + 0.15353885293006897, + 0.0786600187420845 + ], + "credit_map": [ + 0.01991122029721737, + 0.017602719366550446, + 0.017281843349337578, + 0.02497183158993721, + 0.023719578981399536, + 0.020518209785223007, + 0.017977939918637276, + 0.018452277407050133, + 0.020750781521201134, + 0.014850597828626633, + 0.021214444190263748, + 0.017280152067542076, + 0.014130571857094765, + 0.01841769739985466, + 0.013989058323204517, + 0.03920707479119301, + 0.0065928855910897255, + 0.008812745101749897, + 0.00889760535210371, + 0.007305559702217579, + 0.009573171846568584, + 0.007550409063696861, + 0.008436253294348717, + 0.013450577855110168, + 0.009856075048446655, + 0.006208292208611965, + 0.010630802251398563, + 0.007988973520696163, + 0.010396836325526237, + 0.016738025471568108, + 0.012509015388786793, + 0.010303637944161892, + 0.008950410410761833, + 0.009898673743009567, + 0.010451255366206169, + 0.0047837430611252785, + 0.006497472058981657, + 0.0075070820748806, + 0.007815816439688206, + 0.010093532502651215, + 0.004174881614744663, + 0.005019359290599823, + 0.0055036949925124645, + 0.004989622160792351, + 0.004264960531145334, + 0.004255351144820452, + 0.005159230437129736, + 0.0051109930500388145, + 0.009260360151529312, + 0.02376195415854454, + 0.01401766948401928, + 0.012958238832652569, + 0.014415512792766094, + 0.01585680991411209, + 0.01706990972161293, + 0.015467698685824871, + 0.0058458480052649975, + 0.009517406113445759, + 0.009815743193030357, + 0.007089794147759676, + 0.010631070472300053, + 0.007816949859261513, + 0.007885197177529335, + 0.006388034671545029, + 0.013515685684978962, + 0.018209783360362053, + 0.020703332498669624, + 0.025961296632885933, + 0.027503961697220802, + 0.015058915130794048, + 0.018029076978564262, + 0.01455678790807724, + 0.009024912491440773, + 0.007818790152668953, + 0.019187262281775475, + 0.007315231487154961, + 0.00668193306773901, + 0.004890124779194593, + 0.0053338357247412205, + 0.018407927826046944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014285323210060596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202990293502808 + ], + "opposite_logprob_tokens": [ + -4.018202781677246 + ], + "logprob_reward": -0.018202990293502808, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018202990293502808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07060027122497559, + 0.1359873116016388, + 0.03878827765583992, + 0.21633343398571014, + 0.137999027967453, + 0.10784932225942612, + 0.05367784947156906, + 0.07361557334661484, + 0.10091433674097061, + 0.06423459202051163 + ], + "credit_map": [ + 0.008212354965507984, + 0.01091086771339178, + 0.007912415079772472, + 0.009664716199040413, + 0.010761866346001625, + 0.007726410403847694, + 0.008472277782857418, + 0.006939360871911049, + 0.012843344360589981, + 0.01669476181268692, + 0.01977529563009739, + 0.021032342687249184, + 0.01653514988720417, + 0.013580076396465302, + 0.018451012670993805, + 0.017075328156352043, + 0.003693044651299715, + 0.004845735151320696, + 0.005307274404913187, + 0.0045400564558804035, + 0.00469427602365613, + 0.004056177567690611, + 0.006354259792715311, + 0.005297452211380005, + 0.022021271288394928, + 0.01137060672044754, + 0.0711822360754013, + 0.017726827412843704, + 0.013095730915665627, + 0.028228947892785072, + 0.012276336550712585, + 0.04043148085474968, + 0.03634609654545784, + 0.017336701974272728, + 0.02590751461684704, + 0.010492183268070221, + 0.008407111279666424, + 0.014846459962427616, + 0.00872484315186739, + 0.015938112512230873, + 0.00826955121010542, + 0.012753495015203953, + 0.017454195767641068, + 0.020640123635530472, + 0.012977377511560917, + 0.0096891513094306, + 0.013592620380222797, + 0.012472805567085743, + 0.004262508824467659, + 0.008383582346141338, + 0.005518765188753605, + 0.005109018180519342, + 0.0075558871030807495, + 0.00864969938993454, + 0.007835780270397663, + 0.006362604442983866, + 0.009457839652895927, + 0.00815712008625269, + 0.008904047310352325, + 0.010784901678562164, + 0.006867272779345512, + 0.008564081974327564, + 0.011187058873474598, + 0.009693248197436333, + 0.010831624269485474, + 0.010437714867293835, + 0.01599784940481186, + 0.01571308821439743, + 0.014606213197112083, + 0.00991050060838461, + 0.012751519680023193, + 0.010665823705494404, + 0.007711643353104591, + 0.00797529798001051, + 0.012483686208724976, + 0.006400072947144508, + 0.005575207062065601, + 0.004213330335915089, + 0.005135579500347376, + 0.01473977416753769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.018202990293502808 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033707089722156525 + ], + "opposite_logprob_tokens": [ + -3.4087071418762207 + ], + "logprob_reward": -0.033707089722156525, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033707089722156525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08160493522882462, + 0.07140461355447769, + 0.09229202568531036, + 0.22535103559494019, + 0.11120978742837906, + 0.06135529279708862, + 0.044772617518901825, + 0.040286820381879807, + 0.18326576054096222, + 0.08845707774162292 + ], + "credit_map": [ + 0.011245510540902615, + 0.009774347767233849, + 0.008586592972278595, + 0.013287022709846497, + 0.01258020382374525, + 0.007697443012148142, + 0.008843286894261837, + 0.009590528905391693, + 0.00706947548314929, + 0.006168364081531763, + 0.009897243231534958, + 0.007635660469532013, + 0.006525304168462753, + 0.007917007431387901, + 0.007943457923829556, + 0.01824810542166233, + 0.008741459809243679, + 0.008735843002796173, + 0.011505160480737686, + 0.010097133927047253, + 0.008260770700871944, + 0.014761816710233688, + 0.006392019335180521, + 0.023797819390892982, + 0.024660460650920868, + 0.013211864046752453, + 0.07333575189113617, + 0.01884281262755394, + 0.014234756119549274, + 0.029619136825203896, + 0.013513577170670033, + 0.037932686507701874, + 0.03122113272547722, + 0.012708171270787716, + 0.01802162267267704, + 0.008411355316638947, + 0.004662955179810524, + 0.012673490680754185, + 0.0060240658931434155, + 0.017487000674009323, + 0.005524102598428726, + 0.007094954140484333, + 0.008839177899062634, + 0.011177637614309788, + 0.006184061523526907, + 0.008375298231840134, + 0.006272295955568552, + 0.007887763902544975, + 0.002948782639577985, + 0.0066987755708396435, + 0.005338145885616541, + 0.004523507785052061, + 0.005430983379483223, + 0.00731395510956645, + 0.0072686574421823025, + 0.005249808542430401, + 0.0040369294583797455, + 0.004920287523418665, + 0.0058593470603227615, + 0.0040274690836668015, + 0.005641236901283264, + 0.005356507375836372, + 0.005923899356275797, + 0.004521140828728676, + 0.020230360329151154, + 0.018210003152489662, + 0.024269288405776024, + 0.026474766433238983, + 0.025525903329253197, + 0.029931887984275818, + 0.01996190845966339, + 0.018661634996533394, + 0.010486450046300888, + 0.007765398360788822, + 0.023705868050456047, + 0.00542064243927598, + 0.006226932629942894, + 0.005550542380660772, + 0.004167543258517981, + 0.02513369917869568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.033707089722156525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038008034229279 + ], + "opposite_logprob_tokens": [ + -1.9103801250457764 + ], + "logprob_reward": -0.16038008034229279, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038008034229279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06457355618476868, + 0.04814199358224869, + 0.028649145737290382, + 0.40908002853393555, + 0.2710127532482147, + 0.02793571725487709, + 0.05445405840873718, + 0.025500435382127762, + 0.048312924802303314, + 0.02233940176665783 + ], + "credit_map": [ + 0.00601913221180439, + 0.009365824051201344, + 0.00849870964884758, + 0.009469636715948582, + 0.009049991145730019, + 0.008985606953501701, + 0.006918157450854778, + 0.006266504060477018, + 0.0052014244720339775, + 0.003390226047486067, + 0.006555201951414347, + 0.004298534244298935, + 0.004407141823321581, + 0.005831393413245678, + 0.005098089575767517, + 0.013359985314309597, + 0.002608716022223234, + 0.0026359918992966413, + 0.004643124062567949, + 0.005233635660260916, + 0.0027940976433455944, + 0.003307998413220048, + 0.0037186392582952976, + 0.0037069441750645638, + 0.04083608090877533, + 0.021237893030047417, + 0.1298460066318512, + 0.033574748784303665, + 0.024882527068257332, + 0.05477720499038696, + 0.027455832809209824, + 0.07646974921226501, + 0.07530438154935837, + 0.03202156350016594, + 0.041683997958898544, + 0.01911984011530876, + 0.012303235940635204, + 0.030211612582206726, + 0.016020050272345543, + 0.04434807598590851, + 0.0026188543997704983, + 0.002424654783681035, + 0.003726785071194172, + 0.0028930201660841703, + 0.0026627597399055958, + 0.004890813957899809, + 0.0039002145640552044, + 0.004818615969270468, + 0.004403369035571814, + 0.007558914367109537, + 0.007141476031392813, + 0.006473499350249767, + 0.006244601681828499, + 0.008034097962081432, + 0.008229676634073257, + 0.006368427537381649, + 0.0025426242500543594, + 0.0028125622775405645, + 0.0031842405442148447, + 0.002831751946359873, + 0.0026850998401641846, + 0.004571178462356329, + 0.003994205966591835, + 0.0028787741903215647, + 0.005336727946996689, + 0.004445655271410942, + 0.00612681545317173, + 0.006868643220514059, + 0.006690165027976036, + 0.008580505847930908, + 0.0052202981896698475, + 0.005044118035584688, + 0.0028208261355757713, + 0.0017918945522978902, + 0.004271190147846937, + 0.001686129835434258, + 0.0022465342190116644, + 0.002289858413860202, + 0.00250469078309834, + 0.004728279542177916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.16038008034229279 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05666639283299446, + 0.030662821605801582, + 0.03980880230665207, + 0.2461884319782257, + 0.19824177026748657, + 0.060681719332933426, + 0.03715554252266884, + 0.07105115056037903, + 0.21919840574264526, + 0.040344931185245514 + ], + "credit_map": [ + 0.006944212131202221, + 0.007973434403538704, + 0.007437978871166706, + 0.00856386311352253, + 0.00857388973236084, + 0.006444667000323534, + 0.005220448132604361, + 0.005507896654307842, + 0.0029110030736774206, + 0.0032850357238203287, + 0.005201082210987806, + 0.0039025871083140373, + 0.0025344754103571177, + 0.004069954622536898, + 0.0036877549719065428, + 0.005070928949862719, + 0.0048238011077046394, + 0.004274631384760141, + 0.0054631163366138935, + 0.006400425452739, + 0.004087601322680712, + 0.0056676543317735195, + 0.003570117522031069, + 0.005521456245332956, + 0.02741892635822296, + 0.014081819914281368, + 0.08217219263315201, + 0.020315706729888916, + 0.015271554701030254, + 0.03057667799293995, + 0.01459123007953167, + 0.04176032915711403, + 0.04696335271000862, + 0.027965856716036797, + 0.042554955929517746, + 0.014274341985583305, + 0.013294253498315811, + 0.01883269101381302, + 0.010882432572543621, + 0.023473894223570824, + 0.00564050767570734, + 0.008760754950344563, + 0.010173973627388477, + 0.012942264787852764, + 0.005137223284691572, + 0.004912257194519043, + 0.004795589484274387, + 0.00831915158778429, + 0.0029073816258460283, + 0.005042686127126217, + 0.005128583870828152, + 0.003642249619588256, + 0.004917612764984369, + 0.006590158212929964, + 0.004747905768454075, + 0.004178965464234352, + 0.0058860983699560165, + 0.00920732319355011, + 0.008384020999073982, + 0.007339079864323139, + 0.010976030491292477, + 0.009016592055559158, + 0.011712480336427689, + 0.00852952804416418, + 0.022119065746665, + 0.02444063313305378, + 0.028932100161910057, + 0.03174614533782005, + 0.035892512649297714, + 0.030667535960674286, + 0.023768026381731033, + 0.021632373332977295, + 0.003804033389315009, + 0.003916570916771889, + 0.010517661459743977, + 0.0024905456230044365, + 0.003972766455262899, + 0.0026470955926924944, + 0.0028092979919165373, + 0.010186959989368916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.005252490285784006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up shot, shot from a slightly upward angle and employing an extremely shallow depth of field, precisely captures an exquisite antique watch being gently tossed into the air. The watch features a polished rose gold case, an ivory dial adorned with slender Roman numerals and delicate blued-steel Breguet hands, and a rich, dark brown alligator leather strap. As the watch gracefully and slowly turns freely in the air, its domed sapphire crystal and metal case capture and reflect the warm surrounding light, shimmering fleeting, dazzling bursts of light with each rotation. In the background, the figures of friends laughing and chatting are blurred into soft, dynamic silhouettes, while the warm interior lighting transforms into captivating circular circles of light, creating an atmosphere of intimacy, joy, and the essence of life, capturing the beauty of this moment in eternal joy.", + "seed": 25090, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_9709_seed_25090_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": -0.010671459510922432, + "vlm_logprob": -0.010671459510922432, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.010671459510922432, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007720393594354391 + ], + "opposite_logprob_tokens": [ + -5.507720470428467 + ], + "logprob_reward": -0.007720393594354391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007720393594354391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12245538830757141, + 0.04957396909594536, + 0.039303433150053024, + 0.11107095330953598, + 0.06840690225362778, + 0.3460501432418823, + 0.08151917904615402, + 0.06682436913251877, + 0.07377123832702637, + 0.04102436453104019 + ], + "credit_map": [ + 0.013180851005017757, + 0.012026643380522728, + 0.025866907089948654, + 0.012568006291985512, + 0.013075496070086956, + 0.011908558197319508, + 0.01457875408232212, + 0.019250184297561646, + 0.0049638343043625355, + 0.0050034187734127045, + 0.004689875524491072, + 0.006550238002091646, + 0.006492720916867256, + 0.009502185508608818, + 0.006784344092011452, + 0.005587354768067598, + 0.003635774366557598, + 0.0034882184118032455, + 0.004709238652139902, + 0.006016179453581572, + 0.005289220251142979, + 0.006206141319125891, + 0.005785069894045591, + 0.004173594992607832, + 0.012853647582232952, + 0.009876614436507225, + 0.017808493226766586, + 0.021161409094929695, + 0.010078630410134792, + 0.01029991451650858, + 0.017107654362916946, + 0.0118845971301198, + 0.010685933753848076, + 0.004591997247189283, + 0.01539598684757948, + 0.0071883425116539, + 0.004927571397274733, + 0.011336206458508968, + 0.005879752337932587, + 0.008401121012866497, + 0.042843736708164215, + 0.012241245247423649, + 0.08966615051031113, + 0.028971906751394272, + 0.01800360158085823, + 0.03961671143770218, + 0.029649753123521805, + 0.08505706489086151, + 0.012930374592542648, + 0.007538756355643272, + 0.02239120937883854, + 0.00784824788570404, + 0.005563450511544943, + 0.0083537045866251, + 0.008412138558924198, + 0.00848130602389574, + 0.004352239426225424, + 0.003904164070263505, + 0.011671354994177818, + 0.009613985195755959, + 0.005329245701432228, + 0.014961302280426025, + 0.009170672856271267, + 0.007821409963071346, + 0.0037481593899428844, + 0.005827557761222124, + 0.015559462830424309, + 0.008549626916646957, + 0.005917943082749844, + 0.017830094322562218, + 0.010465871542692184, + 0.005872530397027731, + 0.0029363667126744986, + 0.003747625043615699, + 0.004735149908810854, + 0.004669330548495054, + 0.004966652020812035, + 0.006007606163620949, + 0.009582742117345333, + 0.004378896672278643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.007720393594354391 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550539337098598 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.012550539337098598, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550539337098598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09107032418251038, + 0.17982368171215057, + 0.1399085372686386, + 0.08035746216773987, + 0.03996757045388222, + 0.16291682422161102, + 0.12072532624006271, + 0.08103526383638382, + 0.053540874272584915, + 0.05065419152379036 + ], + "credit_map": [ + 0.009085042402148247, + 0.011925711296498775, + 0.011820427142083645, + 0.01250478532165289, + 0.01685471087694168, + 0.009320340119302273, + 0.00906784925609827, + 0.010491456836462021, + 0.020683161914348602, + 0.02257588319480419, + 0.024403871968388557, + 0.025213802233338356, + 0.022908469662070274, + 0.02141890488564968, + 0.024528885260224342, + 0.018090704455971718, + 0.0098440395668149, + 0.013812597841024399, + 0.021443847566843033, + 0.028385091572999954, + 0.01985517144203186, + 0.009583162143826485, + 0.024896124377846718, + 0.012088499963283539, + 0.009066350758075714, + 0.008646802976727486, + 0.008596707135438919, + 0.01821078546345234, + 0.006997900549322367, + 0.00734700495377183, + 0.011836488731205463, + 0.009655422531068325, + 0.006544842384755611, + 0.0024661319330334663, + 0.006880482193082571, + 0.005313423927873373, + 0.002691685687750578, + 0.007032678462564945, + 0.003930690698325634, + 0.005107634700834751, + 0.018697142601013184, + 0.008373213931918144, + 0.04233945533633232, + 0.014894790947437286, + 0.0106367701664567, + 0.016150765120983124, + 0.015813151374459267, + 0.03601152449846268, + 0.010895908810198307, + 0.015113682486116886, + 0.015004327520728111, + 0.024971019476652145, + 0.018813537433743477, + 0.010659720748662949, + 0.014453726820647717, + 0.010813390836119652, + 0.004008350428193808, + 0.0024081540759652853, + 0.02452084608376026, + 0.008114269934594631, + 0.0033091618679463863, + 0.029971329495310783, + 0.004019063897430897, + 0.004684088286012411, + 0.003705647075548768, + 0.007055225316435099, + 0.007993444800376892, + 0.0074918526224792, + 0.006596734747290611, + 0.005270830821245909, + 0.00973878800868988, + 0.005688350647687912, + 0.003332599066197872, + 0.00716436468064785, + 0.004522278904914856, + 0.007246439345180988, + 0.008019699715077877, + 0.008072013966739178, + 0.007062244229018688, + 0.0052345506846904755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.012550539337098598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069071367383003 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.016069071367383003, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069071367383003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.244339719414711, + 0.22825367748737335, + 0.04555220901966095, + 0.037333134561777115, + 0.08149511367082596, + 0.1358269453048706, + 0.06656503677368164, + 0.08682195842266083, + 0.033777959644794464, + 0.04003427177667618 + ], + "credit_map": [ + 0.022271528840065002, + 0.023664385080337524, + 0.04188643395900726, + 0.02921302616596222, + 0.018259692937135696, + 0.03000936098396778, + 0.02894992008805275, + 0.05008536949753761, + 0.02364197187125683, + 0.025634482502937317, + 0.017284758388996124, + 0.02141983062028885, + 0.04769677668809891, + 0.03677051514387131, + 0.03141089528799057, + 0.02439444698393345, + 0.004336006473749876, + 0.00462365010753274, + 0.006163155194371939, + 0.007243473082780838, + 0.004322720225900412, + 0.006971614435315132, + 0.006290489342063665, + 0.00560110155493021, + 0.004824504721909761, + 0.003083476796746254, + 0.004135937429964542, + 0.006347971968352795, + 0.0034204914700239897, + 0.0049372464418411255, + 0.006145586725324392, + 0.00443792250007391, + 0.01916452683508396, + 0.007509210612624884, + 0.020066609606146812, + 0.012437654659152031, + 0.0052846819162368774, + 0.004964383319020271, + 0.00835720170289278, + 0.003710841992869973, + 0.01793629489839077, + 0.004891865886747837, + 0.04041942209005356, + 0.010358982719480991, + 0.005867163185030222, + 0.014825450256466866, + 0.011537129990756512, + 0.02999063953757286, + 0.007848463021218777, + 0.005656143184751272, + 0.01684722863137722, + 0.006950381211936474, + 0.00782772246748209, + 0.0059712352231144905, + 0.00801737792789936, + 0.007446488831192255, + 0.0052628181874752045, + 0.0032866881228983402, + 0.026825308799743652, + 0.007427550852298737, + 0.005223496817052364, + 0.02586350403726101, + 0.004653473850339651, + 0.00827911775559187, + 0.003430599346756935, + 0.002656898694112897, + 0.006634663790464401, + 0.0031659849919378757, + 0.0030713407322764397, + 0.008146191015839577, + 0.003226814093068242, + 0.003445466747507453, + 0.003292446956038475, + 0.004444459918886423, + 0.003851803718134761, + 0.005043698009103537, + 0.005515026859939098, + 0.0068792011588811874, + 0.006709688808768988, + 0.0042979479767382145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.016069071367383003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071204207837582 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071204207837582, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071204207837582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09006878733634949, + 0.08586984127759933, + 0.10660146921873093, + 0.03417159244418144, + 0.11921437084674835, + 0.2602320909500122, + 0.11295963078737259, + 0.06165988743305206, + 0.028560692444443703, + 0.10066168755292892 + ], + "credit_map": [ + 0.012499870732426643, + 0.010110568255186081, + 0.010130136273801327, + 0.008757678791880608, + 0.010273732244968414, + 0.01655631884932518, + 0.006155300419777632, + 0.015585173852741718, + 0.007693313527852297, + 0.010428965091705322, + 0.009063196368515491, + 0.0084633594378829, + 0.016983507201075554, + 0.013770198449492455, + 0.010062208399176598, + 0.009405091404914856, + 0.008738340809941292, + 0.014092608354985714, + 0.011554254218935966, + 0.015413325279951096, + 0.015063169412314892, + 0.01749718375504017, + 0.012274644337594509, + 0.011967924423515797, + 0.00415353337302804, + 0.002807033248245716, + 0.004722265526652336, + 0.005460893735289574, + 0.0038499473594129086, + 0.0036757460329681635, + 0.00565524585545063, + 0.0038469252176582813, + 0.021926725283265114, + 0.0075179580599069595, + 0.03399735316634178, + 0.011265878565609455, + 0.005722433794289827, + 0.014062779024243355, + 0.011955373920500278, + 0.012765863910317421, + 0.034798555076122284, + 0.011819656938314438, + 0.060819823294878006, + 0.018544208258390427, + 0.011078900657594204, + 0.033083587884902954, + 0.020257286727428436, + 0.0698300451040268, + 0.012269752100110054, + 0.01554231345653534, + 0.018443267792463303, + 0.012543530203402042, + 0.010764438658952713, + 0.01130594126880169, + 0.012918136082589626, + 0.019172245636582375, + 0.003818373428657651, + 0.002511913888156414, + 0.01993716135621071, + 0.005511763971298933, + 0.0036090947687625885, + 0.018911857157945633, + 0.002676465082913637, + 0.00468325475230813, + 0.0032151900231838226, + 0.002631844487041235, + 0.00396386394277215, + 0.003660716349259019, + 0.0024264086969196796, + 0.0027140039019286633, + 0.006437594071030617, + 0.0035110702738165855, + 0.009716600179672241, + 0.011143059469759464, + 0.01115006860345602, + 0.013387992046773434, + 0.013374458998441696, + 0.016123678535223007, + 0.01654570922255516, + 0.009220115840435028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.011071204207837582 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of an antique watch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch being gently tossed into the air and turning freely?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch have a polished rose gold case, an ivory dial with Roman numerals, and a dark brown alligator leather strap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing blurred figures of friends laughing and chatting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with an extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946089047938585 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946089047938585, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946089047938585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13906224071979523, + 0.06743119657039642, + 0.08096136152744293, + 0.09429172426462173, + 0.0561409667134285, + 0.07694685459136963, + 0.1692398190498352, + 0.11498472094535828, + 0.12474953383207321, + 0.07619158178567886 + ], + "credit_map": [ + 0.01603292115032673, + 0.013962717726826668, + 0.024117523804306984, + 0.0177744273096323, + 0.01868578977882862, + 0.013391095213592052, + 0.021121632307767868, + 0.01397613249719143, + 0.007575955707579851, + 0.006880567874759436, + 0.009017950855195522, + 0.006885203532874584, + 0.009409469552338123, + 0.012839257717132568, + 0.007038718089461327, + 0.007784072309732437, + 0.008833608590066433, + 0.00846218690276146, + 0.007747676223516464, + 0.01565125212073326, + 0.010248449631035328, + 0.008403167128562927, + 0.013793962076306343, + 0.007821058854460716, + 0.00949806347489357, + 0.009154953062534332, + 0.012813747860491276, + 0.017165079712867737, + 0.0072672124952077866, + 0.012096664868295193, + 0.013764367438852787, + 0.012531637214124203, + 0.009370695799589157, + 0.0068762400187551975, + 0.008895625360310078, + 0.006780679803341627, + 0.004621497821062803, + 0.00804260466247797, + 0.0050173369236290455, + 0.006536286789923906, + 0.009251611307263374, + 0.004316911567002535, + 0.020377716049551964, + 0.009640787728130817, + 0.006044179201126099, + 0.006382626015692949, + 0.006284366827458143, + 0.014648656360805035, + 0.014952095225453377, + 0.02317626215517521, + 0.020584043115377426, + 0.018924735486507416, + 0.02115737646818161, + 0.023686757311224937, + 0.022846706211566925, + 0.023911846801638603, + 0.005612260662019253, + 0.0056314850226044655, + 0.03463812917470932, + 0.011568292044103146, + 0.007582373451441526, + 0.03690987080335617, + 0.006948165129870176, + 0.006094141863286495, + 0.009324575774371624, + 0.010016296990215778, + 0.02834870107471943, + 0.013809875585138798, + 0.010584624484181404, + 0.017435666173696518, + 0.02308039925992489, + 0.012149394489824772, + 0.006019093096256256, + 0.009259960614144802, + 0.00803958997130394, + 0.0070274570025503635, + 0.008433225564658642, + 0.009782699868083, + 0.020145585760474205, + 0.00748396897688508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.005946089047938585 + } + ], + "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/0000128/group_8/index_tensor([5441, 5441])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_8/index_tensor([5441, 5441])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b37e2ab3b6203dc293e8bae122ac4fbe38384ff0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_8/index_tensor([5441, 5441])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle close-up captures a pair of vintage four-wheeled roller skates swaying hesitantly across a rough, dark-gray asphalt surface. The skates are clad in worn, seaweed-green leather boots, their laces slightly frayed. Their polyurethane wheels are a milky white, speckled with the grit from the ground. Warm, golden sunlight bathes the scene in late afternoon light, capturing the dull gleam of the aluminum panels and casting long, soft shadows that accentuate the skates' precarious motion. One skate tilts precariously inward, its outer wheel lifting just a millimeter off the pavement before correcting with a clumsy weight shift, passing the jolt to the other skate. This precarious back-and-forth, the only movement in the frame, creates a concentrated tension. The entire scene is rendered with the rich, cinematic quality of 35mm film, characterized by delicate, organic textures, deep color saturation, and a gentle halo around the sunlit edges of the skates, evoking a nostalgic and slightly anxious atmosphere of precarious equilibrium.", + "seed": 18517, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_8/index_5441_seed_18517_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": -0.010848061367869376, + "vlm_logprob": -0.010848061367869376, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.010848061367869376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01631280966103077 + ], + "opposite_logprob_tokens": [ + -4.516312599182129 + ], + "logprob_reward": -0.01631280966103077, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01631280966103077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16318310797214508, + 0.04085281863808632, + 0.09175118803977966, + 0.022720996290445328, + 0.05345892906188965, + 0.34229183197021484, + 0.07829707115888596, + 0.05845648795366287, + 0.029022641479969025, + 0.11996490508317947 + ], + "credit_map": [ + 0.019288867712020874, + 0.018411505967378616, + 0.02371959201991558, + 0.026241997256875038, + 0.022310473024845123, + 0.01727992482483387, + 0.018178267404437065, + 0.017752502113580704, + 0.005378447473049164, + 0.003274848684668541, + 0.00859152153134346, + 0.008446364663541317, + 0.003848011838272214, + 0.003913543187081814, + 0.0029546860605478287, + 0.00444539962336421, + 0.0082472525537014, + 0.008743979036808014, + 0.014174801297485828, + 0.014065667055547237, + 0.008832610212266445, + 0.011094188317656517, + 0.012497578747570515, + 0.014095114544034004, + 0.002014998346567154, + 0.004912625998258591, + 0.00281907944008708, + 0.0026293371338397264, + 0.0027641444467008114, + 0.0035121680703014135, + 0.0021636257879436016, + 0.001905019860714674, + 0.005262129008769989, + 0.00911323819309473, + 0.007802183739840984, + 0.005288256332278252, + 0.006983467843383551, + 0.008260924369096756, + 0.004192937631160021, + 0.0065557933412492275, + 0.04267353564500809, + 0.07330089807510376, + 0.04045134037733078, + 0.04922468215227127, + 0.05404333397746086, + 0.029043231159448624, + 0.019361237064003944, + 0.034193601459264755, + 0.006716354284435511, + 0.005697515327483416, + 0.01489490456879139, + 0.008074862882494926, + 0.01115286722779274, + 0.009420186281204224, + 0.010898737236857414, + 0.011441648937761784, + 0.005883936304599047, + 0.004952446091920137, + 0.01032286137342453, + 0.011552567593753338, + 0.005699274595826864, + 0.004868747666478157, + 0.0085068978369236, + 0.0066697620786726475, + 0.0025510231498628855, + 0.002460080198943615, + 0.0033573557157069445, + 0.003488196525722742, + 0.0021750200539827347, + 0.005482300650328398, + 0.004238656256347895, + 0.0052700103260576725, + 0.006422212813049555, + 0.015640629455447197, + 0.024589883163571358, + 0.018815016373991966, + 0.01204065140336752, + 0.00906301662325859, + 0.011498624458909035, + 0.02189488336443901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.01631280966103077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954266060143709 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954266060143709, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954266060143709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13848790526390076, + 0.059089984744787216, + 0.09020879119634628, + 0.07500339299440384, + 0.19063882529735565, + 0.1521574854850769, + 0.06862496584653854, + 0.05165726691484451, + 0.08369015157222748, + 0.09044124186038971 + ], + "credit_map": [ + 0.011093470267951488, + 0.018120158463716507, + 0.017711471766233444, + 0.015157106332480907, + 0.031108791008591652, + 0.01950182393193245, + 0.012132637202739716, + 0.013662435114383698, + 0.006212037522345781, + 0.007341294549405575, + 0.007498000282794237, + 0.008260929957032204, + 0.009240520186722279, + 0.008203824050724506, + 0.004723791498690844, + 0.007609587162733078, + 0.009337250143289566, + 0.012278169393539429, + 0.01199683453887701, + 0.009240190498530865, + 0.01016313023865223, + 0.01179942861199379, + 0.014391680248081684, + 0.01100210566073656, + 0.008031207136809826, + 0.015746736899018288, + 0.012017922475934029, + 0.008942737244069576, + 0.008925608359277248, + 0.008260078728199005, + 0.0067467959597706795, + 0.006332302466034889, + 0.02406552992761135, + 0.025896072387695312, + 0.02615293674170971, + 0.02192426659166813, + 0.021320538595318794, + 0.024999285116791725, + 0.02634504996240139, + 0.019935138523578644, + 0.008235812187194824, + 0.018922027200460434, + 0.021019697189331055, + 0.023639511317014694, + 0.01790960505604744, + 0.014076847583055496, + 0.019139958545565605, + 0.02921401709318161, + 0.0073685552924871445, + 0.005208856891840696, + 0.013088826090097427, + 0.01125213224440813, + 0.006604204420000315, + 0.0072738174349069595, + 0.008373923599720001, + 0.00945464987307787, + 0.005246574990451336, + 0.005595333408564329, + 0.010373810306191444, + 0.00659989146515727, + 0.006651764735579491, + 0.0046312385238707066, + 0.0063594626262784, + 0.006199186202138662, + 0.007470380514860153, + 0.005102753639221191, + 0.009757361374795437, + 0.011341563425958157, + 0.006240139249712229, + 0.017237119376659393, + 0.011989586986601353, + 0.014551239088177681, + 0.013167849741876125, + 0.01041297148913145, + 0.011530940420925617, + 0.01251262053847313, + 0.009153586812317371, + 0.00985727272927761, + 0.010325671173632145, + 0.013480328023433685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.005954266060143709 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532292865216732 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532292865216732, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532292865216732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059385526925325394, + 0.0726100355386734, + 0.07869084179401398, + 0.0447939895093441, + 0.21749843657016754, + 0.12971845269203186, + 0.06364080309867859, + 0.0496160164475441, + 0.05664439499378204, + 0.2274014800786972 + ], + "credit_map": [ + 0.005543860606849194, + 0.00816061906516552, + 0.007665344979614019, + 0.007619660347700119, + 0.012638718821108341, + 0.006645961664617062, + 0.004909512121230364, + 0.006201850716024637, + 0.00762291019782424, + 0.012351362966001034, + 0.009926176629960537, + 0.007778808940201998, + 0.013224315829575062, + 0.011101777665317059, + 0.0038754467386752367, + 0.006729239132255316, + 0.007108165416866541, + 0.009468945674598217, + 0.015581644140183926, + 0.009255626238882542, + 0.007182005792856216, + 0.008899698033928871, + 0.013605435378849506, + 0.007589324843138456, + 0.003868505824357271, + 0.005050087813287973, + 0.007949726656079292, + 0.007155074737966061, + 0.003685096511617303, + 0.005370755214244127, + 0.0064502889290452, + 0.005264452192932367, + 0.02608734555542469, + 0.026747126132249832, + 0.031463269144296646, + 0.025432011112570763, + 0.020522408187389374, + 0.024012180045247078, + 0.03529473394155502, + 0.027939369902014732, + 0.015655705705285072, + 0.021130625158548355, + 0.01594099961221218, + 0.015032591298222542, + 0.018259622156620026, + 0.009572268463671207, + 0.01506712008267641, + 0.019059518352150917, + 0.006546809338033199, + 0.004195754416286945, + 0.015336504206061363, + 0.01026955246925354, + 0.006406538654118776, + 0.006606698036193848, + 0.007320544216781855, + 0.006958400830626488, + 0.005155673250555992, + 0.004195889458060265, + 0.008085405454039574, + 0.009273766539990902, + 0.005385315977036953, + 0.003909074701368809, + 0.007759503088891506, + 0.00585138937458396, + 0.00511992396786809, + 0.003979456145316362, + 0.0069597396068274975, + 0.008408674970269203, + 0.00446960236877203, + 0.01153391320258379, + 0.006905979476869106, + 0.009267102926969528, + 0.0339936688542366, + 0.02672525681555271, + 0.025309449061751366, + 0.03432149067521095, + 0.026906531304121017, + 0.02621031180024147, + 0.024105273187160492, + 0.029829489067196846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.012532292865216732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192361384630203 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192361384630203, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192361384630203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07323607057332993, + 0.05890049412846565, + 0.041015416383743286, + 0.07231307029724121, + 0.2187286615371704, + 0.11521288752555847, + 0.10427811741828918, + 0.09361853450536728, + 0.18227781355381012, + 0.04041893407702446 + ], + "credit_map": [ + 0.007661213167011738, + 0.007939988747239113, + 0.01098349504172802, + 0.01208576187491417, + 0.011219307780265808, + 0.00727257551625371, + 0.008314508013427258, + 0.007759221363812685, + 0.007318295072764158, + 0.006241305265575647, + 0.009968594647943974, + 0.008166360668838024, + 0.006350102834403515, + 0.007104580756276846, + 0.0055016763508319855, + 0.008249574340879917, + 0.005879065487533808, + 0.005056655500084162, + 0.004990475717931986, + 0.0047536613419651985, + 0.004662185441702604, + 0.007172583602368832, + 0.004605303984135389, + 0.0038954862393438816, + 0.006595213897526264, + 0.0106813320890069, + 0.010023667477071285, + 0.011476217769086361, + 0.006401911843568087, + 0.009099266491830349, + 0.009193159639835358, + 0.0088422866538167, + 0.025535356253385544, + 0.02697281539440155, + 0.032010942697525024, + 0.027070393785834312, + 0.020995143800973892, + 0.024738479405641556, + 0.035000991076231, + 0.02640453167259693, + 0.013829206116497517, + 0.019488446414470673, + 0.013336904346942902, + 0.018446601927280426, + 0.01825578138232231, + 0.011578922159969807, + 0.009419632144272327, + 0.010857385583221912, + 0.012352817691862583, + 0.008192316628992558, + 0.018440833315253258, + 0.013433802872896194, + 0.013241319917142391, + 0.010647681541740894, + 0.014365097507834435, + 0.013604244217276573, + 0.008737904019653797, + 0.010761553421616554, + 0.01615915820002556, + 0.017284635454416275, + 0.007519796956330538, + 0.00868862122297287, + 0.01175508089363575, + 0.0127117820084095, + 0.019382253289222717, + 0.013869685120880604, + 0.020384281873703003, + 0.0224301777780056, + 0.014189355075359344, + 0.03146662563085556, + 0.02702232450246811, + 0.033533088862895966, + 0.0027820372488349676, + 0.004628405906260014, + 0.006583081558346748, + 0.006643796339631081, + 0.0035221173893660307, + 0.003958264831453562, + 0.004186116624623537, + 0.008115109987556934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.014192361384630203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248576868325472 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248576868325472, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248576868325472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13207696378231049, + 0.09950706362724304, + 0.07359206676483154, + 0.06254011392593384, + 0.14429615437984467, + 0.06399597972631454, + 0.13661013543605804, + 0.06224710866808891, + 0.12837247550487518, + 0.09676199406385422 + ], + "credit_map": [ + 0.02032160386443138, + 0.012283893302083015, + 0.01971995271742344, + 0.01723545417189598, + 0.011920594610273838, + 0.015409831888973713, + 0.015298286452889442, + 0.019887343049049377, + 0.010235442779958248, + 0.012586625292897224, + 0.01339650433510542, + 0.012379953637719154, + 0.01660284399986267, + 0.013654603622853756, + 0.008379070088267326, + 0.01227201521396637, + 0.006799824070185423, + 0.007721490226686001, + 0.01123745460063219, + 0.007678811904042959, + 0.007981362752616405, + 0.010114436969161034, + 0.013721977360546589, + 0.00833671260625124, + 0.006392011418938637, + 0.010639536194503307, + 0.009505022317171097, + 0.007343435660004616, + 0.005966072902083397, + 0.008773580193519592, + 0.0067145018838346004, + 0.007205952424556017, + 0.014905028976500034, + 0.027966972440481186, + 0.019681081175804138, + 0.01881287805736065, + 0.01827419176697731, + 0.020843204110860825, + 0.012875515036284924, + 0.010937286540865898, + 0.00816989503800869, + 0.007928314618766308, + 0.00852967705577612, + 0.007318365387618542, + 0.008378459140658379, + 0.006102617364376783, + 0.010272961109876633, + 0.007295687682926655, + 0.013925254344940186, + 0.01528069656342268, + 0.026793142780661583, + 0.028449969366192818, + 0.011717156507074833, + 0.01017067115753889, + 0.014835006557404995, + 0.015438239090144634, + 0.006448028143495321, + 0.0074799275025725365, + 0.010131164453923702, + 0.011597339063882828, + 0.006455232854932547, + 0.00425134040415287, + 0.008061443455517292, + 0.007822633720934391, + 0.010841263458132744, + 0.008282340131700039, + 0.014687315560877323, + 0.017007317394018173, + 0.00860796868801117, + 0.022275293245911598, + 0.022164367139339447, + 0.024506615474820137, + 0.0058633144944906235, + 0.009595086798071861, + 0.01909925788640976, + 0.015659982338547707, + 0.008246142417192459, + 0.008697042241692543, + 0.01153476070612669, + 0.01806640811264515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.005248576868325472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a pair of vintage four-wheeled roller skates swaying hesitantly across a rough, dark-gray asphalt surface. The skates are clad in worn, seaweed-green leather boots, their laces slightly frayed. Their polyurethane wheels are a milky white, speckled with the grit from the ground. Warm, golden sunlight bathes the scene in late afternoon light, capturing the dull gleam of the aluminum panels and casting long, soft shadows that accentuate the skates' precarious motion. One skate tilts precariously inward, its outer wheel lifting just a millimeter off the pavement before correcting with a clumsy weight shift, passing the jolt to the other skate. This precarious back-and-forth, the only movement in the frame, creates a concentrated tension. The entire scene is rendered with the rich, cinematic quality of 35mm film, characterized by delicate, organic textures, deep color saturation, and a gentle halo around the sunlit edges of the skates, evoking a nostalgic and slightly anxious atmosphere of precarious equilibrium.", + "seed": 18518, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_8/index_5441_seed_18518_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": -0.007989911735057831, + "vlm_logprob": -0.007989911735057831, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.007989911735057831, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007128519471734762 + ], + "opposite_logprob_tokens": [ + -5.382128715515137 + ], + "logprob_reward": -0.007128519471734762, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007128519471734762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0763869509100914, + 0.050823844969272614, + 0.1190812960267067, + 0.17896346747875214, + 0.05916419252753258, + 0.2610374987125397, + 0.12622609734535217, + 0.057110272347927094, + 0.0241762213408947, + 0.047030143439769745 + ], + "credit_map": [ + 0.008120061829686165, + 0.013657321222126484, + 0.009967642836272717, + 0.00846647098660469, + 0.009657308459281921, + 0.009112751111388206, + 0.007378702517598867, + 0.010026693344116211, + 0.004657587502151728, + 0.0083349933847785, + 0.006670456845313311, + 0.006165145430713892, + 0.0058953114785254, + 0.004337442107498646, + 0.008483221754431725, + 0.006279690656810999, + 0.014977476559579372, + 0.019703468307852745, + 0.019281502813100815, + 0.013754298910498619, + 0.013052099384367466, + 0.016393082216382027, + 0.009372233413159847, + 0.012547132559120655, + 0.015382439829409122, + 0.02928127907216549, + 0.017660120502114296, + 0.018949519842863083, + 0.037080325186252594, + 0.017996765673160553, + 0.021068992093205452, + 0.021544013172388077, + 0.005576528143137693, + 0.006294111721217632, + 0.009014100767672062, + 0.008728144690394402, + 0.006247228942811489, + 0.005442440044134855, + 0.010737835429608822, + 0.00712380139157176, + 0.012130575254559517, + 0.00831762608140707, + 0.017464976757764816, + 0.013796348124742508, + 0.01629064418375492, + 0.08600501716136932, + 0.01950136199593544, + 0.0875309482216835, + 0.01145210675895214, + 0.013939891941845417, + 0.014057601802051067, + 0.017707450315356255, + 0.022031588479876518, + 0.019862227141857147, + 0.016469573602080345, + 0.010705661959946156, + 0.006769377738237381, + 0.007176883518695831, + 0.006565352436155081, + 0.003914165776222944, + 0.00753194373100996, + 0.013765311799943447, + 0.004515276290476322, + 0.006871960125863552, + 0.0026412929873913527, + 0.0030482644215226173, + 0.002735550981014967, + 0.0031821923330426216, + 0.003148292424157262, + 0.0029865228570997715, + 0.003223118605092168, + 0.0032109871972352266, + 0.005072855390608311, + 0.005281984340399504, + 0.006865097675472498, + 0.00514213414862752, + 0.0075130765326321125, + 0.007178570609539747, + 0.004411166533827782, + 0.00556525494903326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.007128519471734762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256166215986013 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256166215986013, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256166215986013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08813875913619995, + 0.10540267825126648, + 0.07171276956796646, + 0.21459601819515228, + 0.14030496776103973, + 0.06762441992759705, + 0.07863301783800125, + 0.04811611399054527, + 0.06530582904815674, + 0.1201654002070427 + ], + "credit_map": [ + 0.008971352130174637, + 0.010044842027127743, + 0.010679476894438267, + 0.011689682491123676, + 0.011470191180706024, + 0.01272562611848116, + 0.0095480065792799, + 0.01300958264619112, + 0.01245964877307415, + 0.01456773467361927, + 0.012263703159987926, + 0.012362928129732609, + 0.01735989935696125, + 0.011974341236054897, + 0.011885369196534157, + 0.012529052793979645, + 0.006084634456783533, + 0.009620635770261288, + 0.009470908902585506, + 0.011429439298808575, + 0.010971738956868649, + 0.0077885957434773445, + 0.00846010446548462, + 0.007886714302003384, + 0.026608191430568695, + 0.03236422687768936, + 0.03727314993739128, + 0.01956234499812126, + 0.03335992991924286, + 0.025186043232679367, + 0.017843006178736687, + 0.02239915169775486, + 0.014962724409997463, + 0.017486365512013435, + 0.021338121965527534, + 0.01989935152232647, + 0.016256926581263542, + 0.014499588869512081, + 0.017265165224671364, + 0.01859673298895359, + 0.006214227993041277, + 0.007678763009607792, + 0.009795689024031162, + 0.009501739405095577, + 0.00819452852010727, + 0.007783250883221626, + 0.008213932625949383, + 0.010242286138236523, + 0.014627782627940178, + 0.015286071226000786, + 0.007589360699057579, + 0.009958990849554539, + 0.008190548047423363, + 0.011737716384232044, + 0.0059401607140898705, + 0.005302388221025467, + 0.005586054641753435, + 0.007133485749363899, + 0.005624816287308931, + 0.0034328310284763575, + 0.007059051655232906, + 0.009489180520176888, + 0.004062889609485865, + 0.00572780380025506, + 0.006346451118588448, + 0.01030738465487957, + 0.007837992161512375, + 0.00737624429166317, + 0.011150551028549671, + 0.007678445428609848, + 0.006712974049150944, + 0.00789578165858984, + 0.010125520639121532, + 0.011799889616668224, + 0.021752560511231422, + 0.01329081878066063, + 0.016898665577173233, + 0.01806694082915783, + 0.012190153822302818, + 0.016040842980146408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.005256166215986013 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253464724868536 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.01253464724868536, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253464724868536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10698483139276505, + 0.08244766294956207, + 0.1291343718767166, + 0.173102468252182, + 0.15107987821102142, + 0.1256881207227707, + 0.05293026566505432, + 0.05175882205367088, + 0.05959257110953331, + 0.06728101521730423 + ], + "credit_map": [ + 0.01250453945249319, + 0.010929131880402565, + 0.012465963140130043, + 0.012659238651394844, + 0.01713033765554428, + 0.016125639900565147, + 0.010903705842792988, + 0.014266283251345158, + 0.009400124661624432, + 0.012630072422325611, + 0.010250508785247803, + 0.010374508798122406, + 0.010754603892564774, + 0.008049278520047665, + 0.009234428405761719, + 0.011754141189157963, + 0.011109702289104462, + 0.012897768057882786, + 0.018829533830285072, + 0.02096329815685749, + 0.015532443299889565, + 0.019640576094388962, + 0.014130615629255772, + 0.016030441969633102, + 0.02019062452018261, + 0.023048220202326775, + 0.024769870564341545, + 0.020704062655568123, + 0.021274100989103317, + 0.019990727305412292, + 0.016597487032413483, + 0.026527371257543564, + 0.015062362886965275, + 0.019123103469610214, + 0.02449563518166542, + 0.022175133228302002, + 0.01463742833584547, + 0.013502293266355991, + 0.024199994280934334, + 0.017883922904729843, + 0.005950108636170626, + 0.007093831431120634, + 0.010825219564139843, + 0.0109608368948102, + 0.006369589827954769, + 0.03635482117533684, + 0.00878303311765194, + 0.039350684732198715, + 0.00607378501445055, + 0.010244262404739857, + 0.004946170374751091, + 0.007530747912824154, + 0.005916745401918888, + 0.00890151597559452, + 0.00455571198835969, + 0.004761329386383295, + 0.005343405529856682, + 0.006644552573561668, + 0.006410606205463409, + 0.003862053155899048, + 0.004937527701258659, + 0.013918722048401833, + 0.00544297369197011, + 0.005198980215936899, + 0.0058800517581403255, + 0.006871855352073908, + 0.006902774795889854, + 0.009412533603608608, + 0.006226053461432457, + 0.006913769990205765, + 0.009048270992934704, + 0.008337261155247688, + 0.006216105073690414, + 0.00520330062136054, + 0.01346898265182972, + 0.008517051115632057, + 0.005948523059487343, + 0.010395676828920841, + 0.006854349747300148, + 0.010677028447389603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.01253464724868536 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780462831258774 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780462831258774, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780462831258774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12414245307445526, + 0.09243571758270264, + 0.1796363741159439, + 0.12374959141016006, + 0.10733886063098907, + 0.13480006158351898, + 0.07585722208023071, + 0.07460285723209381, + 0.027296479791402817, + 0.06014038249850273 + ], + "credit_map": [ + 0.011656525544822216, + 0.02438451535999775, + 0.01415960118174553, + 0.014782208018004894, + 0.013409764505922794, + 0.012981175445020199, + 0.013780530542135239, + 0.01898813620209694, + 0.009276414290070534, + 0.013954035937786102, + 0.01449897512793541, + 0.0125211700797081, + 0.01160570327192545, + 0.00943665485829115, + 0.008477604947984219, + 0.012665153481066227, + 0.02092251181602478, + 0.025060510262846947, + 0.027044521644711494, + 0.019935835152864456, + 0.021358242258429527, + 0.03414476662874222, + 0.012881387956440449, + 0.01828859932720661, + 0.010323390364646912, + 0.01717091165482998, + 0.01648818887770176, + 0.012914645485579967, + 0.024287812411785126, + 0.013590929098427296, + 0.010688244365155697, + 0.018285464495420456, + 0.010337294079363346, + 0.01648111268877983, + 0.016244616359472275, + 0.014819128438830376, + 0.010451958514750004, + 0.01134954672306776, + 0.017292527481913567, + 0.010362675413489342, + 0.011913716793060303, + 0.010023808106780052, + 0.015099750831723213, + 0.012351389974355698, + 0.018521655350923538, + 0.02724827080965042, + 0.013471673242747784, + 0.02616981230676174, + 0.008133689872920513, + 0.009405620396137238, + 0.009262824431061745, + 0.00859154760837555, + 0.012016086839139462, + 0.013449128717184067, + 0.008297308348119259, + 0.0067010181955993176, + 0.007180779706686735, + 0.010957188904285431, + 0.007815377786755562, + 0.003160949796438217, + 0.005667442921549082, + 0.02785307727754116, + 0.0049079302698373795, + 0.007060105912387371, + 0.0030884372536092997, + 0.0036062048748135567, + 0.003040107898414135, + 0.003776976140215993, + 0.002819198416545987, + 0.0038127636071294546, + 0.002919109771028161, + 0.004233682062476873, + 0.0058425962924957275, + 0.006560339592397213, + 0.010537570342421532, + 0.007487208116799593, + 0.005979153793305159, + 0.007897801697254181, + 0.007226252928376198, + 0.008609458804130554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.009780462831258774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249762907624245 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249762907624245, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249762907624245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08426009863615036, + 0.06587573140859604, + 0.32302042841911316, + 0.07091066986322403, + 0.08297121524810791, + 0.1033845990896225, + 0.07698515802621841, + 0.043580152094364166, + 0.019427470862865448, + 0.1295844465494156 + ], + "credit_map": [ + 0.008034780621528625, + 0.012689365074038506, + 0.0167921781539917, + 0.01078435406088829, + 0.007530291564762592, + 0.009823157452046871, + 0.008100936189293861, + 0.01050503458827734, + 0.009920943528413773, + 0.0094824880361557, + 0.0095515176653862, + 0.008196874521672726, + 0.006531013175845146, + 0.006779819261282682, + 0.006735172588378191, + 0.008677906356751919, + 0.035829655826091766, + 0.046228792518377304, + 0.05271094664931297, + 0.029870951548218727, + 0.032545704394578934, + 0.07017026096582413, + 0.02293025329709053, + 0.032733868807554245, + 0.006013842765241861, + 0.00859921332448721, + 0.010949837043881416, + 0.009470975957810879, + 0.012027031742036343, + 0.009101998060941696, + 0.006467747502028942, + 0.00828002393245697, + 0.006061367690563202, + 0.008709355257451534, + 0.012268047779798508, + 0.008951590396463871, + 0.010402338579297066, + 0.008326202630996704, + 0.010674181394279003, + 0.017578132450580597, + 0.0069874003529548645, + 0.007922017946839333, + 0.010225847363471985, + 0.008925299160182476, + 0.010082737542688847, + 0.02795901522040367, + 0.008292431943118572, + 0.022989854216575623, + 0.012297808192670345, + 0.012887291610240936, + 0.008944101631641388, + 0.010609779506921768, + 0.00990411639213562, + 0.009243002161383629, + 0.007803678046911955, + 0.005295380484312773, + 0.0042923083528876305, + 0.006762581877410412, + 0.004502002149820328, + 0.0023118385579437017, + 0.004272552207112312, + 0.013646452687680721, + 0.0026456790510565042, + 0.005146737210452557, + 0.0021795162465423346, + 0.002777488436549902, + 0.002448789542540908, + 0.0019517112523317337, + 0.0031898017041385174, + 0.002785921096801758, + 0.0018331753090023994, + 0.002261067507788539, + 0.013535094447433949, + 0.009525423869490623, + 0.024906057864427567, + 0.01534082181751728, + 0.00950893945991993, + 0.016746748238801956, + 0.01536658313125372, + 0.02465478517115116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.005249762907624245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a pair of vintage four-wheeled roller skates swaying hesitantly across a rough, dark-gray asphalt surface. The skates are clad in worn, seaweed-green leather boots, their laces slightly frayed. Their polyurethane wheels are a milky white, speckled with the grit from the ground. Warm, golden sunlight bathes the scene in late afternoon light, capturing the dull gleam of the aluminum panels and casting long, soft shadows that accentuate the skates' precarious motion. One skate tilts precariously inward, its outer wheel lifting just a millimeter off the pavement before correcting with a clumsy weight shift, passing the jolt to the other skate. This precarious back-and-forth, the only movement in the frame, creates a concentrated tension. The entire scene is rendered with the rich, cinematic quality of 35mm film, characterized by delicate, organic textures, deep color saturation, and a gentle halo around the sunlit edges of the skates, evoking a nostalgic and slightly anxious atmosphere of precarious equilibrium.", + "seed": 18519, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_8/index_5441_seed_18519_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": -0.012983362097293139, + "vlm_logprob": -0.012983362097293139, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.012983362097293139, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022809520363807678 + ], + "opposite_logprob_tokens": [ + -4.1478095054626465 + ], + "logprob_reward": -0.022809520363807678, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022809520363807678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09002070873975754, + 0.12861043214797974, + 0.05406197905540466, + 0.38640427589416504, + 0.05954320728778839, + 0.03770766407251358, + 0.052574627101421356, + 0.09892498701810837, + 0.0581522211432457, + 0.03399994224309921 + ], + "credit_map": [ + 0.013454529456794262, + 0.009193506091833115, + 0.02187715843319893, + 0.010535983368754387, + 0.007117272354662418, + 0.008484233170747757, + 0.00934215635061264, + 0.010015874169766903, + 0.017938928678631783, + 0.012049839831888676, + 0.02314528450369835, + 0.015468341298401356, + 0.015812600031495094, + 0.01117526926100254, + 0.01787232980132103, + 0.015147832222282887, + 0.008061925880610943, + 0.005883591249585152, + 0.005446095485240221, + 0.006602826528251171, + 0.006841481197625399, + 0.005333362612873316, + 0.009568607434630394, + 0.006324089132249355, + 0.04749288037419319, + 0.04146489500999451, + 0.03933120146393776, + 0.04926443099975586, + 0.034221503883600235, + 0.042952995747327805, + 0.04582919552922249, + 0.0858471542596817, + 0.007128974888473749, + 0.006199368741363287, + 0.005479593761265278, + 0.005866348277777433, + 0.012942520901560783, + 0.005507017485797405, + 0.010824736207723618, + 0.0055946470238268375, + 0.004090691916644573, + 0.004845160059630871, + 0.006323748268187046, + 0.004009416792541742, + 0.006495086010545492, + 0.0036313009914010763, + 0.0037726510781794786, + 0.004539611749351025, + 0.008098196238279343, + 0.004288749303668737, + 0.006712861359119415, + 0.009541790932416916, + 0.004853630438446999, + 0.005316851194947958, + 0.0088810995221138, + 0.00488144950941205, + 0.01665319688618183, + 0.011981864459812641, + 0.012453003786504269, + 0.014971420168876648, + 0.009560334496200085, + 0.009969453327357769, + 0.013064530678093433, + 0.010271179489791393, + 0.016431935131549835, + 0.00346956355497241, + 0.017032282426953316, + 0.004140614997595549, + 0.0028840084560215473, + 0.005359856877475977, + 0.005016258452087641, + 0.0038177003152668476, + 0.005949034355580807, + 0.0033562099561095238, + 0.004600957036018372, + 0.003975875675678253, + 0.004180520307272673, + 0.003588959341868758, + 0.004749313462525606, + 0.003599072340875864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.022809520363807678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745188031345606 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745188031345606, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745188031345606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16401289403438568, + 0.09479722380638123, + 0.07642094790935516, + 0.11555274575948715, + 0.07629863917827606, + 0.08069638162851334, + 0.07488303631544113, + 0.17608164250850677, + 0.09342151135206223, + 0.04783498868346214 + ], + "credit_map": [ + 0.021510649472475052, + 0.02014864794909954, + 0.027555517852306366, + 0.018252462148666382, + 0.031545285135507584, + 0.013428760692477226, + 0.016059530898928642, + 0.015512031503021717, + 0.013290607370436192, + 0.010495045222342014, + 0.019735412672162056, + 0.011127768084406853, + 0.009737585671246052, + 0.008710219524800777, + 0.01196577399969101, + 0.009734815917909145, + 0.009820342063903809, + 0.010559910908341408, + 0.00980569887906313, + 0.0070563112385571, + 0.013840817846357822, + 0.007721938658505678, + 0.008972174488008022, + 0.00864375103265047, + 0.013676750473678112, + 0.017854049801826477, + 0.009494180791079998, + 0.0108256246894598, + 0.021757710725069046, + 0.011541519314050674, + 0.014699989929795265, + 0.01570291817188263, + 0.010453782044351101, + 0.006709842476993799, + 0.01182149350643158, + 0.008173393085598946, + 0.009321711026132107, + 0.006345825269818306, + 0.013311677612364292, + 0.010160909965634346, + 0.009744315408170223, + 0.00819153431802988, + 0.009553235955536366, + 0.007197186816483736, + 0.011150472797453403, + 0.007479310967028141, + 0.011399495415389538, + 0.015980832278728485, + 0.01157841645181179, + 0.008114940486848354, + 0.009736952371895313, + 0.017834903672337532, + 0.007164821494370699, + 0.008272754959762096, + 0.0062501090578734875, + 0.005930136423557997, + 0.04065482318401337, + 0.013241967186331749, + 0.0535980686545372, + 0.019904373213648796, + 0.009778296574950218, + 0.017951495945453644, + 0.009979354217648506, + 0.010973269119858742, + 0.02507028542459011, + 0.005818625446408987, + 0.024787992238998413, + 0.005186849273741245, + 0.006526297423988581, + 0.012814040295779705, + 0.007172515615820885, + 0.006044907029718161, + 0.007376554422080517, + 0.006881613750010729, + 0.006144324783235788, + 0.006655600853264332, + 0.004412158392369747, + 0.005380687769502401, + 0.005707986652851105, + 0.005276064854115248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.006745188031345606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184728920459747 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184728920459747, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184728920459747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19257472455501556, + 0.09705138206481934, + 0.06569826602935791, + 0.19242922961711884, + 0.13075599074363708, + 0.13652759790420532, + 0.04305871203541756, + 0.06241524592041969, + 0.036524150520563126, + 0.04296465218067169 + ], + "credit_map": [ + 0.0295578483492136, + 0.021552272140979767, + 0.027386728674173355, + 0.016446569934487343, + 0.043995410203933716, + 0.01578802801668644, + 0.022417638450860977, + 0.015430223196744919, + 0.012071814388036728, + 0.010053400881588459, + 0.011954708956182003, + 0.016064796596765518, + 0.010876219719648361, + 0.009064839221537113, + 0.015461526811122894, + 0.011504082009196281, + 0.009162607602775097, + 0.0053568435832858086, + 0.010289683006703854, + 0.00787628535181284, + 0.006803888827562332, + 0.006324669346213341, + 0.01113632507622242, + 0.008747962303459644, + 0.023313894867897034, + 0.02823518216609955, + 0.01896257884800434, + 0.03164547681808472, + 0.02680995687842369, + 0.018392013385891914, + 0.025992123410105705, + 0.019078003242611885, + 0.018216613680124283, + 0.010233367793262005, + 0.026517968624830246, + 0.013677560724318027, + 0.014677947387099266, + 0.010577921755611897, + 0.023373711854219437, + 0.013480893336236477, + 0.01631954312324524, + 0.014315513893961906, + 0.01694393716752529, + 0.011012010276317596, + 0.015607161447405815, + 0.014028084464371204, + 0.014485752210021019, + 0.03381558880209923, + 0.0062775020487606525, + 0.003859439166262746, + 0.004663311410695314, + 0.00395736750215292, + 0.00588475726544857, + 0.004380266647785902, + 0.01031701359897852, + 0.0037190550938248634, + 0.011966017074882984, + 0.006997218355536461, + 0.0143890967592597, + 0.010655166581273079, + 0.004556823056191206, + 0.005067629273980856, + 0.004942748695611954, + 0.003840548451989889, + 0.0054743122309446335, + 0.003480365499854088, + 0.004223999101668596, + 0.002613987075164914, + 0.0032057848293334246, + 0.006323724985122681, + 0.004915768746286631, + 0.006286207586526871, + 0.007411424070596695, + 0.004652488976716995, + 0.005543867591768503, + 0.006200830917805433, + 0.003793783485889435, + 0.006271137855947018, + 0.004547912161797285, + 0.00454320851713419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.018184728920459747 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541239149868488 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541239149868488, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541239149868488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13028517365455627, + 0.07873156666755676, + 0.09930679202079773, + 0.05158953741192818, + 0.0898825004696846, + 0.11328130960464478, + 0.07145102322101593, + 0.08319563418626785, + 0.10259800404310226, + 0.17967846989631653 + ], + "credit_map": [ + 0.01940881460905075, + 0.014282779768109322, + 0.023332133889198303, + 0.01792372763156891, + 0.011788753792643547, + 0.012139611877501011, + 0.016438426449894905, + 0.014970937743782997, + 0.008889260701835155, + 0.008851714432239532, + 0.012817754410207272, + 0.014708898961544037, + 0.006043076049536467, + 0.006988941226154566, + 0.011479195207357407, + 0.008952734060585499, + 0.012541250325739384, + 0.010394257493317127, + 0.01347158569842577, + 0.009919261559844017, + 0.013727929443120956, + 0.009623226709663868, + 0.017522109672427177, + 0.012107165530323982, + 0.008635585196316242, + 0.004640616942197084, + 0.005703306291252375, + 0.006651140283793211, + 0.006841811817139387, + 0.004678234923630953, + 0.007945604622364044, + 0.006493234541267157, + 0.010265606455504894, + 0.015848280861973763, + 0.013325338251888752, + 0.008844047784805298, + 0.011899166740477085, + 0.008032347075641155, + 0.012746547348797321, + 0.008921165019273758, + 0.016858968883752823, + 0.009777815081179142, + 0.013455572538077831, + 0.0106970788910985, + 0.009662085212767124, + 0.012301533482968807, + 0.011758215725421906, + 0.028770027682185173, + 0.010398691520094872, + 0.009510086849331856, + 0.006595855113118887, + 0.005294003523886204, + 0.01367437094449997, + 0.009090646170079708, + 0.01067136786878109, + 0.006216003559529781, + 0.01744275912642479, + 0.00959730613976717, + 0.017164435237646103, + 0.009418650530278683, + 0.007449251599609852, + 0.009301131591200829, + 0.006934028584510088, + 0.005888066720217466, + 0.021962063387036324, + 0.00991001259535551, + 0.019309019669890404, + 0.009259093552827835, + 0.007319329772144556, + 0.010590864345431328, + 0.011089508421719074, + 0.01315811276435852, + 0.03247803449630737, + 0.01773969829082489, + 0.0275490153580904, + 0.020753920078277588, + 0.018572423607110977, + 0.020204594358801842, + 0.0269309189170599, + 0.015449857339262962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.012541239149868488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636134020984173 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.004636134020984173, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636134020984173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.234939306974411, + 0.05333726480603218, + 0.08286914229393005, + 0.03997092321515083, + 0.09800822287797928, + 0.31262707710266113, + 0.02274395525455475, + 0.06928432732820511, + 0.06058405712246895, + 0.025635745376348495 + ], + "credit_map": [ + 0.04855455830693245, + 0.02181348204612732, + 0.04676491394639015, + 0.026581572368741035, + 0.02563115768134594, + 0.024174565449357033, + 0.023589663207530975, + 0.017829377204179764, + 0.008913235738873482, + 0.0048332177102565765, + 0.01229801494628191, + 0.005782297346740961, + 0.0048951925709843636, + 0.005560247227549553, + 0.00547032942995429, + 0.005584727972745895, + 0.012100319378077984, + 0.007570210844278336, + 0.012478476390242577, + 0.010548878461122513, + 0.010870755650103092, + 0.006969403941184282, + 0.012551776133477688, + 0.009779314510524273, + 0.005998942535370588, + 0.006586264353245497, + 0.0035757520236074924, + 0.0035286385100334883, + 0.006269208621233702, + 0.00464217271655798, + 0.0050073545426130295, + 0.004362589679658413, + 0.01296378392726183, + 0.011647799983620644, + 0.015158289112150669, + 0.012380055151879787, + 0.011015038006007671, + 0.008875615894794464, + 0.014427147805690765, + 0.011540487408638, + 0.040414050221443176, + 0.030531056225299835, + 0.038833487778902054, + 0.029625725001096725, + 0.03310514986515045, + 0.030763771384954453, + 0.03718390315771103, + 0.07216989994049072, + 0.0031842400785535574, + 0.00305765843950212, + 0.0022727574687451124, + 0.002337300917133689, + 0.0024352495092898607, + 0.0030659798067063093, + 0.003457402577623725, + 0.0029333638958632946, + 0.017044810578227043, + 0.004894444718956947, + 0.02176499180495739, + 0.00793273001909256, + 0.003513679141178727, + 0.00590166263282299, + 0.004138567950576544, + 0.00409343047067523, + 0.019896287471055984, + 0.002375486074015498, + 0.021116608753800392, + 0.002704976825043559, + 0.0021989091765135527, + 0.0036920260172337294, + 0.0038548961747437716, + 0.004744859412312508, + 0.003922432195395231, + 0.0028814030811190605, + 0.0033364538103342056, + 0.0037595692556351423, + 0.002334773074835539, + 0.003424539929255843, + 0.003326421370729804, + 0.002650150330737233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.004636134020984173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a pair of vintage four-wheeled roller skates swaying hesitantly across a rough, dark-gray asphalt surface. The skates are clad in worn, seaweed-green leather boots, their laces slightly frayed. Their polyurethane wheels are a milky white, speckled with the grit from the ground. Warm, golden sunlight bathes the scene in late afternoon light, capturing the dull gleam of the aluminum panels and casting long, soft shadows that accentuate the skates' precarious motion. One skate tilts precariously inward, its outer wheel lifting just a millimeter off the pavement before correcting with a clumsy weight shift, passing the jolt to the other skate. This precarious back-and-forth, the only movement in the frame, creates a concentrated tension. The entire scene is rendered with the rich, cinematic quality of 35mm film, characterized by delicate, organic textures, deep color saturation, and a gentle halo around the sunlit edges of the skates, evoking a nostalgic and slightly anxious atmosphere of precarious equilibrium.", + "seed": 18520, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_8/index_5441_seed_18520_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": -0.017030030488967896, + "vlm_logprob": -0.017030030488967896, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.017030030488967896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01825636811554432 + ], + "opposite_logprob_tokens": [ + -4.143256187438965 + ], + "logprob_reward": -0.01825636811554432, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01825636811554432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08252807706594467, + 0.09704232960939407, + 0.06611375510692596, + 0.07377379387617111, + 0.12259048223495483, + 0.13995587825775146, + 0.029543790966272354, + 0.3029612600803375, + 0.06396433711051941, + 0.021526282653212547 + ], + "credit_map": [ + 0.015400068834424019, + 0.006833061575889587, + 0.013911845162510872, + 0.019754083827137947, + 0.0033905119635164738, + 0.004985813517123461, + 0.008187168277800083, + 0.010065517388284206, + 0.009600932709872723, + 0.0070334551855921745, + 0.018876517191529274, + 0.010086610913276672, + 0.016970522701740265, + 0.014373970218002796, + 0.008070241659879684, + 0.012030069716274738, + 0.007674571126699448, + 0.005904246121644974, + 0.01237042061984539, + 0.007479683496057987, + 0.006840302608907223, + 0.012892665341496468, + 0.006121722515672445, + 0.006830138154327869, + 0.020311083644628525, + 0.004642749670892954, + 0.022117286920547485, + 0.011987066827714443, + 0.004333634860813618, + 0.003163701854646206, + 0.003809755202382803, + 0.003408508375287056, + 0.015351255424320698, + 0.012437161058187485, + 0.015557175502181053, + 0.020952802151441574, + 0.01827007532119751, + 0.01218064408749342, + 0.011602058075368404, + 0.016239294782280922, + 0.012177594937384129, + 0.012113072909414768, + 0.02657483145594597, + 0.021676918491721153, + 0.013488185591995716, + 0.013940537348389626, + 0.022089621052145958, + 0.017895113676786423, + 0.0023226477205753326, + 0.001879328046925366, + 0.004532919730991125, + 0.003176763653755188, + 0.003758716629818082, + 0.0060275341384112835, + 0.0039305840618908405, + 0.003915294539183378, + 0.03204752877354622, + 0.04926632344722748, + 0.057226262986660004, + 0.04522942751646042, + 0.03233564645051956, + 0.027412686496973038, + 0.02946232445538044, + 0.029981035739183426, + 0.006765422876924276, + 0.004588683135807514, + 0.01603611558675766, + 0.004952861461788416, + 0.0030455540399998426, + 0.008959145285189152, + 0.006911450065672398, + 0.012705100700259209, + 0.003968383651226759, + 0.0020898275543004274, + 0.0027923656161874533, + 0.0020346292294561863, + 0.0018759331433102489, + 0.002443309174850583, + 0.003633260726928711, + 0.002688572509214282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.01825636811554432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642767556011677 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642767556011677, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642767556011677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11988893896341324, + 0.07400377094745636, + 0.1879432499408722, + 0.04246224835515022, + 0.027330169454216957, + 0.08776751160621643, + 0.05022331699728966, + 0.0939328521490097, + 0.28083890676498413, + 0.035609059035778046 + ], + "credit_map": [ + 0.023562537506222725, + 0.010287064127624035, + 0.016746852546930313, + 0.0225799772888422, + 0.010413577780127525, + 0.006797037087380886, + 0.015287063084542751, + 0.014214824885129929, + 0.006514264270663261, + 0.007378743030130863, + 0.009784833528101444, + 0.010373633354902267, + 0.01129611860960722, + 0.010740927420556545, + 0.011016697622835636, + 0.006898551248013973, + 0.025320447981357574, + 0.01613626815378666, + 0.041411712765693665, + 0.019250502809882164, + 0.017805295065045357, + 0.03339148312807083, + 0.01664366014301777, + 0.01798388548195362, + 0.010468271560966969, + 0.0033086389303207397, + 0.010691171512007713, + 0.006588982883840799, + 0.00285917054861784, + 0.0032859169878065586, + 0.002655492164194584, + 0.0026046033017337322, + 0.0025847505312412977, + 0.0015889741480350494, + 0.00643759174272418, + 0.0027565660420805216, + 0.004032223019748926, + 0.003414169419556856, + 0.003483364824205637, + 0.0030325318221002817, + 0.007570276502519846, + 0.010054903104901314, + 0.015060944482684135, + 0.013187972828745842, + 0.009936208836734295, + 0.01009250059723854, + 0.013003560714423656, + 0.008861146867275238, + 0.003725941525772214, + 0.003850162262097001, + 0.007471390999853611, + 0.0057481625117361546, + 0.006575210951268673, + 0.0120460893958807, + 0.003848427440971136, + 0.006957930512726307, + 0.011207877658307552, + 0.01466158777475357, + 0.019649669528007507, + 0.01230962947010994, + 0.007805778179317713, + 0.007159609813243151, + 0.011040755547583103, + 0.010097944177687168, + 0.02359617128968239, + 0.009264837019145489, + 0.09461008012294769, + 0.020254701375961304, + 0.010432863608002663, + 0.037474825978279114, + 0.01622103713452816, + 0.06898438930511475, + 0.004986943677067757, + 0.002957869553938508, + 0.00804983638226986, + 0.0031815075781196356, + 0.003966828808188438, + 0.00458809919655323, + 0.003846710314974189, + 0.0040312642231583595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.008642767556011677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059687361121178 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.016059687361121178, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059687361121178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08016624301671982, + 0.030686482787132263, + 0.06787954270839691, + 0.021562812849879265, + 0.026303833350539207, + 0.038264546543359756, + 0.03212139010429382, + 0.11835446208715439, + 0.5503935813903809, + 0.03426703065633774 + ], + "credit_map": [ + 0.018081847578287125, + 0.005757944192737341, + 0.015593556687235832, + 0.018227573484182358, + 0.005840419791638851, + 0.005203451029956341, + 0.0056386590003967285, + 0.0058227949775755405, + 0.0030268242117017508, + 0.002570060547441244, + 0.003996924962848425, + 0.003314607311040163, + 0.0057382723316550255, + 0.006440095603466034, + 0.0031484949868172407, + 0.002451204461976886, + 0.007215470541268587, + 0.0063305506482720375, + 0.012551761232316494, + 0.00785953551530838, + 0.006295901723206043, + 0.012737218290567398, + 0.0071471394039690495, + 0.007741968147456646, + 0.003915597219020128, + 0.002053184900432825, + 0.0029122894629836082, + 0.0025968707632273436, + 0.0020580871496349573, + 0.0031472332775592804, + 0.0029858427587896585, + 0.0018937104614451528, + 0.0034331989008933306, + 0.00220599421299994, + 0.004531446378678083, + 0.0036704351659864187, + 0.002915114164352417, + 0.003252319060266018, + 0.0027007556054741144, + 0.0035945731215178967, + 0.0030286889523267746, + 0.006191720254719257, + 0.00515300827100873, + 0.0056225452572107315, + 0.005209058988839388, + 0.0038340017199516296, + 0.005412967875599861, + 0.0038125566206872463, + 0.0026705702766776085, + 0.0018269637366756797, + 0.004744141828268766, + 0.004443378187716007, + 0.004958020523190498, + 0.0068008266389369965, + 0.0030389202293008566, + 0.003638571361079812, + 0.011929982341825962, + 0.020175565034151077, + 0.02561834827065468, + 0.015933139249682426, + 0.012622666545212269, + 0.009452614933252335, + 0.010760212317109108, + 0.011861941777169704, + 0.044507306069135666, + 0.014174711890518665, + 0.19609691202640533, + 0.03654243052005768, + 0.019552363082766533, + 0.06439992785453796, + 0.025320546701550484, + 0.1497994214296341, + 0.002680768957361579, + 0.002621921245008707, + 0.007460004184395075, + 0.002559773623943329, + 0.004857153166085482, + 0.003920602612197399, + 0.0038687328342348337, + 0.006298077292740345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.016059687361121178 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809867799282074 + ], + "opposite_logprob_tokens": [ + -3.2880985736846924 + ], + "logprob_reward": -0.03809867799282074, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809867799282074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13175033032894135, + 0.2156873643398285, + 0.06636285781860352, + 0.1851225197315216, + 0.030674772337079048, + 0.15222755074501038, + 0.03801270201802254, + 0.08421886712312698, + 0.0570707768201828, + 0.03887225687503815 + ], + "credit_map": [ + 0.029798336327075958, + 0.01042113732546568, + 0.023599131032824516, + 0.013149631209671497, + 0.012288949452340603, + 0.007978157140314579, + 0.014509720727801323, + 0.02000526152551174, + 0.017444180324673653, + 0.014408444985747337, + 0.031881753355264664, + 0.029478125274181366, + 0.03311178833246231, + 0.03230352699756622, + 0.028236020356416702, + 0.028823526576161385, + 0.008331059478223324, + 0.00627396535128355, + 0.01089590322226286, + 0.006465986371040344, + 0.007245489861816168, + 0.011730111204087734, + 0.008668455295264721, + 0.006751887500286102, + 0.04055943712592125, + 0.01759866438806057, + 0.046404365450143814, + 0.02660105749964714, + 0.013950175605714321, + 0.014265306293964386, + 0.013327756896615028, + 0.012415764853358269, + 0.003407695796340704, + 0.0017680564196780324, + 0.004017523489892483, + 0.003583214245736599, + 0.0034230968449264765, + 0.004758767317980528, + 0.0036563382018357515, + 0.006060081068426371, + 0.014559635892510414, + 0.013250518590211868, + 0.026062918826937675, + 0.023617936298251152, + 0.01795257441699505, + 0.017439555376768112, + 0.021472398191690445, + 0.017872007563710213, + 0.0030681395437568426, + 0.0025434112176299095, + 0.007369492202997208, + 0.006476573646068573, + 0.004336287267506123, + 0.004476230591535568, + 0.004646472632884979, + 0.005096094682812691, + 0.008703706786036491, + 0.014366165734827518, + 0.013933788985013962, + 0.012187250889837742, + 0.009327908977866173, + 0.007691211998462677, + 0.009248930960893631, + 0.008759901858866215, + 0.005422270391136408, + 0.0071520693600177765, + 0.00802706927061081, + 0.005811008624732494, + 0.0054063936695456505, + 0.012706278823316097, + 0.0066452911123633385, + 0.005900397431105375, + 0.004018857609480619, + 0.003917803056538105, + 0.00717940554022789, + 0.0032470119185745716, + 0.005589089822024107, + 0.004447409883141518, + 0.004967402201145887, + 0.005505278240889311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.03809867799282074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a pair of vintage four-wheeled roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skates clad in worn, seaweed-green leather boots with slightly frayed laces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the polyurethane wheels milky white and speckled with grit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature warm, golden sunlight casting long, soft shadows on a rough, dark-gray asphalt surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30767959356307983, + 0.11795423179864883, + 0.03588297590613365, + 0.018358413130044937, + 0.014013382606208324, + 0.0327814556658268, + 0.056177616119384766, + 0.06664307415485382, + 0.30007150769233704, + 0.05043771490454674 + ], + "credit_map": [ + 0.05041133984923363, + 0.0249341931194067, + 0.05143354460597038, + 0.06670207530260086, + 0.0200601015239954, + 0.023239953443408012, + 0.0322389230132103, + 0.038659486919641495, + 0.007401552516967058, + 0.00392443872988224, + 0.012950586155056953, + 0.008952293545007706, + 0.02185206301510334, + 0.02413984201848507, + 0.018658606335520744, + 0.02007487043738365, + 0.0056187063455581665, + 0.0022054605651646852, + 0.005987260024994612, + 0.0028370607178658247, + 0.0035269774962216616, + 0.006229399237781763, + 0.003855963470414281, + 0.0056221517734229565, + 0.002505493350327015, + 0.003079489339143038, + 0.002746326383203268, + 0.0033773647155612707, + 0.0017761442577466369, + 0.0013854136923328042, + 0.0017899781232699752, + 0.0016982047818601131, + 0.0012486319756135345, + 0.0009274767362512648, + 0.0022188120055943727, + 0.0019021909683942795, + 0.0019007428782060742, + 0.0017909938469529152, + 0.001623671967536211, + 0.0024008636828511953, + 0.002839958295226097, + 0.004697070922702551, + 0.005111925303936005, + 0.004140113014727831, + 0.003926143981516361, + 0.004699795972555876, + 0.003977634944021702, + 0.0033888183534145355, + 0.003288406878709793, + 0.0028265235014259815, + 0.009464860893785954, + 0.012274319306015968, + 0.006080151069909334, + 0.007523230277001858, + 0.00729316147044301, + 0.007426969241350889, + 0.0027775706257671118, + 0.005663411691784859, + 0.009821778163313866, + 0.012802940793335438, + 0.004523557610809803, + 0.0035302836913615465, + 0.019325245171785355, + 0.008198295719921589, + 0.02452680841088295, + 0.0071308971382677555, + 0.10843519866466522, + 0.020553968846797943, + 0.00980265624821186, + 0.03771257773041725, + 0.0159064419567585, + 0.07600297778844833, + 0.003461727872490883, + 0.0037479873280972242, + 0.011244315654039383, + 0.003080022521317005, + 0.007688605226576328, + 0.0057242074981331825, + 0.005616172682493925, + 0.009874680079519749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.004092651419341564 + } + ], + "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/0000128/group_9/index_tensor([7519, 7519])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_9/index_tensor([7519, 7519])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..94ed3d3fc91b1d8a9d862eb6c9d704c59b864162 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_9/index_tensor([7519, 7519])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot, the camera hovering just above the water, pans swiftly to the right, tracking a majestic blue heron as it traverses the shallow, clear waters of a tranquil marsh during golden hour. With its gray-brown body, long, sinuous neck formed in a tight S-curve, and crest of dark feathers, the heron moves with breathtaking speed. Its piercing yellow eyes lock onto unseen prey ahead with predatory intensity, its dagger-like yellow beak poised to strike. Its long, slender legs pave the water, creating a V-shaped wake and spraying shimmering droplets into the air, capturing the soft, warm light of the setting sun. The rapid panning motion, combined with the heron's speed, creates a natural motion blur against the background of submerged reeds and disturbed water, effectively isolating the subject. This visual effect powerfully emphasizes the swift, abrupt nature of its movements, creating an atmosphere of tension and drama as a primeval hunt unfolds, while preserving the sharp, breathtaking clarity of the heron's highly focused head and eyes.", + "seed": 11255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_9/index_7519_seed_11255_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": -0.10072408271953463, + "vlm_logprob": -0.10072408271953463, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.10072408271953463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0349370576441288 + ], + "opposite_logprob_tokens": [ + -3.6599371433258057 + ], + "logprob_reward": -0.0349370576441288, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0349370576441288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09127594530582428, + 0.13248729705810547, + 0.06829463690519333, + 0.11896511912345886, + 0.11447973549365997, + 0.12247327715158463, + 0.1930396854877472, + 0.03421300649642944, + 0.05754607543349266, + 0.06722527742385864 + ], + "credit_map": [ + 0.009867692366242409, + 0.01591138355433941, + 0.007753793615847826, + 0.006924200337380171, + 0.01783609949052334, + 0.010995397344231606, + 0.01180192083120346, + 0.01018545776605606, + 0.008673571981489658, + 0.011989075690507889, + 0.00848935917019844, + 0.012708540074527264, + 0.014039089903235435, + 0.028644738718867302, + 0.01524539478123188, + 0.03269752115011215, + 0.009943642653524876, + 0.009083690121769905, + 0.005225194618105888, + 0.009379511699080467, + 0.010136470198631287, + 0.010873082093894482, + 0.005741602275520563, + 0.007911443710327148, + 0.03754794970154762, + 0.006953570991754532, + 0.03745877370238304, + 0.008709137327969074, + 0.0040441881865262985, + 0.008987228386104107, + 0.00777951255440712, + 0.00748475780710578, + 0.030805490911006927, + 0.005125888157635927, + 0.03031705692410469, + 0.011620723642408848, + 0.006533446721732616, + 0.010746624320745468, + 0.006759501527994871, + 0.012571005150675774, + 0.015945225954055786, + 0.009422700852155685, + 0.011930309236049652, + 0.013941632583737373, + 0.013796951621770859, + 0.01369169820100069, + 0.02507222257554531, + 0.018672535195946693, + 0.027930643409490585, + 0.017555322498083115, + 0.02880912460386753, + 0.029928000643849373, + 0.024625053629279137, + 0.025671688839793205, + 0.018328705802559853, + 0.02019115723669529, + 0.0041283732280135155, + 0.002479883376508951, + 0.005915546324104071, + 0.004895724821835756, + 0.0027928717900067568, + 0.0036760533694177866, + 0.005338629707694054, + 0.004985923413187265, + 0.00717242993414402, + 0.007145426701754332, + 0.01173577643930912, + 0.004638845566660166, + 0.007814925163984299, + 0.007779417559504509, + 0.00604972941800952, + 0.00520952558144927, + 0.011107519268989563, + 0.006177166942507029, + 0.007727181538939476, + 0.006829963997006416, + 0.008832530118525028, + 0.007604086771607399, + 0.00864678155630827, + 0.010300048626959324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.0349370576441288 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289241135120392 + ], + "opposite_logprob_tokens": [ + -1.0539240837097168 + ], + "logprob_reward": -0.4289241135120392, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289241135120392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15144626796245575, + 0.1361253559589386, + 0.09796745330095291, + 0.15919983386993408, + 0.07104019075632095, + 0.05788049474358559, + 0.023726843297481537, + 0.150310218334198, + 0.05968140810728073, + 0.09262187778949738 + ], + "credit_map": [ + 0.010113169439136982, + 0.011979802511632442, + 0.013430238701403141, + 0.013055015355348587, + 0.019567161798477173, + 0.030030453577637672, + 0.021170392632484436, + 0.03210004046559334, + 0.009265875443816185, + 0.018594322726130486, + 0.007567541673779488, + 0.0127254044637084, + 0.016071569174528122, + 0.02962595783174038, + 0.013495765626430511, + 0.028778910636901855, + 0.01643798127770424, + 0.012481000274419785, + 0.008862465620040894, + 0.013331472873687744, + 0.012222138233482838, + 0.013705486431717873, + 0.00962082389742136, + 0.011306079104542732, + 0.03210774064064026, + 0.01704948954284191, + 0.03598180040717125, + 0.014990060590207577, + 0.01454794779419899, + 0.015656249597668648, + 0.011932495050132275, + 0.016934053972363472, + 0.014782734215259552, + 0.004057407379150391, + 0.009297828190028667, + 0.00680624321103096, + 0.006054819095879793, + 0.00811883807182312, + 0.007519388105720282, + 0.014402927830815315, + 0.012253159657120705, + 0.00399580504745245, + 0.004895350895822048, + 0.006526786834001541, + 0.0066877915523946285, + 0.00599482050165534, + 0.011419421061873436, + 0.006107362452894449, + 0.002588993404060602, + 0.002009378746151924, + 0.0026030060835182667, + 0.004112912807613611, + 0.003109244629740715, + 0.0037063935305923223, + 0.0023585939779877663, + 0.0032383217476308346, + 0.022773897275328636, + 0.011810903437435627, + 0.02661752514541149, + 0.019327877089381218, + 0.018632562831044197, + 0.014507723972201347, + 0.020303260535001755, + 0.01633647084236145, + 0.00665042782202363, + 0.007474562153220177, + 0.01070469617843628, + 0.005947713740170002, + 0.009005727246403694, + 0.007956507615745068, + 0.006703061051666737, + 0.005238713696599007, + 0.01521704439073801, + 0.010143005289137363, + 0.01591913029551506, + 0.009297328069806099, + 0.009176543913781643, + 0.008544077165424824, + 0.009784387424588203, + 0.014540360309183598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.4289241135120392 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630014326423407 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007630014326423407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630014326423407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05990741029381752, + 0.06931339204311371, + 0.025838026776909828, + 0.16992613673210144, + 0.04545878246426582, + 0.054115455597639084, + 0.475253701210022, + 0.030220987275242805, + 0.02645093761384487, + 0.04351518303155899 + ], + "credit_map": [ + 0.006500122603029013, + 0.008445858024060726, + 0.006579146720468998, + 0.006639398168772459, + 0.009734993800520897, + 0.0068353405222296715, + 0.007823898456990719, + 0.007348652463406324, + 0.004259918816387653, + 0.009934434667229652, + 0.004866760689765215, + 0.007463524583727121, + 0.009897005744278431, + 0.012883768416941166, + 0.007780671585351229, + 0.012227312661707401, + 0.0033666237723082304, + 0.002792205661535263, + 0.002547735348343849, + 0.0032950148452073336, + 0.00465793302282691, + 0.0037137940526008606, + 0.002891834592446685, + 0.0025728861801326275, + 0.04852303862571716, + 0.010920576751232147, + 0.05159539729356766, + 0.01308818906545639, + 0.008965651504695415, + 0.013354268856346607, + 0.010851189494132996, + 0.012627829797565937, + 0.007917752489447594, + 0.0037903001066297293, + 0.007258995436131954, + 0.004257439635694027, + 0.0044625853188335896, + 0.008260350674390793, + 0.003485866589471698, + 0.0060254959389567375, + 0.006896249484270811, + 0.004542862065136433, + 0.006712311878800392, + 0.00688169477507472, + 0.006661233492195606, + 0.0047880010679364204, + 0.009933991357684135, + 0.0076991114765405655, + 0.058441080152988434, + 0.045561935752630234, + 0.07944723963737488, + 0.0745147094130516, + 0.050306085497140884, + 0.05645492300391197, + 0.041890379041433334, + 0.06863734126091003, + 0.003859923454001546, + 0.0021549900993704796, + 0.006399191450327635, + 0.0033580276649445295, + 0.002960415091365576, + 0.0030195508152246475, + 0.003583999816328287, + 0.00488488981500268, + 0.0034142066724598408, + 0.0027113654650747776, + 0.006603756919503212, + 0.0026543315034359694, + 0.002721082651987672, + 0.003133108839392662, + 0.0025744163431227207, + 0.0026386685203760862, + 0.007137668784707785, + 0.004175748210400343, + 0.006264427676796913, + 0.003842312144115567, + 0.005548064596951008, + 0.0053258794359862804, + 0.005651016253978014, + 0.005570064298808575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.007630014326423407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070479527115822 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070479527115822, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070479527115822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09714677929878235, + 0.1325165331363678, + 0.09972784668207169, + 0.1460222601890564, + 0.13424421846866608, + 0.09708433598279953, + 0.13769327104091644, + 0.06640490144491196, + 0.047467589378356934, + 0.04169221967458725 + ], + "credit_map": [ + 0.010206539183855057, + 0.013879220932722092, + 0.009096669033169746, + 0.009302034974098206, + 0.018450694158673286, + 0.01141596119850874, + 0.01501554250717163, + 0.009780127555131912, + 0.007326515857130289, + 0.012426265515387058, + 0.006472822278738022, + 0.009305372834205627, + 0.014538557268679142, + 0.03380535915493965, + 0.014004473574459553, + 0.034637171775102615, + 0.015354812145233154, + 0.014710109680891037, + 0.0101028922945261, + 0.01060333289206028, + 0.017365574836730957, + 0.012620550580322742, + 0.00813960563391447, + 0.010830970481038094, + 0.042446695268154144, + 0.009128643199801445, + 0.04553752392530441, + 0.011210629716515541, + 0.00818066019564867, + 0.011212275363504887, + 0.009149834513664246, + 0.009155990555882454, + 0.03826860338449478, + 0.006406155414879322, + 0.041196804493665695, + 0.014854534529149532, + 0.007145467214286327, + 0.008618315681815147, + 0.007181019056588411, + 0.010573314502835274, + 0.013779849745333195, + 0.0065832012332975864, + 0.010811672545969486, + 0.011886739172041416, + 0.011228607967495918, + 0.008145006373524666, + 0.020316973328590393, + 0.014332286082208157, + 0.01781270280480385, + 0.013400179333984852, + 0.02270190231502056, + 0.021559463813900948, + 0.015720108523964882, + 0.015975715592503548, + 0.011741132475435734, + 0.01878207176923752, + 0.007225183304399252, + 0.0054828934371471405, + 0.006723602768033743, + 0.010483433492481709, + 0.010061192326247692, + 0.008297807537019253, + 0.011204029433429241, + 0.006926756352186203, + 0.006083750165998936, + 0.005848274566233158, + 0.009479066357016563, + 0.004488409496843815, + 0.006124005652964115, + 0.007200722582638264, + 0.004106765612959862, + 0.0041365972720086575, + 0.0066267261281609535, + 0.0034939241595566273, + 0.005202257074415684, + 0.00434374762699008, + 0.0055586728267371655, + 0.004509722348302603, + 0.005158858839422464, + 0.0067983088083565235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.016070479527115822 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058748587965965 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.016058748587965965, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058748587965965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06912688165903091, + 0.079862080514431, + 0.07640288025140762, + 0.04650607705116272, + 0.17535574734210968, + 0.1457027941942215, + 0.10268738865852356, + 0.06786731630563736, + 0.08798214048147202, + 0.14850664138793945 + ], + "credit_map": [ + 0.0058154333382844925, + 0.008821063674986362, + 0.004917407874017954, + 0.005482235457748175, + 0.010361031629145145, + 0.010488821193575859, + 0.012111138552427292, + 0.011129751801490784, + 0.006528165191411972, + 0.012544744648039341, + 0.00804383959621191, + 0.009314307942986488, + 0.012130429968237877, + 0.009398963302373886, + 0.012352712452411652, + 0.009548915550112724, + 0.011491062119603157, + 0.009900867938995361, + 0.009062639437615871, + 0.006957337260246277, + 0.0123650673776865, + 0.00973973423242569, + 0.009480180218815804, + 0.007405990269035101, + 0.00962293241173029, + 0.004284857306629419, + 0.010900909081101418, + 0.004468214698135853, + 0.004307643044739962, + 0.0046425508335232735, + 0.0038232672959566116, + 0.0044557033106684685, + 0.049213431775569916, + 0.00897774938493967, + 0.04666927084326744, + 0.01773006096482277, + 0.009505607187747955, + 0.009619568474590778, + 0.012833168730139732, + 0.020806899294257164, + 0.01983053795993328, + 0.013060495257377625, + 0.015198314562439919, + 0.015988662838935852, + 0.018224148079752922, + 0.01688367873430252, + 0.027267294004559517, + 0.019249670207500458, + 0.012754748575389385, + 0.009862075559794903, + 0.017960915341973305, + 0.014762620441615582, + 0.009520438499748707, + 0.01099564041942358, + 0.009654663503170013, + 0.01717628352344036, + 0.009016270749270916, + 0.004821368958801031, + 0.012704157270491123, + 0.008921044878661633, + 0.00833046529442072, + 0.005726969800889492, + 0.011051777750253677, + 0.007295258343219757, + 0.011722126975655556, + 0.009486852213740349, + 0.023191269487142563, + 0.01029448863118887, + 0.007523805368691683, + 0.006389363668859005, + 0.010143548250198364, + 0.009230682626366615, + 0.020975589752197266, + 0.013646858744323254, + 0.027181340381503105, + 0.014949867501854897, + 0.012591082602739334, + 0.012139424681663513, + 0.01774948462843895, + 0.02927299775183201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.016058748587965965 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, the camera hovering just above the water, pans swiftly to the right, tracking a majestic blue heron as it traverses the shallow, clear waters of a tranquil marsh during golden hour. With its gray-brown body, long, sinuous neck formed in a tight S-curve, and crest of dark feathers, the heron moves with breathtaking speed. Its piercing yellow eyes lock onto unseen prey ahead with predatory intensity, its dagger-like yellow beak poised to strike. Its long, slender legs pave the water, creating a V-shaped wake and spraying shimmering droplets into the air, capturing the soft, warm light of the setting sun. The rapid panning motion, combined with the heron's speed, creates a natural motion blur against the background of submerged reeds and disturbed water, effectively isolating the subject. This visual effect powerfully emphasizes the swift, abrupt nature of its movements, creating an atmosphere of tension and drama as a primeval hunt unfolds, while preserving the sharp, breathtaking clarity of the heron's highly focused head and eyes.", + "seed": 11256, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_9/index_7519_seed_11256_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": -0.19317425033077598, + "vlm_logprob": -0.19317425033077598, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.19317425033077598, + "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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0303451307117939 + ], + "opposite_logprob_tokens": [ + -3.9053452014923096 + ], + "logprob_reward": -0.0303451307117939, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0303451307117939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15868528187274933, + 0.12383221089839935, + 0.04553937911987305, + 0.12306218594312668, + 0.13514824211597443, + 0.15458433330059052, + 0.03797772526741028, + 0.1372794210910797, + 0.05270988121628761, + 0.0311813373118639 + ], + "credit_map": [ + 0.01823904924094677, + 0.03393235802650452, + 0.010247724130749702, + 0.013175789266824722, + 0.017518319189548492, + 0.028749685734510422, + 0.01787368208169937, + 0.018948674201965332, + 0.012061121873557568, + 0.01352882944047451, + 0.015356050804257393, + 0.01432847324758768, + 0.018084246665239334, + 0.013091782107949257, + 0.01717076450586319, + 0.020210959017276764, + 0.004892708733677864, + 0.00775791984051466, + 0.005724402144551277, + 0.0057090576738119125, + 0.004062641877681017, + 0.004892962984740734, + 0.005414946004748344, + 0.007084743119776249, + 0.010047673247754574, + 0.01127956248819828, + 0.011247936636209488, + 0.025836240500211716, + 0.016882987692952156, + 0.011955301277339458, + 0.020686792209744453, + 0.015125696547329426, + 0.009148625656962395, + 0.008484726771712303, + 0.007626968901604414, + 0.006253210362046957, + 0.007690667174756527, + 0.039990831166505814, + 0.009408066980540752, + 0.046545159071683884, + 0.013169036246836185, + 0.01961151696741581, + 0.0174519345164299, + 0.026242025196552277, + 0.026674263179302216, + 0.019863301888108253, + 0.0207882821559906, + 0.010783982463181019, + 0.004384462721645832, + 0.004856541752815247, + 0.005336804315447807, + 0.005359665956348181, + 0.005371051840484142, + 0.004337610676884651, + 0.0051347375847399235, + 0.0031968567054718733, + 0.023780982941389084, + 0.020381098613142967, + 0.011583028361201286, + 0.012711036019027233, + 0.019516820088028908, + 0.020421667024493217, + 0.01443425938487053, + 0.014450537040829659, + 0.004713122732937336, + 0.006812490522861481, + 0.0046955011785030365, + 0.007240410428494215, + 0.00661733653396368, + 0.005378813948482275, + 0.008802908472716808, + 0.008449306711554527, + 0.004990322049707174, + 0.004227520897984505, + 0.004271421115845442, + 0.003448482370004058, + 0.0032003377564251423, + 0.0029110440518707037, + 0.003578098025172949, + 0.004554112907499075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0303451307117939 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983086347579956 + ], + "opposite_logprob_tokens": [ + -0.5233086347579956 + ], + "logprob_reward": -0.8983086347579956, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983086347579956, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08008258789777756, + 0.18075281381607056, + 0.025202356278896332, + 0.06377683579921722, + 0.1413235366344452, + 0.08085774630308151, + 0.042489342391490936, + 0.06133116036653519, + 0.2849085330963135, + 0.03927511349320412 + ], + "credit_map": [ + 0.00796038843691349, + 0.00848184060305357, + 0.008540697395801544, + 0.010506589896976948, + 0.00786074809730053, + 0.010732707567512989, + 0.016573810949921608, + 0.009425804950296879, + 0.019910313189029694, + 0.02733866684138775, + 0.022962920367717743, + 0.018162274733185768, + 0.026834547519683838, + 0.019768748432397842, + 0.02217189222574234, + 0.023603448644280434, + 0.0035483259707689285, + 0.003916589543223381, + 0.0028030804824084044, + 0.003015054389834404, + 0.003194266464561224, + 0.0028924725484102964, + 0.002935245167464018, + 0.00289732008241117, + 0.004736179020255804, + 0.005861338693648577, + 0.008042567409574986, + 0.010415170341730118, + 0.009563229978084564, + 0.00602684123441577, + 0.010387693531811237, + 0.008743812330067158, + 0.010803109966218472, + 0.011126836761832237, + 0.009585529565811157, + 0.008812426589429379, + 0.022542104125022888, + 0.031632523983716965, + 0.012845673598349094, + 0.03397533670067787, + 0.007905139587819576, + 0.010743231512606144, + 0.009346802718937397, + 0.014714713208377361, + 0.011398457922041416, + 0.009518515318632126, + 0.010580300353467464, + 0.006650589406490326, + 0.004000745248049498, + 0.003781030187383294, + 0.005602494813501835, + 0.0071198297664523125, + 0.00660160044208169, + 0.005675603169947863, + 0.005922598764300346, + 0.003785435575991869, + 0.009939775802195072, + 0.007492454256862402, + 0.0051370589062571526, + 0.005540140904486179, + 0.009927140548825264, + 0.009256047196686268, + 0.007132112514227629, + 0.006906421855092049, + 0.022769171744585037, + 0.03281967714428902, + 0.030276816338300705, + 0.03544740378856659, + 0.04301000013947487, + 0.042701441794633865, + 0.04337167367339134, + 0.03451233357191086, + 0.0051037161611020565, + 0.004700827412307262, + 0.00479138595983386, + 0.006403288338333368, + 0.0039602587930858135, + 0.004592183046042919, + 0.004899018909782171, + 0.004824432078748941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.8983086347579956 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "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": [ + 0.06396908313035965, + 0.18912234902381897, + 0.05776707082986832, + 0.23528873920440674, + 0.05167777091264725, + 0.02756102755665779, + 0.019498310983181, + 0.19850610196590424, + 0.1383415162563324, + 0.018268022686243057 + ], + "credit_map": [ + 0.005130586214363575, + 0.009386778809130192, + 0.007116029039025307, + 0.010613047517836094, + 0.008939508348703384, + 0.006942452397197485, + 0.00821826234459877, + 0.007622418459504843, + 0.020594526082277298, + 0.03396240621805191, + 0.022003406658768654, + 0.019611669704318047, + 0.026182886213064194, + 0.019174890592694283, + 0.023033928126096725, + 0.024558626115322113, + 0.006043619010597467, + 0.012641179375350475, + 0.007043025456368923, + 0.005386854521930218, + 0.006793679669499397, + 0.00669532036408782, + 0.006308732554316521, + 0.00685465894639492, + 0.02176540158689022, + 0.014462226070463657, + 0.021660372614860535, + 0.04598158970475197, + 0.028106436133384705, + 0.02935142070055008, + 0.03965044766664505, + 0.034310854971408844, + 0.005385891534388065, + 0.006381328217685223, + 0.006078209728002548, + 0.005706632975488901, + 0.010003761388361454, + 0.005933726672083139, + 0.006539869587868452, + 0.005648351740092039, + 0.003354481654241681, + 0.004276636056602001, + 0.003548281267285347, + 0.0030799577943980694, + 0.004387462045997381, + 0.00352511671371758, + 0.0033192646224051714, + 0.0020698283333331347, + 0.0019238484092056751, + 0.002374958246946335, + 0.0022911541163921356, + 0.0025605312548577785, + 0.0027843168936669827, + 0.0025738917756825686, + 0.003112809034064412, + 0.0018768019508570433, + 0.033700671046972275, + 0.029798125848174095, + 0.022561538964509964, + 0.019858133047819138, + 0.026315636932849884, + 0.02755124308168888, + 0.020671896636486053, + 0.018048854544758797, + 0.0109378257766366, + 0.013027587905526161, + 0.011800624430179596, + 0.014292638748884201, + 0.026416653767228127, + 0.022696318104863167, + 0.024127105250954628, + 0.015042764134705067, + 0.003014278830960393, + 0.0028061328921467066, + 0.001921473303809762, + 0.0020017442293465137, + 0.002769715618342161, + 0.0020216808188706636, + 0.002058440586552024, + 0.001674556522630155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "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 a low-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588379353284836 + ], + "opposite_logprob_tokens": [ + -3.8955883979797363 + ], + "logprob_reward": -0.020588379353284836, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588379353284836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05499893054366112, + 0.08180982619524002, + 0.04156302660703659, + 0.6303335428237915, + 0.03992520645260811, + 0.03915960714221001, + 0.011179277673363686, + 0.045929137617349625, + 0.04057203233242035, + 0.014529448002576828 + ], + "credit_map": [ + 0.004737112205475569, + 0.008470357395708561, + 0.004684108775109053, + 0.008621028624475002, + 0.009333635680377483, + 0.007729727774858475, + 0.006218142341822386, + 0.005204819608479738, + 0.00907569844275713, + 0.013469036668539047, + 0.011072155088186264, + 0.00906429160386324, + 0.012427193112671375, + 0.008341141045093536, + 0.009680528193712234, + 0.008679780177772045, + 0.003848814405500889, + 0.010109259746968746, + 0.0037507456727325916, + 0.004573742859065533, + 0.004942761734127998, + 0.004814279731363058, + 0.003968717996031046, + 0.005554702132940292, + 0.05700869485735893, + 0.047576576471328735, + 0.06284536421298981, + 0.1353776603937149, + 0.07920614629983902, + 0.05335221439599991, + 0.11092904955148697, + 0.08403778821229935, + 0.0026636580005288124, + 0.002593402052298188, + 0.004208083730190992, + 0.003408377291634679, + 0.004216950852423906, + 0.008954720571637154, + 0.003044577082619071, + 0.010835433378815651, + 0.00544779421761632, + 0.005539380945265293, + 0.004818886052817106, + 0.005125137511640787, + 0.005249679088592529, + 0.0048086997121572495, + 0.004616177175194025, + 0.0035538524389266968, + 0.0010952982120215893, + 0.001327231526374817, + 0.0018446915782988071, + 0.001442114356905222, + 0.001540740835480392, + 0.0014906745636835694, + 0.0015513679245486856, + 0.0008871579775586724, + 0.007958253845572472, + 0.0058800275437533855, + 0.005338320042937994, + 0.004395965952426195, + 0.00734656723216176, + 0.006311976816505194, + 0.004056872334331274, + 0.004641152918338776, + 0.003765807021409273, + 0.00354969734326005, + 0.004583937115967274, + 0.004937070421874523, + 0.006247419398277998, + 0.005823547951877117, + 0.005678215529769659, + 0.005986339412629604, + 0.0021681005600839853, + 0.002160680014640093, + 0.0018021931173279881, + 0.0015152059495449066, + 0.0016258679097518325, + 0.001647452125325799, + 0.0018264668760821223, + 0.0017834811005741358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.020588379353284836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531350366771221 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531350366771221, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531350366771221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10246100276708603, + 0.060156818479299545, + 0.010970104485750198, + 0.21619302034378052, + 0.32446420192718506, + 0.06661024689674377, + 0.06554025411605835, + 0.06819059699773788, + 0.04775555804371834, + 0.037658218294382095 + ], + "credit_map": [ + 0.011245567351579666, + 0.015792760998010635, + 0.011288891546428204, + 0.015448948368430138, + 0.010589038021862507, + 0.01277931872755289, + 0.01453467272222042, + 0.010781816206872463, + 0.005512113682925701, + 0.007693888153880835, + 0.008332587778568268, + 0.007417875342071056, + 0.008186220191419125, + 0.006424630526453257, + 0.007386191748082638, + 0.009203310124576092, + 0.0016360508743673563, + 0.0015929461224004626, + 0.0011754161678254604, + 0.0011573692318052053, + 0.0012832534266635776, + 0.001451770425774157, + 0.0012609101831912994, + 0.0014123881701380014, + 0.019214143976569176, + 0.012852753512561321, + 0.022471431642770767, + 0.04622921720147133, + 0.02608799748122692, + 0.022127561271190643, + 0.03845078498125076, + 0.028759121894836426, + 0.023260056972503662, + 0.024899818003177643, + 0.01742425188422203, + 0.01832054927945137, + 0.0365586131811142, + 0.08326215296983719, + 0.025757895782589912, + 0.09498085081577301, + 0.006283529102802277, + 0.00656273914501071, + 0.008610288612544537, + 0.009610065259039402, + 0.01131768710911274, + 0.008912956342101097, + 0.010609379969537258, + 0.004703599959611893, + 0.00557039026170969, + 0.006986076477915049, + 0.007914554327726364, + 0.008517863228917122, + 0.009646578691899776, + 0.009672112762928009, + 0.011311220936477184, + 0.0059214564971625805, + 0.01193400751799345, + 0.010362044908106327, + 0.007123555988073349, + 0.007026345003396273, + 0.009686731733381748, + 0.00869607925415039, + 0.007222470827400684, + 0.006139364559203386, + 0.004137821961194277, + 0.0069892448373138905, + 0.004263343755155802, + 0.005780133418738842, + 0.0065803853794932365, + 0.006428873632103205, + 0.006707384716719389, + 0.006868371739983559, + 0.004387787077575922, + 0.004496505483984947, + 0.006111673079431057, + 0.006352388300001621, + 0.004309381358325481, + 0.0032399026677012444, + 0.0048584542237222195, + 0.0039021263364702463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.012531350366771221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, the camera hovering just above the water, pans swiftly to the right, tracking a majestic blue heron as it traverses the shallow, clear waters of a tranquil marsh during golden hour. With its gray-brown body, long, sinuous neck formed in a tight S-curve, and crest of dark feathers, the heron moves with breathtaking speed. Its piercing yellow eyes lock onto unseen prey ahead with predatory intensity, its dagger-like yellow beak poised to strike. Its long, slender legs pave the water, creating a V-shaped wake and spraying shimmering droplets into the air, capturing the soft, warm light of the setting sun. The rapid panning motion, combined with the heron's speed, creates a natural motion blur against the background of submerged reeds and disturbed water, effectively isolating the subject. This visual effect powerfully emphasizes the swift, abrupt nature of its movements, creating an atmosphere of tension and drama as a primeval hunt unfolds, while preserving the sharp, breathtaking clarity of the heron's highly focused head and eyes.", + "seed": 11257, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_9/index_7519_seed_11257_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": -0.09728453531861306, + "vlm_logprob": -0.09728453531861306, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.09728453531861306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04223087430000305 + ], + "opposite_logprob_tokens": [ + -3.5422308444976807 + ], + "logprob_reward": -0.04223087430000305, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04223087430000305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16526533663272858, + 0.07285062223672867, + 0.03391285613179207, + 0.0695200189948082, + 0.1299896538257599, + 0.23967613279819489, + 0.13118767738342285, + 0.018973222002387047, + 0.031530093401670456, + 0.10709438472986221 + ], + "credit_map": [ + 0.022631114348769188, + 0.018595490604639053, + 0.009701451286673546, + 0.008447889238595963, + 0.03238064423203468, + 0.03334122523665428, + 0.018465392291545868, + 0.02170213684439659, + 0.010546890087425709, + 0.013564173132181168, + 0.004875904880464077, + 0.006610444746911526, + 0.009449709206819534, + 0.014089886099100113, + 0.007438707631081343, + 0.006274904124438763, + 0.005707744974642992, + 0.00522835087031126, + 0.0033889685291796923, + 0.003880190197378397, + 0.004424709361046553, + 0.004607072565704584, + 0.004040767904371023, + 0.002635051729157567, + 0.008465251885354519, + 0.009297379292547703, + 0.00931357592344284, + 0.006654196884483099, + 0.008541036397218704, + 0.012860221788287163, + 0.004744009114801884, + 0.009644348174333572, + 0.01408844068646431, + 0.014519443735480309, + 0.00935629103332758, + 0.0037867643404752016, + 0.007444639224559069, + 0.049672119319438934, + 0.00507564190775156, + 0.02604631334543228, + 0.022628242149949074, + 0.030310658738017082, + 0.026121046394109726, + 0.045133013278245926, + 0.04378410428762436, + 0.019507328048348427, + 0.03211088478565216, + 0.020080847665667534, + 0.02079044096171856, + 0.015265480615198612, + 0.009050373919308186, + 0.011310706846415997, + 0.026890156790614128, + 0.022194916382431984, + 0.0157623328268528, + 0.009923265315592289, + 0.002371544949710369, + 0.0017263911431655288, + 0.004350070841610432, + 0.001862474367953837, + 0.0030953804962337017, + 0.002428988926112652, + 0.0013991595478728414, + 0.001739210681989789, + 0.006978352554142475, + 0.010391970165073872, + 0.0017882594838738441, + 0.001453156117349863, + 0.0038141091354191303, + 0.003174105891957879, + 0.001699834130704403, + 0.002230302896350622, + 0.024000192061066628, + 0.01267628837376833, + 0.0124515974894166, + 0.008053921163082123, + 0.012956720776855946, + 0.009091484360396862, + 0.007295695133507252, + 0.0205684844404459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.04223087430000305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38707494735717773 + ], + "opposite_logprob_tokens": [ + -1.1370749473571777 + ], + "logprob_reward": -0.38707494735717773, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38707494735717773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18049819767475128, + 0.16172437369823456, + 0.06299213320016861, + 0.08096861094236374, + 0.18165691196918488, + 0.08440662175416946, + 0.10705025494098663, + 0.022517129778862, + 0.06259327381849289, + 0.055592503398656845 + ], + "credit_map": [ + 0.021580440923571587, + 0.01657385565340519, + 0.011688902974128723, + 0.01331022847443819, + 0.02819422446191311, + 0.03342399001121521, + 0.031118232756853104, + 0.024608319625258446, + 0.025515025481581688, + 0.03596403822302818, + 0.01303363312035799, + 0.016785459592938423, + 0.018724661320447922, + 0.0210379958152771, + 0.015771985054016113, + 0.014891578815877438, + 0.01008665096014738, + 0.012097969651222229, + 0.00591204222291708, + 0.006463462952524424, + 0.008800556883215904, + 0.009795844554901123, + 0.004489586688578129, + 0.005346023943275213, + 0.013528231531381607, + 0.010005902498960495, + 0.009080936200916767, + 0.008777312934398651, + 0.010945504531264305, + 0.012634970247745514, + 0.004229177720844746, + 0.011766578070819378, + 0.019399749115109444, + 0.021878469735383987, + 0.01379710528999567, + 0.006233304738998413, + 0.011183664202690125, + 0.06532016396522522, + 0.008087259717285633, + 0.03575720638036728, + 0.007581132929772139, + 0.010849370621144772, + 0.010760628618299961, + 0.013128750957548618, + 0.015848953276872635, + 0.00953084696084261, + 0.009289709851145744, + 0.007417233660817146, + 0.012972937896847725, + 0.020889950916171074, + 0.005292035173624754, + 0.006881198845803738, + 0.019614050164818764, + 0.018760671839118004, + 0.010987830348312855, + 0.011651589535176754, + 0.002484941389411688, + 0.0028119601774960756, + 0.0030039127450436354, + 0.0034741293638944626, + 0.005041246768087149, + 0.002561799483373761, + 0.0016613848274573684, + 0.0014777564210817218, + 0.016357142478227615, + 0.021478446200489998, + 0.0036269002594053745, + 0.00306933862157166, + 0.0073059797286987305, + 0.005404516123235226, + 0.0025098773185163736, + 0.002841078443452716, + 0.012059921398758888, + 0.005832176189869642, + 0.006544716190546751, + 0.005715979728847742, + 0.004498265217989683, + 0.0048465062864124775, + 0.005957886576652527, + 0.010137057863175869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.38707494735717773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067370180040597916 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.0067370180040597916, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067370180040597916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09826294332742691, + 0.07091611623764038, + 0.042363740503787994, + 0.17834560573101044, + 0.0327884741127491, + 0.1648557186126709, + 0.07685550302267075, + 0.12647153437137604, + 0.046403445303440094, + 0.1627369523048401 + ], + "credit_map": [ + 0.011947069317102432, + 0.011106809601187706, + 0.009735953994095325, + 0.008400507271289825, + 0.013738137669861317, + 0.012398726306855679, + 0.019274089485406876, + 0.01166164968162775, + 0.01067492924630642, + 0.010694652795791626, + 0.004806117620319128, + 0.0069337692111730576, + 0.00782019179314375, + 0.012809764593839645, + 0.008984137326478958, + 0.008192550390958786, + 0.0076407636515796185, + 0.008495467714965343, + 0.003626639721915126, + 0.0035406379029154778, + 0.006559839006513357, + 0.006535432767122984, + 0.0036266851238906384, + 0.0023382792714983225, + 0.01831258088350296, + 0.02208581566810608, + 0.027004428207874298, + 0.019773032516241074, + 0.025815097615122795, + 0.024446524679660797, + 0.018176358193159103, + 0.022731782868504524, + 0.003634779481217265, + 0.0043194410391151905, + 0.0024332497268915176, + 0.0014263296034187078, + 0.0027206859085708857, + 0.010961314663290977, + 0.0023652499075978994, + 0.0049274214543402195, + 0.013760642148554325, + 0.0169522725045681, + 0.02830832079052925, + 0.018645528703927994, + 0.033147040754556656, + 0.01718490943312645, + 0.021072329953312874, + 0.015784667804837227, + 0.01120566576719284, + 0.01295133214443922, + 0.006544950418174267, + 0.007128581870347261, + 0.013528586365282536, + 0.01258053258061409, + 0.0079214908182621, + 0.004994361195713282, + 0.019651401787996292, + 0.01326039433479309, + 0.01799081079661846, + 0.01754472777247429, + 0.019501954317092896, + 0.016283003613352776, + 0.011198512278497219, + 0.011040722951292992, + 0.010417629964649677, + 0.014254546724259853, + 0.002998190699145198, + 0.0026684687472879887, + 0.0071992711164057255, + 0.004117271397262812, + 0.0022657259833067656, + 0.0024823404382914305, + 0.05084247142076492, + 0.020900219678878784, + 0.01730562373995781, + 0.01099249068647623, + 0.014737534336745739, + 0.021331820636987686, + 0.01198967732489109, + 0.014637106098234653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.0067370180040597916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808273538947105 + ], + "opposite_logprob_tokens": [ + -3.529808282852173 + ], + "logprob_reward": -0.029808273538947105, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808273538947105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10268457233905792, + 0.06289681047201157, + 0.17830400168895721, + 0.06454512476921082, + 0.20873518288135529, + 0.12690098583698273, + 0.17956215143203735, + 0.016943227499723434, + 0.014155523851513863, + 0.04527243226766586 + ], + "credit_map": [ + 0.013713918626308441, + 0.013112040236592293, + 0.008037439547479153, + 0.00699342368170619, + 0.017381707206368446, + 0.021447978913784027, + 0.009360444732010365, + 0.012637623585760593, + 0.009693882428109646, + 0.013975716196000576, + 0.005582326557487249, + 0.006305736489593983, + 0.007100680842995644, + 0.00966524239629507, + 0.005925587844103575, + 0.004647643305361271, + 0.031149322167038918, + 0.039180655032396317, + 0.01387390773743391, + 0.01892947033047676, + 0.017502320930361748, + 0.0320899598300457, + 0.010196313261985779, + 0.015382049605250359, + 0.009108354337513447, + 0.008895380422472954, + 0.008049733936786652, + 0.0060462672263383865, + 0.009015356190502644, + 0.011637402698397636, + 0.003301844233646989, + 0.008490786887705326, + 0.022307666018605232, + 0.022323956713080406, + 0.015660228207707405, + 0.006212793756276369, + 0.011283418163657188, + 0.07883596420288086, + 0.007615762762725353, + 0.04449539631605148, + 0.013769811019301414, + 0.015855146571993828, + 0.013608074747025967, + 0.022863900288939476, + 0.02534504234790802, + 0.011624285951256752, + 0.013998950831592083, + 0.009835776872932911, + 0.016788195818662643, + 0.02475898712873459, + 0.010824456810951233, + 0.01550090778619051, + 0.04265492409467697, + 0.032570190727710724, + 0.016086678951978683, + 0.02037779986858368, + 0.002053162781521678, + 0.002197012770920992, + 0.0017071319743990898, + 0.00269063888117671, + 0.0033524837344884872, + 0.0020952834747731686, + 0.0014594867825508118, + 0.001388025819323957, + 0.002268133219331503, + 0.001931305043399334, + 0.0015749584417790174, + 0.0009345374419353902, + 0.0029390458948910236, + 0.0015466720797121525, + 0.0015140622854232788, + 0.0014468092704191804, + 0.009695816785097122, + 0.004696004092693329, + 0.004793290514498949, + 0.003758565755560994, + 0.006886675022542477, + 0.005023830570280552, + 0.006829072255641222, + 0.00358917866833508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.029808273538947105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057156339287758 + ], + "opposite_logprob_tokens": [ + -3.89557147026062 + ], + "logprob_reward": -0.02057156339287758, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057156339287758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19571012258529663, + 0.06876718252897263, + 0.21390359103679657, + 0.08945542573928833, + 0.11341501027345657, + 0.035822633653879166, + 0.09530213475227356, + 0.027978822588920593, + 0.09621838480234146, + 0.06342669576406479 + ], + "credit_map": [ + 0.02961154468357563, + 0.023280000314116478, + 0.014260262250900269, + 0.0151261230930686, + 0.029332729056477547, + 0.028114842250943184, + 0.035624489188194275, + 0.020360121503472328, + 0.009836641140282154, + 0.009990684688091278, + 0.0037582560908049345, + 0.004853498190641403, + 0.012089110910892487, + 0.01658036932349205, + 0.007728108204901218, + 0.003930509556084871, + 0.033552952110767365, + 0.030235154554247856, + 0.021653054282069206, + 0.022836867719888687, + 0.030772624537348747, + 0.02648325264453888, + 0.023448815569281578, + 0.02492087334394455, + 0.008232415653765202, + 0.0128479627892375, + 0.012261784635484219, + 0.00995731633156538, + 0.012054020538926125, + 0.013969453051686287, + 0.008294617757201195, + 0.011837851256132126, + 0.012391550466418266, + 0.011959679424762726, + 0.008667721413075924, + 0.0036307445261627436, + 0.006355136167258024, + 0.0418609082698822, + 0.004302112851291895, + 0.02424716204404831, + 0.004137858748435974, + 0.005713233724236488, + 0.003933959174901247, + 0.004775119014084339, + 0.005877998191863298, + 0.004388380330055952, + 0.0034699977841228247, + 0.003526087151840329, + 0.009315995499491692, + 0.01686067320406437, + 0.007187831215560436, + 0.009641285985708237, + 0.016055533662438393, + 0.020651260390877724, + 0.009314099326729774, + 0.00627544429153204, + 0.003498104168102145, + 0.0034255541395395994, + 0.0034314622171223164, + 0.0037926824297755957, + 0.005333330482244492, + 0.0037390964571386576, + 0.0026395199820399284, + 0.00211907085031271, + 0.02122112177312374, + 0.031251754611730576, + 0.006538270972669125, + 0.004114846233278513, + 0.012506495229899883, + 0.009425869211554527, + 0.004308972042053938, + 0.006851047743111849, + 0.013518054038286209, + 0.008033597841858864, + 0.012199233286082745, + 0.005674493033438921, + 0.00905313715338707, + 0.0050798687152564526, + 0.0048742229118943214, + 0.004994082264602184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.02057156339287758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, the camera hovering just above the water, pans swiftly to the right, tracking a majestic blue heron as it traverses the shallow, clear waters of a tranquil marsh during golden hour. With its gray-brown body, long, sinuous neck formed in a tight S-curve, and crest of dark feathers, the heron moves with breathtaking speed. Its piercing yellow eyes lock onto unseen prey ahead with predatory intensity, its dagger-like yellow beak poised to strike. Its long, slender legs pave the water, creating a V-shaped wake and spraying shimmering droplets into the air, capturing the soft, warm light of the setting sun. The rapid panning motion, combined with the heron's speed, creates a natural motion blur against the background of submerged reeds and disturbed water, effectively isolating the subject. This visual effect powerfully emphasizes the swift, abrupt nature of its movements, creating an atmosphere of tension and drama as a primeval hunt unfolds, while preserving the sharp, breathtaking clarity of the heron's highly focused head and eyes.", + "seed": 11258, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_9/index_7519_seed_11258_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": -0.20376660814508796, + "vlm_logprob": -0.20376660814508796, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.20376660814508796, + "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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023674197494983673 + ], + "opposite_logprob_tokens": [ + -4.148674011230469 + ], + "logprob_reward": -0.023674197494983673, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.023674197494983673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3810000419616699, + 0.064629465341568, + 0.11229777336120605, + 0.09397291392087936, + 0.0829792395234108, + 0.048137906938791275, + 0.07795856893062592, + 0.029490265995264053, + 0.03087252750992775, + 0.07866127043962479 + ], + "credit_map": [ + 0.012128756381571293, + 0.028338633477687836, + 0.01643095538020134, + 0.0209849551320076, + 0.05009031668305397, + 0.1219843327999115, + 0.08906888961791992, + 0.041973214596509933, + 0.006051952484995127, + 0.00897478312253952, + 0.00879391748458147, + 0.006619511172175407, + 0.01064971461892128, + 0.007822192274034023, + 0.008327473886311054, + 0.007389924954622984, + 0.011879272758960724, + 0.015926653519272804, + 0.011637239716947079, + 0.015975233167409897, + 0.016763459891080856, + 0.013661481440067291, + 0.01410690788179636, + 0.012347531504929066, + 0.008749356493353844, + 0.009719130583107471, + 0.015048918314278126, + 0.011309695430099964, + 0.011992626823484898, + 0.014806578867137432, + 0.011524805799126625, + 0.01082181092351675, + 0.006435436196625233, + 0.008981545455753803, + 0.01895628310739994, + 0.008304176852107048, + 0.007820316590368748, + 0.009358638897538185, + 0.011598709039390087, + 0.011524141766130924, + 0.003928689751774073, + 0.0060054888017475605, + 0.0054054963402450085, + 0.005977989174425602, + 0.01151924580335617, + 0.006810480263084173, + 0.0044164215214550495, + 0.0040740990079939365, + 0.005727595183998346, + 0.007119724527001381, + 0.009493967518210411, + 0.021305449306964874, + 0.012914065271615982, + 0.008807739242911339, + 0.007083095610141754, + 0.005506937392055988, + 0.0033370107412338257, + 0.0021878534462302923, + 0.00328547484241426, + 0.003696808824315667, + 0.003061453578993678, + 0.004569636192172766, + 0.002119717886671424, + 0.007232313975691795, + 0.004188361577689648, + 0.003397768596187234, + 0.003979734145104885, + 0.004048963543027639, + 0.004221414215862751, + 0.003992887679487467, + 0.002901880769059062, + 0.004141520243138075, + 0.007679130416363478, + 0.012253410182893276, + 0.007169411517679691, + 0.006739537697285414, + 0.01281640026718378, + 0.010994741693139076, + 0.012767564505338669, + 0.008241080678999424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.023674197494983673 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742594957351685 + ], + "opposite_logprob_tokens": [ + -0.47425949573516846 + ], + "logprob_reward": -0.9742594957351685, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742594957351685, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5402101278305054, + 0.03420617803931236, + 0.0575595498085022, + 0.02706332691013813, + 0.02738463692367077, + 0.04734904691576958, + 0.17958290874958038, + 0.017582330852746964, + 0.020998632535338402, + 0.048063334077596664 + ], + "credit_map": [ + 0.014484909363090992, + 0.03991413488984108, + 0.019886380061507225, + 0.023916855454444885, + 0.06959883868694305, + 0.18482299149036407, + 0.13228337466716766, + 0.055302634835243225, + 0.0034555799793452024, + 0.004442988429218531, + 0.0039586275815963745, + 0.0034106788225471973, + 0.005027323495596647, + 0.004591880366206169, + 0.003995793871581554, + 0.005323302932083607, + 0.006326849106699228, + 0.005630097817629576, + 0.0046147545799613, + 0.00560888322070241, + 0.007916349917650223, + 0.008754607290029526, + 0.008390801027417183, + 0.010317202657461166, + 0.0025831758975982666, + 0.0030147230718284845, + 0.0027932918164879084, + 0.0031703757122159004, + 0.005892808549106121, + 0.003705307375639677, + 0.0035565467551350594, + 0.002347097732126713, + 0.0030292565934360027, + 0.003507996676489711, + 0.004141965880990028, + 0.002617397578433156, + 0.0038624524604529142, + 0.0035553406924009323, + 0.002712610177695751, + 0.0039576152339577675, + 0.0048074801452457905, + 0.00448608910664916, + 0.004919473081827164, + 0.005594369489699602, + 0.006957476958632469, + 0.005983284208923578, + 0.006402462720870972, + 0.008198411203920841, + 0.01581207849085331, + 0.011992656625807285, + 0.030578965321183205, + 0.047023385763168335, + 0.02318660169839859, + 0.02066565304994583, + 0.01638374663889408, + 0.013939816504716873, + 0.001878145383670926, + 0.002333519048988819, + 0.001304174424149096, + 0.0019801047164946795, + 0.0026208986528217793, + 0.0023577921092510223, + 0.002617371268570423, + 0.0024903255980461836, + 0.0018111704848706722, + 0.0019308574264869094, + 0.002313900040462613, + 0.0030654233414679766, + 0.0035627358593046665, + 0.0026432566810399294, + 0.0029987196903675795, + 0.0026725702919065952, + 0.004460775759071112, + 0.006826247088611126, + 0.004269841592758894, + 0.006015385035425425, + 0.005990239791572094, + 0.006327223032712936, + 0.007248234935104847, + 0.006925384514033794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.9742594957351685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638040162622929 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008638040162622929, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638040162622929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5946037769317627, + 0.031244589015841484, + 0.045457396656274796, + 0.044670797884464264, + 0.05094354972243309, + 0.04272967204451561, + 0.07872331142425537, + 0.048445962369441986, + 0.02048068307340145, + 0.04270030930638313 + ], + "credit_map": [ + 0.015957524999976158, + 0.04387187212705612, + 0.023059949278831482, + 0.027522681280970573, + 0.07695712894201279, + 0.20227447152137756, + 0.14460499584674835, + 0.060355134308338165, + 0.0033144939225167036, + 0.003704284317791462, + 0.003180548083037138, + 0.0031567090190947056, + 0.0058154091238975525, + 0.0037408526986837387, + 0.0041952235624194145, + 0.004137067589908838, + 0.006422854028642178, + 0.00672880606725812, + 0.004955129697918892, + 0.007638196926563978, + 0.0044894469901919365, + 0.0056657674722373486, + 0.00481624249368906, + 0.004740954842418432, + 0.004636884201318026, + 0.004920830950140953, + 0.0044959113001823425, + 0.004859776236116886, + 0.008005433715879917, + 0.006649883463978767, + 0.006749697029590607, + 0.004352377261966467, + 0.00508405826985836, + 0.005157853942364454, + 0.006492569111287594, + 0.004208953119814396, + 0.008726607076823711, + 0.007610808126628399, + 0.005735297221690416, + 0.007927404716610909, + 0.004779996816068888, + 0.0036336113698780537, + 0.0041444567032158375, + 0.006489183753728867, + 0.010442830622196198, + 0.006541314069181681, + 0.0037280411925166845, + 0.0029702347237616777, + 0.005054183769971132, + 0.008205918595194817, + 0.01062218938022852, + 0.016996024176478386, + 0.012199205346405506, + 0.0092654749751091, + 0.007774313446134329, + 0.008606002666056156, + 0.006085142493247986, + 0.005302825011312962, + 0.0038488046266138554, + 0.0050764428451657295, + 0.007081268355250359, + 0.006079195532947779, + 0.008122289553284645, + 0.006849995348602533, + 0.0019065685337409377, + 0.0019018262391909957, + 0.0021481141448020935, + 0.002307184971868992, + 0.0044021448120474815, + 0.00291440449655056, + 0.002847825875505805, + 0.0020526123698800802, + 0.00411292864009738, + 0.0058386242017149925, + 0.003922830335795879, + 0.004073393996804953, + 0.006437536329030991, + 0.005579798482358456, + 0.006396917160600424, + 0.006338279694318771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.008638040162622929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463933777064085 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463933777064085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463933777064085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2044239491224289, + 0.04282691702246666, + 0.22613872587680817, + 0.07501595467329025, + 0.15504558384418488, + 0.06230716407299042, + 0.09858685731887817, + 0.046045079827308655, + 0.025262238457798958, + 0.06434755027294159 + ], + "credit_map": [ + 0.011964455246925354, + 0.024321403354406357, + 0.01807096041738987, + 0.031134579330682755, + 0.02430913597345352, + 0.04354787990450859, + 0.027659863233566284, + 0.023415658622980118, + 0.005799735430628061, + 0.00655565457418561, + 0.0051802280358970165, + 0.0036074589006602764, + 0.0068209124729037285, + 0.005507227964699268, + 0.005006537772715092, + 0.004349159076809883, + 0.022009044885635376, + 0.03192133456468582, + 0.019221853464841843, + 0.028626268729567528, + 0.03714919090270996, + 0.03433430939912796, + 0.024873878806829453, + 0.02800283208489418, + 0.008180733770132065, + 0.005832553841173649, + 0.015320748090744019, + 0.009368248283863068, + 0.010281305760145187, + 0.008137128315865993, + 0.010379444807767868, + 0.00751578388735652, + 0.01490996964275837, + 0.020794792100787163, + 0.019707584753632545, + 0.012277553789317608, + 0.035090815275907516, + 0.02295330911874771, + 0.015902595594525337, + 0.013408943079411983, + 0.005668139550834894, + 0.004558451473712921, + 0.005978149827569723, + 0.009944358840584755, + 0.01155941653996706, + 0.008718115277588367, + 0.010529008693993092, + 0.005351521540433168, + 0.009138859808444977, + 0.012311122380197048, + 0.010096549056470394, + 0.013005745597183704, + 0.018872002139687538, + 0.012124696746468544, + 0.00967140682041645, + 0.013366472907364368, + 0.0044838013127446175, + 0.005496188532561064, + 0.0037990149576216936, + 0.004410282243043184, + 0.00912046991288662, + 0.006381345447152853, + 0.0040577370673418045, + 0.008296237327158451, + 0.002089678542688489, + 0.002225901698693633, + 0.0022766205947846174, + 0.0032243255991488695, + 0.0038785883225500584, + 0.0034839326981455088, + 0.0038151126354932785, + 0.004268075805157423, + 0.006586723029613495, + 0.008890085853636265, + 0.0069737667217850685, + 0.007485399022698402, + 0.007511254865676165, + 0.009067585691809654, + 0.008082370273768902, + 0.009750361554324627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.00463933777064085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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-angle medium shot with the camera hovering just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning swiftly to the right while tracking a blue heron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the heron have a gray-brown body, a long neck in a tight S-curve, and a crest of dark feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the heron's piercing yellow eyes and dagger-like yellow beak visible and focused sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the heron's legs creating a V-shaped wake and spraying shimmering droplets in shallow clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621969562023878 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.007621969562023878, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621969562023878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49754613637924194, + 0.036914028227329254, + 0.05485846474766731, + 0.05374171584844589, + 0.08878052234649658, + 0.037601348012685776, + 0.12307573109865189, + 0.02154405787587166, + 0.02338893711566925, + 0.06254912167787552 + ], + "credit_map": [ + 0.016102969646453857, + 0.038528040051460266, + 0.021437829360365868, + 0.03451421484351158, + 0.06345942616462708, + 0.16364343464374542, + 0.11132711172103882, + 0.048533063381910324, + 0.0027589318342506886, + 0.003157371189445257, + 0.004366051405668259, + 0.0038682580925524235, + 0.004892710130661726, + 0.004545801784843206, + 0.0066638668067753315, + 0.006661032326519489, + 0.006259399000555277, + 0.006198996677994728, + 0.007555788848549128, + 0.0060846335254609585, + 0.00789601169526577, + 0.006163999903947115, + 0.007032896392047405, + 0.007666734512895346, + 0.00524552259594202, + 0.0046523031778633595, + 0.010131853632628918, + 0.006687580142170191, + 0.006866253912448883, + 0.005114325787872076, + 0.008853818289935589, + 0.006190056446939707, + 0.008947287686169147, + 0.007205653935670853, + 0.015896134078502655, + 0.012363414280116558, + 0.006470166612416506, + 0.01112252939492464, + 0.015474354848265648, + 0.01130097359418869, + 0.003870725864544511, + 0.004584777634590864, + 0.004108770284801722, + 0.00553885055705905, + 0.00658878730610013, + 0.00544597115367651, + 0.004001179244369268, + 0.0034622850362211466, + 0.012490591034293175, + 0.015417995862662792, + 0.014631122350692749, + 0.024211790412664413, + 0.020966194570064545, + 0.016290219500660896, + 0.008137286640703678, + 0.010930521413683891, + 0.003948274068534374, + 0.0014907214790582657, + 0.0026183133013546467, + 0.0023913742043077946, + 0.0030330526642501354, + 0.002790020778775215, + 0.0024115717969834805, + 0.002860728418454528, + 0.0029616644605994225, + 0.0019133388996124268, + 0.003253031987696886, + 0.0023947826121002436, + 0.0037883666809648275, + 0.0032588522881269455, + 0.0033827307634055614, + 0.0024361652322113514, + 0.005328233353793621, + 0.007958794012665749, + 0.006398462690412998, + 0.006937659345567226, + 0.007224323693662882, + 0.008252277038991451, + 0.01351756602525711, + 0.006931797135621309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.007621969562023878 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file